TW200538734A - Systems and methods for treating, diagnosing and predicting the occurrence of a medical condition - Google Patents

Abstract

Methods and systems are provided that use clinical information, molecular information and computer-generated morphometric information in a predictive model for predicting the occurrence (e.g., recurrence) of a medical condition, for example, cancer.

Description

200538734 (1) 九、發明說明200538734 (1) IX. Description of the invention

本申請案爲主張2004年二月27日提出申請美國臨時 專利申請第60/548,322號，和2004年六月24日提出申請 的60/577,05 1號的優先權之2 005年二月25日提出申請的 美國專利申請第—/ —，—號，名稱’’Methods and Systems for Predicting Occurrence of an Event”之部份接續案；爲 主張2003年十一月17日提出申請的美國臨時專利申請第 60/5 20,8 1 5號優先權的2004年十一月17日提出申請的美 國專利申請第1〇/991，8 97號之部份接續案；爲2003年七 月21日提出申請的美國專利申請案第1 0/624,23 3號之部 份接續案；及主張2003年十一月18日提出申請的美國臨 時專利申請第60/520,939號優先權之2004年十一月17日 提出申請的美國專利申請案第1 0/99 1,240號的部份接續案 ;且主張2004年三月12日提出申請的美國臨時專利申請 第 60/5 52,497 號，2004 年六月 4 日提出申請的 60/5 77,05 1號，2004年八月11日提出申請的60/600,764 號，2004年十月20日提出申請的60/620,514號，和2005 年二月9日提出申請的60/65 1,779號之優先權·，所有彼等 都以彼等的全文以引用方式倂於本文。 【發明所屬之技術領域】 本發明具體實例係有關使用臨床資訊，分子資訊和電 腦產生的形態學資訊一預測模型中預測醫療狀況發生（例 如，疾病或對治療的回應性或無回應性）的方法和系統。 -5- 200538734 (2) 例如，於一具體實例中，本發明包括使用，臨 形態學資訊以治療、診斷、和預測***癌發 系統。 【先前技術】 醫師都被要求作出許多醫療決定，從例如， 及何時可能經歷一醫療狀況到在患者已被診斷有 如何治療患者。對一患者決定恰當的治療過程可 者，例如，存活及/或恢復之機率。類似地，預 的發生可以有利地促使個人規劃該事件。例如， 者是否可能經歷疾病的發生（如，復發）可以使 患者建議一恰當的治療過程。 傳統上，醫師都極爲依賴彼等的專業和訓練 診斷和預測醫療狀況的發生。例如，病理學 Gleason 計分系統（Gleason scoring system)來 腺癌的進展和侵害級次，其中係根據醫師在顯微 察到的***組織之外觀來評等該癌。對於更不 列腺組織樣品給予較高的G1 e a s ο η計分[1 ]。雖穿 評等廣被病理學家視爲係可靠者，不過其爲一種 分系統。特別者，觀察相同組織樣品的不同病理 作出柔和的解說。 幫助醫師在醫療診斷中的習用工具在範圍和 有限制。例如，在患者進行斷根***切除術 醫師進行有關***癌治療的決定所用之工具 ，分子和 之方法和 患者是否 狀況之後 能增加患 測一事件 預測一患 醫師對該 來治療， 家係使用 評估前列 鏡下所觀 分化的前 ^ Gleason 主觀性計 學家可能 應用上都 .後，幫助 ί限於以血 -6 - 200538734 (3) 清爲基礎的PSA篩選檢驗及一般化的圖解表。一種由 Katten et al.美國專利第6，4 0 9，6 6 4號開發出的手術後圖 解表廣爲泌尿科醫師所用且可用來預測經斷根***切除 術處理過的患者疾病復發的7年槪率。此種圖解表只提供 有關生物化學性衰竭的可能性之資訊（亦即，PSA含量之 增加），而不能預測臨床衰竭（死亡）。再者，此種圖解 表只預測患者是否在7年內可能復發，而不能預測患者的 狀況在該期間內何時可能復發。在此圖解表中的豫後變數 包括治療前的血清PSA含量，Gleason計分及病理學家對 ***囊侵入，外科邊際，精囊侵入，和淋巴結狀態之顯 微鏡評估。在有疾病復發的臨床症象，上升的血清PSA， 或佐劑治療起始之時，即記錄治療失敗。不過，此等圖解 表具有數項限制。最顯著的限制爲即使是此等圖解表中最 好者也只有比在有完善辨識（和諧指數=1 . 〇 )的模型與不 具辨識能力的模型（和諧指數=〇. 5 )之間的中間者稍微佳 而已。再者，具有在中等範圍內的圖解表預測（7年無進 展存活，30 — 70% )的患者中有約30%係不確定者，以該 預測的正確性不超過擲銅板之故。 電腦實施的影像處理和分析中之技術已出現而提供明 顯增加的計算能力。於許多應用中’從單一影像自動摘取 大量在量上連續數値化的特性之能力已付諸實現。一特性 X，於，對於某些ACB，該特性的數値組係包括介於A與 B之間的所有數X之時’即稱爲係連續數値化者。目前已 針對採自細胞學樣品的影像開發出癌影像分析系統[2] [3] 200538734 (4) 。不過’此種系統只捕捉細胞因 觀察到的所有構造資訊，更不必 子資訊了。對於時常在診斷中（ 比個別細胞的外觀更具重要作用 元素構造，尙未提出分析用之癌 理學家係委諸人工技術來分析前 癌的病理學等級[4]。習用癌影像 Ϊ 影像典型地比細胞影像更複雜且 知識之事實而更形加劇。 基於前述，有需要提供用有 斷和預測醫療狀況發生，回應和 方法。也有需要提供利用組織層 預測醫療狀況發生所用之電腦執] 【發明內容】 本發明具體實例係提供預測 系統和方法。如本文中所用者， 括’例如，預測一患者是否會及 症的發生（如，復發），預測一 (如’新的藥學藥物）是否可能 種適當醫療狀況的發生。以本發 由醫生或其他個體用來，例如對 或用以診斷患者的醫療狀況。 於本發明一方面中，係提出 而不能利用到組織層次可 說將該資訊組合臨床和分 2口’在 Gleason分析中） 的組織層次之不同病理學 影像分析系統。因而，病 列腺的形狀和尺寸以決定 分析系統的缺乏更因組織 需要了解廣博的領域專業 改良的預測力以治療，診 其他醫療現象用之系統和 次的資訊來治療，診斷和 ί系統和方法。 醫療狀況發生用之自動化 醫療狀況發生之預測可包 /或何時會經歷疾病如癌 患者對於一或更多種療法 發生回應，及預測任何他 明具體實施進行的預測可 患者選擇恰當的療程及/ 用以產生一可預測醫療狀 -8 - 200538734 (5) 況發生的模型之系統和方法。預測用模型的產生可包括使 用分析工具以一群彼等的結果爲至少部份已知的患者之數 據來訓練一支援向量機（S V Μ )或神經網路。於一具體實 例中，訓練用數據包括臨床數據，分子數據，和電腦產生 的形態學數據。如本文中所用者，一特別類型的數據（例 如’臨床’分子，或形態學）可包括一或更多該類型的特 性。此外’形態學數據係經定義爲包括關聯於或衍生自組 ，織的電子（數位）影像之任何電腦產生數據，包括但不限 於有關組織或其部份的構造性質（如，面積，長度，寬度 ’緻密性，和密度），組織或其部份的光譜性質（如紅， 綠，藍（RGB )色道値，亮度和管道組織圖），及組織影 像及/或經鑑定的組織部份之非整數性質（fractal properties )(例如上皮內介面的非整數維，腔輪廓）， 子波（wavelet )分解係數及/或其他影像數據轉換的統計 學性質。於其他具體實例中，訓練數據只包括電腦產生的 • 形態學數據或臨床數據與電腦產生的形態學數據之組合。 於一具體實例中，該系統和方法係經提供用以根據一 或多項與基質（Stroma )，細胞質，上皮核，基質核，腔 ’紅血球，組織矯作物，或組織背景，或彼等的組合相關 聯的電腦產生之形態學特性產生一預測模型。該預測模型 可以只根據電腦產生的形態學特性或與表4中所列一或更 多項臨床特性及/或表6中所列一或多項分子特性之組合 而產生。例如，可將該一或更多特性輸入到分析工具中， 定出該等特性對於相關模型預測醫療狀況的能力之影像。 -9- 200538734 (6) 增加模型預測能力之特性可以包括在最後模型內，而不會 增加（例如，或者減低）預測力的特性則從考慮中移除。 使用上述形態學特性單獨者或分別組合表4及/或表6中 所列的臨床及/或形態學特性作爲發展預測模型所用基礎 可以將醫生，其他個人，及/或自動化處理設備（如，組 織影像分析系統）等資源焦注於取得更可能與結果相關聯 因而可用於最後預測模型中的患者特性數據。 # 於本發明另一方面中，提供一預測模型用以評估一患 者的資料組而評定該患者醫療狀況的發生風險，此處該預 測模型係只根據電腦產生的形態學數據或組合著臨床數據 及/或分子數據。例如，該預測模型可接收患者資料組作 爲輸入，且可輸出一 ”計分”（score )指示出患者經歷與醫 療狀況相關聯的一或多項結果之可能性。 於一具體實例中，提供一種預測模型以預測疾病的發 生或復發’此處該模型係根據一或多項關聯於基質、細胞 ^ 質、上皮核、基質核、腔、紅血球、組織矯作物、或組織 背景、或彼等的組合之電腦產生的形態學特性。該預測模 型可以只根據此等電腦產生的形態學特性或與表4中所列 一或多項臨床特性及/或與表6中所列的一或多項分子特 性之組合。 於另一具體實例中，提供一預測模型以預測***癌 復發，此處該模型係根據圖6中所列的一或多項臨床及/ 或分子特性及針對一或更多下列病學理物件的一或更多項 形態學特性··紅血球、上皮核、基質、腔、細胞質、和組 -10- 200538734 (7) 織背景。 於又另一具體實例中，提供一預測模型以 癌復發，此處該模型係根據圖9中所列的一或 /或分子特性及針對一或更多下列病理學物件 項形態學特性：紅血球、上皮核、基質、腔、 於另一具體實例中，提供一預測模型以預 存活率，此處該模型係根據圖1 1中所列的一 • 床及/或分子特性及針對一或更多下面病理學 更多項形態學特性：紅血球、上皮核、和基質 於另一具體實例中，該預測模型可定出一 否爲正常或不正常或可預測一患者是否可能經 除術後的臨床失敗。 於另一方面中，提供系統和方法其中係於 每一點測量患者數據且用本發明預測模型予以 的診斷或治療可根據每一評定所得的結果之比 I 種比較可槪述於，例如，電腦的報告輸出中供 個人所用。例如，可提供系統和方法以篩選一 抑制劑化合物。患者的第一資料組可用一預測 定，此處該模型係根據臨床數據、分子數據、 的形態學數據。可以給患者投服一試驗化合物 驗化合物之後，可從患者得到第二資料組且以 評定。之後可將第一資料組的評定結果與第二 定組果相比較。第二資料組結果相對於第一資 可指不出該試驗化合物爲一種抑制劑化合物。 預測*** ，多項臨床及 的一或更多 和細胞質。 測***癌 或更多項臣品 物件的一或 〇 組織樣品是 歷***切 眾多時點的 評定。患者 較而定。此 醫生或其他 醫療狀況的 模型予以評 和電腦產生 。於投予試 該預測模型 資料組的評 料組之改變 -11 -This application claims priority of US Provisional Patent Application No. 60 / 548,322 filed on February 27, 2004, and No. 60 / 577,05 filed on June 24, 2004. February 25, 005 Part of the continuation of the US Patent Application No.-/-,-, filed on the date, with the name "Methods and Systems for Predicting Occurrence of an Event"; a U.S. provisional patent application filed on November 17, 2003 Partial continuation of US Patent Application No. 10/991, 8 97 filed on November 17, 2004 with priority 60/5 20,8 1 5; filed for July 21, 2003 Partial continuation of U.S. Patent Application No. 10 / 624,23 3; and U.S. Provisional Patent Application No. 60 / 520,939 filed on November 18, 2003 for priority of November 17, 2004 Partial continuation of U.S. Patent Application No. 10/99 1,240 filed on the same day; and U.S. Provisional Patent Application No. 60/5 52,497 filed on March 12, 2004, filed on June 4, 2004 Application 60/5 77,05 No. 60, filed on August 11, 2004 No. 600,764, No. 60 / 620,514 filed on October 20, 2004, and No. 60/65, 1,779 filed on February 9, 2005. All of them are cited in their entirety. [Technical field to which the invention belongs] Specific examples of the present invention are related to the use of clinical information, molecular information, and computer-generated morphological information in a prediction model to predict the occurrence of a medical condition (eg, disease or responsiveness or no response to treatment) Method and system. 2005-5-34 (2) For example, in a specific example, the present invention includes a system for treating, diagnosing, and predicting the onset of prostate cancer using morphological information. [Prior Art] Physicians have been Many medical decisions are required, from, for example, when and when a medical condition may be experienced to how the patient has been diagnosed. The decision on the appropriate course of treatment for a patient may be, for example, the chance of survival and / or recovery. Similarly Pre-occurrence can be beneficial for individuals to plan the event. For example, whether the person is likely to experience the occurrence of a disease (eg, relapse) Suggest an appropriate course of treatment. Traditionally, physicians have relied heavily on their expertise and training to diagnose and predict the occurrence of medical conditions. For example, the pathology Gleason scoring system to the progression and severity of adenocarcinoma Among them, the cancer was evaluated based on the appearance of prostate tissue microscopically examined by the physician. A higher G1 e a s ο η score was given to more spleen tissue samples [1]. Although the wear rating is widely regarded by pathologists as reliable, it is a subsystem. In particular, observe the different pathology of the same tissue sample and make a soft explanation. The tools that help physicians in medical diagnosis are limited in scope and scope. For example, after the patient's radical prostatectomy physician makes a decision about prostate cancer treatment, the tools, molecules, and methods and whether the patient's condition can increase the incidence of an event to predict an affected physician to treat it. The Gleason subjective computer scientists who have observed the differentiation under the microscope may apply it all. After that, the help is limited to PS-6 screening tests based on blood-6-200538734 (3) Qing and generalized diagrams. A post-surgical diagram developed by Katten et al. U.S. Patent No. 6,409,646, which is widely used by urologists and can be used to predict 7 years of disease recurrence in patients treated with mastectomy Rate. Such a chart provides information only on the possibility of biochemical failure (ie, an increase in PSA content), and does not predict clinical failure (death). Furthermore, this chart only predicts whether a patient is likely to relapse within 7 years and cannot predict when the patient's condition may recur during that period. Post-heavy variables in this diagram include serum PSA levels before treatment, Gleason scores, and pathologist microscopic assessment of prostate capsule invasion, surgical margin, seminal vesicle invasion, and lymph node status. When clinical signs of disease recurrence, rising serum PSA, or adjuvant therapy are initiated, treatment failure is recorded. However, these charts have several restrictions. The most significant limitation is that even the best of these diagrams is only in the middle between a model with perfect identification (harmony index = 1.0) and a model without identification (harmony index = 0.5). Those are slightly better. Furthermore, about 30% of patients with chart-based predictions in the middle range (7-year non-progressive survival, 30-70%) are uncertain, and the accuracy of the prediction does not exceed the reason of copper plate throwing. Computer-implemented technologies for image processing and analysis have emerged to provide significantly increased computing power. In many applications, the ability to automatically extract a large number of quantitatively continuous features from a single image has been realized. A characteristic X, then, for some ACBs, when the number system of this characteristic includes all numbers X between A and B 'is called a system continuous number generator. Cancer image analysis systems have been developed for images taken from cytological samples [2] [3] 200538734 (4). But ‘this system only captures all the structural information observed by the cell, not to mention sub-information. For the element structure that is often used in diagnosis (more important than the appearance of individual cells), 尙 did not propose the use of artificial techniques to analyze the pathological grade of precancerous cancer pedigree [4]. Conventional cancer imaging 典型 typical images The fact that ground is more complex and knowledgeable than cell imaging is exacerbated. Based on the foregoing, there is a need to provide interruptions and predict the occurrence of medical conditions, responses and methods. There is also a need to provide a computer implementation to use tissue layers to predict the occurrence of medical conditions] [ SUMMARY OF THE INVENTION A specific example of the present invention is to provide a prediction system and method. As used herein, including 'for example, predicting whether a patient will be affected (eg, relapse), and predicting (eg,' new pharmaceutical drug ') It is possible that an appropriate medical condition occurs. The present invention is used by a doctor or other individual, for example, to diagnose or diagnose a patient's medical condition. In one aspect of the present invention, the information is proposed and cannot be used at the organizational level. Combination of clinical and sub-portal (in Gleason analysis) tissue-level different pathological imaging analysis systems. Therefore, the shape and size of the diseased glands determine the lack of an analysis system. The organization needs to understand the broad field of professionally improved predictive power to treat, diagnose, and diagnose other medical phenomena. method. Automated medical condition occurrence predictions can include / or when a patient will experience a disease such as a cancer patient responding to one or more therapies, and predicting any other specific implementation of the prediction can allow the patient to choose the appropriate course of treatment and / System and method for generating a model for predicting medical conditions-200538734 (5). The generation of a prediction model may include using an analysis tool to train a support vector machine (SVM) or neural network with data from a group of patients whose results are at least partially known. In a specific example, the training data includes clinical data, molecular data, and computer-generated morphological data. As used herein, a particular type of data (e.g., a 'clinical' molecule, or morphology) may include one or more characteristics of that type. In addition, 'morphological data is defined as including any computer-generated data associated with or derived from electronic (digital) images of tissues, including but not limited to the structural properties (e.g., area, length, 'Width' compactness, and density), the spectral properties of the tissue or part of it (such as red, green, blue (RGB) color channels, brightness, and piping organization maps), and tissue images and / or identified tissue parts Non-integer properties (such as the non-integer dimension of the intraepithelial interface, cavity profile), wavelet decomposition coefficients, and / or other statistical properties of image data conversion. In other specific examples, the training data includes only computer-generated morphological data or a combination of clinical data and computer-generated morphological data. In a specific example, the system and method are provided to interact with one or more of stroma, cytoplasm, epithelial nucleus, stromal nucleus, lumen 'red blood cells, tissue correction crops, or tissue background, or a combination thereof. The associated computer-generated morphological properties produce a predictive model. The prediction model may be generated based on computer-generated morphological characteristics or a combination with one or more clinical characteristics listed in Table 4 and / or one or more molecular characteristics listed in Table 6. For example, the one or more characteristics can be input into an analysis tool to determine an image of the characteristics' ability to predict the medical condition of the relevant model. -9- 200538734 (6) Features that increase the predictive power of the model can be included in the final model, while features that do not increase (eg, or reduce) the predictive power are removed from consideration. Using the above morphological characteristics individually or in combination with the clinical and / or morphological characteristics listed in Table 4 and / or Table 6 as a basis for developing a predictive model, doctors, other individuals, and / or automated processing equipment (eg, Organizational image analysis systems) and other resources focus on obtaining patient characteristics data that are more likely to be correlated with results and can be used in the final prediction model. # In another aspect of the present invention, a prediction model is provided for evaluating a patient's data set to assess the risk of the patient's medical condition. Here, the prediction model is based on computer-generated morphological data or combined with clinical data. And / or molecular data. For example, the predictive model may receive a patient profile as an input, and may output a "score" indicating the likelihood that the patient will experience one or more results associated with the medical condition. In a specific example, a predictive model is provided to predict the occurrence or recurrence of a disease. 'Here the model is based on one or more associations with the stroma, cytoplasm, epithelium, stromal nucleus, cavity, red blood cells, tissue crops, or Computer-generated morphological characteristics of organizational background, or a combination of them. The prediction model may be based on these computer-generated morphological characteristics alone or in combination with one or more clinical characteristics listed in Table 4 and / or with one or more molecular characteristics listed in Table 6. In another specific example, a predictive model is provided for predicting prostate cancer recurrence, where the model is based on one or more of the clinical and / or molecular characteristics listed in FIG. 6 and one or more of the following pathological objects: Or more morphological characteristics · red blood cells, epithelial nucleus, stroma, cavity, cytoplasm, and group-10-200538734 (7) weaving background. In yet another specific example, a predictive model is provided for cancer recurrence, where the model is based on one or / or molecular characteristics listed in FIG. 9 and morphological characteristics for one or more of the following pathological items: red blood cells , Epithelial nucleus, stroma, cavity, in another specific example, a predictive model is provided to pre-live, here the model is based on a bed and / or molecular characteristics listed in FIG. More morphological characteristics of the following pathologies: red blood cells, epithelial nuclei, and stroma. In another specific example, the prediction model can determine whether a patient is normal or abnormal or predict whether a patient is likely to undergo postoperative surgery. Clinical failure. In another aspect, systems and methods are provided in which patient data is measured at each point and a diagnosis or treatment performed using the predictive model of the present invention can be based on the ratio of the results of each assessment. A comparison can be described in, for example, a computer The report output is for personal use. For example, systems and methods can be provided to screen an inhibitor compound. The patient's first data set can be determined using a predictor, where the model is based on clinical data, molecular data, and morphological data. After a test compound can be administered to a patient, a second data set can be obtained from the patient and evaluated. The evaluation results of the first data set can then be compared with the results of the second set. Compared with the first data set, the results of the second data set cannot indicate that the test compound is an inhibitor compound. Predict the prostate, multiple clinical and one or more and cytoplasm. One or 〇 tissue samples for prostate cancer or more objects are evaluated at many points in the prostatectomy. Patients vary. This model of a doctor or other medical condition is evaluated and computer generated. Changes in the evaluation group of the prediction model data group during the investment test -11-

200538734 (8) 於本發明又另一方面中，提供一檢驗套組用上 診斷及/或預測醫療狀況的發生。此等檢驗套組5 院，其他醫療設施、或任何其他適當場所。該檢馬 接收患者數據（例如，包括臨床數據、分子數據 電腦產生的形態學數據），對預測模型比較患者白 例如，在檢驗套組的記憶體內編程者）及輸出比車 於某些具體實例中，該分子數據及/或電腦產生的 數據可在接收患者的組織樣品後以分析措施產生。 數據可經由將組織樣品的電子影響分割成一或更多 將該一或更多物件分類或一或更多物件類別（如， 腔、紅血球、等）及對該一或更多物件類別進行-項測量而產生。於某些具體實例中，該檢驗套組p 輸入用以將，例如，更新資料接收到預測模型。於 體實例中，該檢驗套組可包括一輸出用以，例如， 料，例如可用於患者帳目及/或使用追踪之資料， 裝置或場所。 【實施方式】 較佳具體實例之詳細說明 本發明具體實例係有關在一預測模型中使用單 腦產生之形態學資訊或組合著臨床資訊及/或分子 預測醫療狀況發生之方法和系統。例如，於本發明 實例中，係使用臨床、分子和電腦產生形態學資訪 ***癌的發。於其他具體實例中，本發明中提供 治療、 放在醫 套組可 、及/或 數據（ 結果。 形態學 形態學 物件， 基質、 或更多 包括一 某些具 傳遞資 到另一 獨的電 資訊以 一具體 來預測 的請述 -12- 200538734 (9) 係用以預測其他醫療狀況，例如，其他類型的疾病（如， 上皮和混合贅瘤，包括***、結腸、肺、膀胱、肝、胰、 腎細胞、和柔軟組織）之發生，與患者對一或多種療法（ 如藥學藥物）的回應性或無回應性。此等預測可以由醫生 或其他個人用來，例如，選擇對一患者的恰當療程及/或 診斷患者的醫療狀況。 於本發明一方面中，可提供一分析工具包括支援向量 • 機（SVM )及/或神經網路以決定出在臨床、分子、和電 腦產生形態學特性與醫療狀況之間的關係。相關聯的特性 可形成一模型，其可用來預測該狀況的發生或復發。例如 ，可以使用分析工具來根據一群其針對一醫療狀況的結果 (例如，癌復發時間）至少部份已知之患者所得資料產生 一預測模型。然後可以使用該模型來評估一新患者的資料 以預測該醫療狀況於該新患者之發生。於某些具體實例中 ，該分析工具可以使用只爲該三資料型的亞組（如，只含 Β 臨床和形態學資料者）來產生預測模型。 本發明具體實例所用的臨床、分子及/或形態學資料 可包括與一醫療狀況的診斷、治療及/或預測相關聯的任 何臨床、分子、及/或形態學資料。針對與***癌復發 和存活相關聯分析過以產生預測模型之特性都在下面與， 例如，表1、2、4及/或6相關聯地說明過。要了解者， 此等特性中至少有某些（例如，上皮與混合贅瘤）可提供 一基礎用以發展用於其他醫療狀況（例如，***、結腸、 肺、膀胱、肝、胰、腎細胞、和柔軟組織）之預測模型。 -13- 200538734 (10) 例如，一或多項表1、2、4及/或6中的特性可用 某種其他醫療狀況的患者進行評估後，輸入到分析 測定出該等特性是否與該醫療狀況相關聯。可增加 測醫療狀況發生的能力之特性即可包括在最後模型 不會增加（例如，會減低）模型預測力的特性即可 中移除。使用表1、2、4及/或6中的特性作爲發 預測模型所用基礎可將醫生、其他個體、及/自動 設備（如，組織影像分析系統）等資源焦注在取得 與結果相關聯因而可用於最後預測模型中的患者資 者，圖6、9和1 1中顯示出經定出與***癌復發 相關聯之特性。要了解者，此等特性可直接包括在 列腺癌復發及/或存活所用的最後模型中，及/或用 出用於其他醫療狀況的預測模型。 形態學資料可包括指示出，例如，組織樣品的 造及/或光譜性質之電腦產生資料。於一具體實例 態學資料可包括用於基質、細胞質、上皮核、基質 、紅血球、組織橋作物、組織背景、或彼等的組合 學特性之資料。於本發明一方面中，提供一種組織 析系統以從組織影像獲得形態學特性之測量値。此 可爲使用 the Definiens Cellenger 軟體之 MAGICtm 此一系統可接收H&E染色影像作爲輸入，且可輸 像中的病理學物件之各種形態學特性測量値。有關 像得到態學特性所用系統和方法之額外細節要在下 圖3中予以說明。 來對有 工具以 模型預 中，而 從考慮 展出一 化處理 更可能 料。再 和存活 預測前 於發展 各種構 中，形 核、腔 等形態 影像分 等系統 系統。 出對影 從一影 面有關 -14- 200538734 (11) 臨床特性可包括或根據一或更多患者的資料例如年齡 、種族、體重、醫療史、基因型與疾病狀態，此處的疾病 狀態係指臨床和病理學階段特性及對該疾病程序現有特定 地收集到之任何其他臨床特性。通常，臨床資料係由醫生 在檢驗患者及/或患者的組織或細胞之過程中收集到者， 臨床資料也可包括對一特別醫療範疇更具特異性的臨床資 料。例如，於***癌範疇中，該臨床資料可包括指示出 ί ***特異抗原（PSA )的血液中濃度之資料、數位直腸 檢驗結果、Gleason計分、及/或可能對***癌更具特異 性之其他臨床資料。通常，當表1、2、4及/或6及/或圖 6、9及/或1 1中的任何特性（亦即，臨床、形態學及/或 分子）經應用於***外的醫療範疇之時，可能不考慮對 ***更具特異性的來自此等表及/或圖之特性。視情況 地，可用對目標醫療範疇更具特異性的特性取代***特 異性特性。例如，其他的組織學疾病-特異性特性/徵象 I 可包括壞死區（例如，***原位（in situ)管癌），上皮 細胞（如，***、肺）的尺寸，形狀和區域樣式/分布、 分化程度（例如，非小細胞肺癌（N S C L C，在***和結腸 兩者中可看到的各種腺癌所看到的黏液蛋白（mucin )產 生）的鱗狀分化），細胞的形態學/顯微鏡分布（例如， 乳癌中的襯裡管，NSCLC中的襯裡細支氣管）、以及發炎 程度和類型（例如，***和NSCLC在與***比較之下 各具不同的特性）。 分子特性可包括或根據可指示出生物分子包括核酸、 -15- 200538734 (12) 多肽、醣類、類固醇類和其他小分子或上述的組合，例如 糖蛋白和蛋白質一 RNA複合物，之存在、不存在、相對增 加或減少或相對位置之資料。測量此等分子之位置可包括 腺體、腫瘤、基質、及/或其他位置，且可能決定於特別 的醫療範疇。通常，分子資料係使用常用的分子生物學和 生物化學技術包括南方氏、西方氏和北方式點漬、聚合酶 鏈型反應（PCR )、免疫組織化學、和免疫螢光等。此外 ,，可以使用原位雜交（in situ hybridization)來顯示出分 子生物學特性的豐盛度和位置。組織原位雜交所用的範例 方法和系統都載於上引2003年七月21日提出申請，各稱 爲” Methods and compositions for the preparation and use of fixed-treated cell-lines and tissue in fluorescence in situ hybridization”的美國專利申請第10/624,233號之中。 圖1 A和1 B顯示出使用預測模型來預測患者醫療狀況 發生之範例系統。當，例如，一醫療診斷實驗室對與一遠 ί 程存取裝置相結合的醫生或其他個體提供一醫療決定的支 援時，可以使用圖1Α中的安排。當，例如，要提供包括 預測模型的檢驗套組給一設施例如醫院、其他醫療設施、 或其他適當場所使用時，可以使用圖1 B中的安排。 參考圖1A，預測模型102係座落於診斷設施104之 中。預測模型可包括任何適當的硬體、軟體、或彼等的組 合用以接收患者資料，評估該資料以對該患者預測一醫療 狀況的發生（例如，復發）、及輸出評估結果。於另一具 體實例中，可以使用模型1 02來預測患者對特別的一或多 -16- 200538734 (13) 種療法之回應性。診斷設施1 0 4可從遠程存取裝置1 0 6透 過網際網路服務提供者（ISP ) 108和通信網路1 10和1 12 接收患者資料，且可將該資料輸入到預測模型1 02供評估 所用。從遠程位置接收和評估患者資料所用的其他安排當 然也是可能者（例如，通過另一連接例如電話線或透過物 理郵遞）。位於遠處的醫生或個人可用任何適當方式取得 患者資料且可使用遠程存取裝置1 06將資料傳遞給診斷設 施1 〇4。於某些具體實例中，患者資料可至少部份由診斷 設施1 04或另一設施所產生。例如，診斷設施〗〇4可從遠 程存取裝置106或其他裝置接收數位化版的H&E染色影 像且可根據該影像產生該患者的形態學資料。於另一例子 中’可由診斷設施1 04接收及處理實際的組織樣品以產生 形態學資料。於其他例子中，可由第三者接收新患者的影 像或組織，根據該影像或組織產生形態學資料、及提供該 形態學資料給診斷設施1 04。要從組織影像及/或樣品產生 形態學資料所用的適當影像處理工具要在下面與圖3相關 處說明之。 診斷設施1 0 4可透過，例如，通過I S P 1 0 8和通信網 路1 1 0和1 1 2或另一種方式例如物理郵遞或電話通知而提 供評估結果給與遠程存取裝置1 〇 6相結合的醫生或個人。 該結果可包括診斷”計分”（例如，患者會經歷一或多種與 醫療狀況相關聯的結果之可能性的指示，例如事件復發的 預測時間），指示出預測模型1 02所分析的一或更多特性 係與醫療狀況相關聯者之資訊，指示出預測模型的敏感度 -17- 200538734 (14) 及/或特異性之資訊、或其他的適當診斷資訊或彼等的組 合。例如’圖2顯示出可由預測模型輸出的一虛構患者的 報告之例子。如所示者，該報告係將患者的結果機率（例 如’則列腺癌復發；亦即，y -軸）對時間（單位月）（X —軸）作圖。於此實施例中，患者具有” 5 2 0 ”之計分，此 將該患者置於高風險類項之中。此種報告可由醫生或其他 個體用來幫助決定更精細的臨床-診斷腫瘤等級，發展出 B —有效的手段來小分類患者及對個別患者最後產生更正確 的（且恰當的）治療選項規則系統。該報告也可用來幫助 醫生或個人對患者解釋患者風險。 遠程存取裝置1 06可爲能夠傳遞資訊給診斷設施1 〇4 及/或自診斷設施1 0 4接收資料之任何遠程裝置例如，個 人電腦、無線裝置例如膝上型電腦、手機、或個人數位輔 助器（PDA )、或任何他種適當遠程存取裝置。於圖1A 的系統中可包括多個遠程存取裝置1 〇 6 (例如，使在相應 B 的多個遠程位置之多個醫生或其他個體能與診斷設施1: 4 溝通資料），不過在圖1A中只包括一個遠程存取裝置 106以避免圖式過於複雜。診斷設施1〇4可包括一能夠接 收且處理得自及/或發到遠程存取裝置1 06的通訊之伺服 器。此種伺服器可包括一離散的計算硬體及/或儲存組件 ，不過也可以爲一種軟體應用或爲硬體與軟體的組合。該 伺服器可以使用一或更多個電腦來執行。 每一通訊聯結1 1 〇和1 1 2可爲任何適當的有線或無線 通信路徑或諸通路之組合例如，區域性網路、寬域網路、 •18- 200538734 (15) 電話網路、有線電視網路、網內網路、或網際網路。某些 適當的無線通訊網路可爲全球之移動通信系統（GSM )網 路、時分多重存取（TDMA )網路、碼分多重存取（ COMA)網路、藍芽網路、或任何他種適當的無線網路。 圖1 B顯示提供內含本發明預測模型的檢驗套組1 22 給設施124使用之系統，該設施124可爲一醫院、醫生診 所、或其他適當場所。檢驗套組1 22可包括任何適當的硬 體、軟體、或彼等的組合（如，個人電腦），彼等皆係調 適成可接收患者資訊（如，臨床、形態學和分子資訊中至 少一者），使用預測模型評定該患者資料（如，在檢驗套 組的記憶體內編程者）、及輸出評定結果。例如，檢驗套 組1 2 2可包括一電腦可讀式媒體，其中編碼著電腦可執行 指令以實施預測模型的功能。該預測模型可爲事先產生的 預定模型（例如，由另一系統或應用例如圖1 C中的系統 產生者）。於某些具體實例中，檢驗套組1 2 2可視需要包 括一影像處理工具，其能夠產生對應於來自，例如，一組 織樣品或影像’的形態學特性之資訊。一'種適當的影像處 理工具爲下面關於圖3中所說明者。於其他具體實例中， 檢驗套組1 22可接收呈輸入形式，來自，例如，輸入裝置 (如，鍵盤）或另一裝置或場所，之形態學特性預包裝資 料。檢驗套組1 2 2可視情況包括一輸入以接收，例如更新 資料到預測模型。該檢驗套組也可視需要包括一輸出用以 傳遞數據，例如患者帳單及/或使用追踪有用的資料，到 主設施或其他適當裝置或場所。帳單資料可包括，例如， -19- 200538734 (16) 由檢驗套組評定的患者醫療保險資訊（例如，姓名、保險 提供者、和帳戶號碼）。此種資訊可用於，例如，檢驗套 組提供者以每用基礎索取套組費用之時及/或提供者需要 患者的保險資訊以對保險提供者提出請求之時。 圖1c顯示出產生預測模型用之範例系統。該系統包 括分析工具132 (例如，包括一支援向量機（SVM )及/或 神經網路）和其結果至少部份已知的患者之數據庫134。 φ 分析工具1 3 2可包括任何適當的硬體、軟體、或彼等的組 合用以定出在來的數據庫134的資料與醫療狀況之前的相 互關係。圖1 C中的系統也可包括影像處理工具1 3 6，其 能夠根據，例如經數位化版的H&E染色組織影像，實際 的組織樣品、或兩者，產生形態學資料。工具1 3 6能針對 其數據已包括在數據庫1 3 4內的已知患者產生形態學資料 。一種適當的影像處理工具1 3 6要在下面相關於圖3處予 以說明。 Φ 數據庫1 3 4可包括任何適當的患者資料，例如臨床特 性、形態學特性、分子特性、或彼等的組合所用的資料。 數據庫1 3 4也可包括指示出患者結果之資料，例如該患者 是否及何時會經歷疾病復發。例如，數據庫1 3 4可包括患 者的檢資料（亦即，其結果完全已知的患者資料）例如業 已經歷醫療狀況復發的患者之資料。數據庫1 3 4可替代地 或加添地包括患者的已檢資料（亦即，其結果未完全知悉 的患者之資料）例如一或更多次到醫生處的追踪看診中沒 有顯不疾病復發跡象的患者之資料。分析工具1 3 2對已檢 -20. 200538734 (17) 資料的使用可增加可取用以產生預測模型的數據量。因而 ，可以有利地改良模型的可靠性和預測力。支援向量機（ S VM )和神經網路（NNci )中可以使用已檢和未檢兩種資 料者之例子會在下面述及。 於一具體實例中，分析工具1 3 2可包括一支援向量機 (SVM )。於此一具體實例中，工具132較佳地包括一能 夠對已檢資料實施支援向量回歸的SVM ( SVRc )。如在 φ 申請中的美國專利申請第1〇/991，240號中所述者，於 SVRc中，提供一新穎修改過的損失/補償函數以用在SVM 內而SVM可利用已檢資料。包括來自數據庫134的已知 患者之臨床、分子及/或形態學特性在內之資料可輸入到 SVM以定出預測模型所用參數。該等參數可指示出輸入特 性的相對重要性，且可經調整以使SVM預測已知患者的 結果之能力最大化。有關SVM對於定出特性與醫療狀況 的相互關聯之用途的額外細節經載於[5]和[6]中。 # 分析工具132對SVRc的使用可包括從數據庫134獲 得指示患者狀態的資訊之多維-非線性向量，此處該等向 量中至少一者缺乏針對相應患者的事件發生時間之指示。 之後，分析工具1 3 2可使用該等向量實施回歸產生以核爲 基礎的模型（Kernel-based model)，以根據資訊向量中 所包含的至少某些資訊提供事件時間預測相關的輸出値。 分析工具1 3 2可以使用損失函數用於每一含有已檢資料的 向量，該函數不同於工具1 3 2對包括未檢資料的向量所用 之損失函數。已檢資料樣品可用不同方式處置，因爲其只 -21 - 200538734 (18) 可提供”單側資訊’’（one-sided information )之故。例如， 在存活時間預測的情況中，已檢資料樣品典型地只指示出 該事件在一所給時間內沒有發生，且沒有指明在該所給時 間後，若有時/何時會發生。 分析工具1 3 2對已檢資料所用的損失函數可爲下面所 示者200538734 (8) In yet another aspect of the present invention, a test kit is provided for diagnosing and / or predicting the occurrence of a medical condition. These inspection suites are 5 hospitals, other medical facilities, or any other appropriate location. The checker receives patient data (for example, including clinical data, molecular data and computer-generated morphological data), compares the patient model with the prediction model (for example, programmers in the memory of the test suite), and outputs comparisons to certain specific examples. In this case, the molecular data and / or computer-generated data can be generated by analyzing the patient's tissue samples. Data may be obtained by segmenting the electronic influence of a tissue sample into one or more categories of the one or more objects or one or more object categories (e.g., cavity, red blood cells, etc.) and performing an item on the one or more object categories Resulting from measurement. In some specific examples, the test set p input is used to receive, for example, updated data into a predictive model. In a specific example, the test kit may include an output for, for example, materials such as data, devices, or locations that may be used for patient accounts and / or usage tracking. [Embodiment] Detailed description of preferred specific examples The specific examples of the present invention are related to a method and system for predicting the occurrence of a medical condition by using morphological information generated by a single brain or combining clinical information and / or molecules in a prediction model. For example, in the examples of the present invention, clinical, molecular, and computer-generated morphological resources were used to interview the development of prostate cancer. In other specific examples, the present invention provides treatments, medical kits, and / or data (results. Morphology, morphology, matrix, or more including one or more devices that pass information to another unique electrical source. The information is described with a specific prediction-12- 200538734 (9) is used to predict other medical conditions, for example, other types of diseases (such as epithelial and mixed neoplasms, including breast, colon, lung, bladder, liver, Pancreas, kidney cells, and soft tissue), and patients' responsiveness or non-response to one or more therapies (such as pharmaceutical drugs). These predictions can be used by doctors or other individuals, for example, to choose a patient The appropriate course of treatment and / or diagnosis of the patient's medical condition. In one aspect of the present invention, an analysis tool may be provided including support vector machines (SVM) and / or neural networks to determine clinical, molecular, and computer-generated morphology The relationship between the medical characteristics and the medical condition. The associated characteristics can form a model that can be used to predict the occurrence or recurrence of the condition. For example, analytical tools can be used A prediction model is generated based on data obtained from a group of patients whose outcomes for a medical condition (eg, cancer recurrence time) are at least partially known. The model can then be used to evaluate a new patient's data to predict the medical condition for the new patient. In some specific examples, the analysis tool can use only a subset of the three data types (for example, those with only clinical and morphological data B) to generate a prediction model. The clinical, Molecular and / or morphological data may include any clinical, molecular, and / or morphological data associated with the diagnosis, treatment, and / or prediction of a medical condition. Analysis was performed to generate predictions for associations with prostate cancer recurrence and survival The characteristics of the model are described below in association with, for example, Tables 1, 2, 4, and / or 6. It should be understood that at least some of these characteristics (for example, epithelial and mixed neoplasms) can provide a The foundation is used to develop predictive models for other medical conditions such as breast, colon, lung, bladder, liver, pancreas, kidney cells, and soft tissue. -13- 200538734 (10) For example, one or more of the characteristics in Tables 1, 2, 4, and / or 6 can be evaluated for patients with some other medical condition, and then entered into the analysis to determine whether the characteristics are associated with the medical condition. Features that increase the ability to measure the occurrence of a medical condition can be included in features that do not increase (eg, decrease) the model's predictive power in the final model. Use the features in Tables 1, 2, 4, and / or 6 As the basis for developing the prediction model, resources such as doctors, other individuals, and / or automated equipment (such as tissue imaging analysis systems) can be focused on obtaining patient funding that is correlated with the results and can be used in the final prediction model. Figure 6, The characteristics associated with prostate cancer recurrence are shown in 9 and 11. It should be understood that these characteristics can be directly included in the final model used for the recurrence and / or survival of prostate cancer, and / or used Predictive models for other medical conditions. The morphological data may include computer-generated data indicating, for example, the fabrication and / or spectral properties of a tissue sample. In a specific example, the morphological data may include data for stroma, cytoplasm, epithelium, stroma, red blood cells, tissue bridge crops, tissue background, or their combinatorial characteristics. In one aspect of the present invention, a tissue analysis system is provided to obtain a measurement of morphological characteristics from a tissue image. This can be MAGICtm using the Definiens Cellenger software. This system can receive H & E stained images as input, and can measure various morphological characteristics of pathological objects in the images. Additional details about the systems and methods used to obtain the morphological properties are illustrated in Figure 3 below. It is more likely to have a tool to predict the model, but to consider the standardization process. Before the survival and survival prediction, various morphological image classification systems such as nucleation and cavity were developed. (1) The clinical characteristics may include or based on the information of one or more patients such as age, race, weight, medical history, genotype and disease status, where the disease status is Refers to clinical and pathological stage characteristics and any other clinical characteristics currently specifically collected for the disease program. Generally, clinical data is collected by a doctor during the examination of a patient and / or patient's tissues or cells. Clinical data may also include clinical data that is more specific to a particular medical area. For example, in the category of prostate cancer, the clinical data may include information indicating the blood concentration of a prostate-specific antigen (PSA), digital rectal test results, a Gleason score, and / or may be more specific for prostate cancer. Other clinical information. Generally, when any of the characteristics (ie, clinical, morphological, and / or molecular) in Tables 1, 2, 4 and / or 6 and / or Figures 6, 9 and / or 11 are applied to the medical field outside the prostate At this time, characteristics from these tables and / or maps that are more specific to the prostate may not be considered. Optionally, prostate-specific properties can be replaced with properties that are more specific to the target medical category. For example, other histological diseases-specific characteristics / signs I may include necrotic areas (eg, breast cancer in situ), size, shape, and area pattern / distribution of epithelial cells (eg, breast, lung) Degree of differentiation (for example, squamous differentiation of non-small cell lung cancer (NSCLC, mucin produced by various adenocarcinomas seen in both breast and colon), morphology / microscopy of cells Distribution (eg, lined tubes in breast cancer, lined bronchioles in NSCLC), and degree and type of inflammation (eg, breasts and NSCLC each have different characteristics compared to the prostate). Molecular properties may include or be based on the indication that biomolecules include nucleic acids, -15-200538734 (12) polypeptides, carbohydrates, steroids, and other small molecules or combinations thereof, such as the presence of glycoproteins and protein-RNA complexes, No data exists, relative increase or decrease or relative position. The location at which these molecules are measured may include glands, tumors, stroma, and / or other locations, and may depend on the particular medical domain. In general, molecular information systems use common molecular biology and biochemical techniques including Southern, Western, and Northern spotting, polymerase chain reaction (PCR), immunohistochemistry, and immunofluorescence. In addition, in situ hybridization can be used to show the abundance and location of molecular biological characteristics. The example methods and systems used for tissue in situ hybridization are set out in the above-mentioned application filed on July 21, 2003, each referred to as "Methods and compositions for the preparation and use of fixed-treated cell-lines and tissue in fluorescence in situ hybridization" "In U.S. Patent Application No. 10 / 624,233. Figures 1 A and 1 B show example systems that use predictive models to predict the occurrence of a patient's medical condition. The arrangement in Figure 1A can be used, for example, when a medical diagnostic laboratory provides medical decision support to a doctor or other individual in combination with a remote access device. When, for example, a test kit including a predictive model is to be provided for use at a facility such as a hospital, other medical facility, or other appropriate location, the arrangement in Figure 1B may be used. Referring to FIG. 1A, a prediction model 102 is located in a diagnostic facility 104. The predictive model may include any appropriate hardware, software, or combination thereof to receive patient data, evaluate the data to predict the occurrence of a medical condition (e.g., relapse) for the patient, and output the results of the evaluation. In another specific example, model 102 can be used to predict patient response to a particular one or more -16- 200538734 (13) therapies. The diagnostic facility 1 0 4 can receive patient data from a remote access device 10 6 through an Internet Service Provider (ISP) 108 and a communication network 1 10 and 1 12 and can input this data into the predictive model 10 02 for Used for evaluation. Other arrangements for receiving and evaluating patient data from remote locations are of course possible (for example, via another connection such as a telephone line or physical mail). Remotely located doctors or individuals can obtain patient data in any suitable way and can use remote access devices 106 to pass the data to diagnostic facilities 104. In some specific examples, patient data may be generated at least in part by diagnostic facility 104 or another facility. For example, the diagnostic facility may receive a digitalized H & E stained image from the remote access device 106 or other device and may generate morphological data of the patient based on the image. In another example, 'actual tissue samples may be received and processed by diagnostic facility 104 to generate morphological data. In other examples, a third party may receive an image or tissue of a new patient, generate morphological data based on the image or tissue, and provide the morphological data to the diagnostic facility 104. Appropriate image processing tools to generate morphological data from tissue images and / or samples are described below in relation to Figure 3. The diagnostic facility 104 can provide the evaluation results to the remote access device 106 through, for example, the ISP 108 and the communication network 110 and 12 or another method such as physical mail or telephone notification. Combined doctor or individual. The results may include a diagnostic "score" (for example, an indication of the likelihood that a patient will experience one or more results associated with a medical condition, such as the predicted time of event recurrence), indicating one or Further characteristics are information related to the medical condition, indicating the sensitivity of the predictive model-17- 200538734 (14) and / or specific information, or other appropriate diagnostic information or a combination of them. For example, 'FIG. 2 shows an example of a fictitious patient report that can be output by a predictive model. As shown, the report plots the patient's outcome probability (for example, 'regression of leidenoma; that is, the y-axis) versus time (unit months) (X-axis). In this embodiment, the patient has a "520" score, which places the patient in a high-risk category. Such reports can be used by doctors or other individuals to help determine more elaborate clinical-diagnostic tumor grades and develop B-effective means to classify patients and ultimately produce more correct (and appropriate) treatment option rule systems for individual patients . The report can also be used to help doctors or individuals explain patient risks to patients. The remote access device 106 may be any remote device capable of transmitting information to the diagnostic facility 104 and / or the self-diagnostic facility 104 to receive data, such as a personal computer, a wireless device such as a laptop, a mobile phone, or a personal digital device. Assistant (PDA), or any other suitable remote access device. The system in FIG. 1A may include multiple remote access devices 1 06 (for example, to enable multiple doctors or other individuals at multiple remote locations of the corresponding B to communicate with the diagnostic facility 1: 4), but in the figure Only one remote access device 106 is included in 1A to avoid overly complicated drawings. The diagnostic facility 104 may include a server capable of receiving and processing communications from and / or to the remote access device 106. Such a server may include a discrete computing hardware and / or storage component, but may also be a software application or a combination of hardware and software. The server can be run using one or more computers. Each communication link 1 1 0 and 1 1 2 may be any suitable wired or wireless communication path or combination of paths. For example, a local area network, a wide area network, • 18- 200538734 (15) telephone network, wired TV network, intranet, or internet. Some suitable wireless communication networks may be global mobile communication system (GSM) networks, time division multiple access (TDMA) networks, code division multiple access (COMA) networks, Bluetooth networks, or any other An appropriate wireless network. Fig. 1B shows a system for providing a test set 1 22 containing a predictive model of the present invention to a facility 124, which may be a hospital, doctor's office, or other suitable location. The test kit 1 22 may include any suitable hardware, software, or combination thereof (eg, a personal computer), all of which are adapted to receive patient information (eg, at least one of clinical, morphological, and molecular information). ), Use the predictive model to evaluate the patient data (eg, programmers in the memory of the test suite), and output the evaluation results. For example, the inspection kit 1 2 2 may include a computer-readable medium in which computer-executable instructions are encoded to perform the functions of a predictive model. The predictive model may be a predetermined model generated in advance (e.g., produced by another system or application such as the system in Fig. 1C). In some specific examples, the inspection kit 1 2 may optionally include an image processing tool capable of generating information corresponding to morphological characteristics from, for example, a group of tissue samples or images'. One suitable image processing tool is described below with respect to FIG. 3. In other specific examples, the inspection kit 1 22 may receive pre-packaged data in the form of an input from, for example, an input device (eg, a keyboard) or another device or location. The test suite 1 2 2 may optionally include an input to receive, such as updating data to a predictive model. The test kit may also include an output to transfer data, such as patient billing and / or usage tracking useful information, to the main facility or other appropriate device or location as needed. Billing information may include, for example, -19- 200538734 (16) Patient medical insurance information (eg, name, insurance provider, and account number) assessed by the test suite. Such information can be used, for example, when the test suite provider is requesting the package fee on a per-use basis and / or when the provider requires patient's insurance information to make a request to the insurance provider. Figure 1c shows an example system for generating a prediction model. The system includes an analysis tool 132 (e.g., including a support vector machine (SVM) and / or a neural network) and a database 134 of patients whose results are at least partially known. The φ analysis tool 1 3 2 may include any suitable hardware, software, or a combination thereof to determine the interrelationship between the information in the incoming database 134 and the medical condition. The system in FIG. 1C may also include image processing tools 136, which can generate morphological data based on, for example, a digitalized version of H & E stained tissue images, actual tissue samples, or both. Tool 1 3 6 can generate morphological data for known patients whose data is already included in database 1 3 4. A suitable image processing tool 1 3 6 is described below in relation to FIG. 3. Φ Database 1 3 4 may include any suitable patient data, such as data for clinical characteristics, morphological characteristics, molecular characteristics, or a combination thereof. The database 134 may also include information indicating the outcome of the patient, such as whether and when the patient will experience a relapse. For example, the database 134 may include patient test data (i.e., patient data whose results are fully known), such as data for patients who have experienced a recurrence of the medical condition. The database 1 3 4 may alternatively or additionally include the patient's examination data (ie, the patient's information whose results are not fully known), such as one or more follow-up visits to the doctor without showing recurrence of the disease Signs of patient information. Analytical tools 1 3 2 Examined -20. 200538734 (17) The use of data can increase the amount of data available to generate predictive models. Thus, the reliability and predictive power of the model can be advantageously improved. Examples of those who can use both checked and unchecked data in support vector machines (SVM) and neural networks (NNci) are described below. In a specific example, the analysis tool 1 2 may include a support vector machine (SVM). In this specific example, the tool 132 preferably includes an SVM (SVRc) capable of performing support vector regression on the checked data. As described in U.S. Patent Application No. 10 / 991,240 in the φ application, a novel modified loss / compensation function is provided in the SVRc for use in the SVM and the SVM can utilize the inspected data. Information including clinical, molecular, and / or morphological characteristics of known patients from the database 134 can be input to the SVM to determine the parameters used in the prediction model. These parameters may indicate the relative importance of the input characteristics and may be adjusted to maximize the ability of the SVM to predict outcomes for known patients. Additional details on the use of SVMs to determine the inter-relationship between characteristics and medical conditions are set out in [5] and [6]. # Analysis tool 132's use of SVRc may include obtaining a multi-dimensional-non-linear vector of information indicative of the state of the patient from database 134, where at least one of these vectors lacks an indication of when the event occurred for the corresponding patient. After that, the analysis tool 1 3 2 can use these vectors to implement regression to generate a kernel-based model to predict the relevant output time based on at least some of the information contained in the information vector. The analysis tool 1 3 2 can use a loss function for each vector containing the checked data, which is different from the loss function used by the tool 1 2 2 for vectors containing unchecked data. Samples of inspected data can be handled in different ways, because it only provides “one-sided information”. For example, in the case of prediction of survival time, samples of inspected data Typically only indicates that the event did not occur within a given time, and does not indicate if / when it will happen after that given time. Analysis tool 1 3 2 The loss function used for the checked data can be as follows Shown

Ζα^(/(χ)，3；，5 = 1)— 0 CM 一 e) e>s: e<-ss 此處e = /⑻-L且 /(x) = WrO(x)+6 爲一對特性空間F的線性回歸函數。此處，W爲F中 的一向量，且Φ (X)係將輸入X對F中的向量作圖。 相異者，工具1 3 2對未檢資料所用的損失函數可爲：Zα ^ (/ (χ), 3 ;, 5 = 1) — 0 CM-e) e > s: e < -ss where e = / ⑻-L and / (x) = WrO (x) +6 is A linear regression function of a pair of characteristic spaces F. Here, W is a vector in F, and Φ (X) plots the input X against the vector in F. Differently, the loss function used by tool 1 3 2 for unchecked data can be:

Loss(f (x%y9s = 0)=2 < 〇Loss (f (x% y9s = 0) = 2 < 〇

Cn[en-e)Cn (en-e)

於上面的說明中，W和b係經由解開經最適化問題而 得者，其通用形式爲： min 1 W，b 2In the above description, W and b are obtained by solving the optimization problem. The general form is: min 1 W, b 2

WT w S.t. y. - + 6) < ε 不過，此方程式假設凸最優化問題總是可行的，此不 盡如此。再者，也需要使回歸估計中有小的誤差。基於此 -22- 200538734 (19) 等理由，乃對SVRc使用損失函數。損失可讓回歸估計有 某些選擇之餘地。理想上，所建立的模型要剛好能準確地 計算所有的結果，此爲不可行者。損失函數可容許對理想WT w S.t. y.-+ 6) < ε However, this equation assumes that the convex optimization problem is always feasible, which is not the case. Furthermore, it is necessary to make small errors in the regression estimation. For this reason, -22- 200538734 (19) etc., the loss function is used for SVRc. Losses allow some choice in regression estimation. Ideally, the model established should be able to accurately calculate all the results, which is not feasible. The loss function allows

値有一誤差範圍，此範圍係受鬆弛變數f和f +，及補償C 所控制。偏離理想値，但在f和f #所界定的範圍內之誤 差經計及，但彼等的基値（Contribution )係由（予以減 緩。情況愈誤差，補償愈大。誤差愈小（愈靠近理想値） ，補償愈小。此種隨著誤差增加補償之槪念導致一斜率， 且C即控制此斜率。雖然有多種損失函數可供使用，對於 ε -不敏性損失函數，通用方程式即轉換成爲： min P = + s.L γί-(ν/τΦ(χί) + δ)<ε + ξί ξηξ；>09 i = h-l 對於根據本發明的ε -不敏性損失函數（對已檢和未 檢資料分別應用不同的損失函數），此方程式變成： 以· γί-(\ντΦ(χί) + Β)<ε^ξί (\νΓΦ(Ή)ι.“;+( ξΡ>〇9 / = 1-/ 此處 ΘΌ(1 - w) εΡ =^ + (1-5.)^ 最優化準則對於y -値與f ( x )相差大於ε的資料點 給予補償。鬆弛變數，f和f +分別對應於此過分偏差對 正偏差和負偏差之大小。此補償機制具有兩成分，一是用 於未檢資料（亦即，未經合檢（right-censored)者）且另 -23- 200538734 (20) 一係用於已檢資料。於此，兩成分係以稱爲ε 一不敏性損 失函數之損失函數形式表出。 有關對已檢資料實施支援向量回歸（SVRc )所用系統 和方法之額外細節都載於上引2004年十一月17日提出申 請的美國專利申請第10/991，240號，與2003年十一月18 曰提出申請的美國臨時專利申請第1〇/520,939號之中。 於另一具體實例中，分析工具丨3 2可包括一神經網路 B 。於此一具體實例中，工具1 3 2較佳地包括一能夠利用已 檢資料的神經網路。此外，該神經網路較佳地係使用實質 根據和諧指數（CI )的近似（例如，導數）的目標函數來 訓練一相關的模型（NNci )。雖則CI長久以來即用爲存 活分析所用的效能指標[7]，不過CI訓練神經網路之用途 則在以前尙未經提出。在過去使用CI訓練目標函數的困 難在於CI係不可微者且不能以梯度法予以最優化。如在 2 00 5年二月25日提出申請，其申請中的美國專利申請第 , —/ —，__ 號，名稱 ’’Methods and Systems for PredictingThere is an error range, which is controlled by the relaxation variables f and f + and the compensation C. Departure from the ideal, but the errors within the range defined by f and f # are taken into account, but their basis (Contribution) is (to slow down. The more error the situation, the greater the compensation. The smaller the error (closer) (Ideally 补偿), the smaller the compensation. This idea of compensation as the error increases leads to a slope, and C controls this slope. Although there are many loss functions available, for the ε-insensitivity loss function, the general equation is Translates to: min P = + sL γί- (ν / τΦ (χί) + δ) < ε + ξί ξηξ; > 09 i = hl for the ε-insensitivity loss function according to the present invention (for the detected and Unchecked data apply different loss functions respectively), this equation becomes: γί-(\ ντΦ (χί) + Β) < ε ^ ξί (\ νΓΦ (Ή) ι. "; + (ΞΡ > 〇9 / = 1- / where ΘΌ (1-w) εP = ^ + (1-5.) ^ The optimization criterion compensates for data points where the difference between y-値 and f (x) is greater than ε. Relaxation variables, f and f + Respectively correspond to the magnitude of this excessive deviation to positive and negative deviations. This compensation mechanism has two components, one is for unchecked data (that is, without (Right-censored) and another 23- 200538734 (20) One is used for the checked data. Here, the two components are expressed in the form of a loss function called ε-insensitivity loss function. Additional details of the systems and methods used to implement Support Vector Regression (SVRc) for checked data are set out in U.S. Patent Application No. 10 / 991,240, filed November 17, 2004, and November 18, 2003 The U.S. Provisional Patent Application No. 10 / 520,939 filed in the application. In another specific example, the analysis tool 3 2 may include a neural network B. In this specific example, the tool 1 3 2 is preferred The neural network includes a neural network capable of utilizing the checked data. In addition, the neural network preferably trains a related model (NNci) using an objective function substantially based on an approximation (eg, a derivative) of the harmony index (CI). Although CI has long been used as a performance indicator for survival analysis [7], the use of CI for training neural networks has not been proposed before. The difficulty in using CI to train objective functions in the past is that CI is not differentiable and Not with gradients It is optimized as filed in the February 25, 2005, US Patent Application its first application, - / -, __ number, the name '' Methods and Systems for Predicting

Occurrence of an Event”，中所述者，此種障礙可經由使 用CI的近似作爲目標函數予以克服。 例如，當分析工具1 3 2包括一用來預測***癌復發 的神經網路之時，該神經網路可處理一組其針對***復 發的結果爲至少部份已知的患者之輸入資料以產生以輸出 。針對該神經網路的輸入所選的特別特性可透過使用上述 SVRc (例如，用分析工具132的支援向量機來執行）或使 用另一種適當的特性選擇法來選擇。工具1 3 2的誤差模組 -24- 200538734 (21) 可決定出在輸出與對應於輸入資料的合意輸出之間的誤差 (例如，在預測結果與一患者的已知結果之間的差異）。 之後，分析工具132可使用實質根據CI的近似之目標函 數來評等神經網路的效能。分析工具1 3 2可根據目標函數 的結果調整神經網路的加權連接（如，特性的相對重要性 )。有關適配神經網路的加權連接來調整特性與預測結果 的相互關係之額外細節經載於[8]和[9]之中。 # 和諧指數可表成下列形式： _ ΙΩΙ 此處 i(Uj)4l:ti>tj }5 ]1〇:其他情耐 且可根據在患者i和j分別分別所得豫後估計&與& 之間的配對比較。於此例中，係由滿足下列條件的所有 患者配對{i，j}所構成：Occurrence of an Event ", such obstacles can be overcome by using an approximation of CI as the objective function. For example, when the analysis tool 1 2 2 includes a neural network to predict the recurrence of prostate cancer, the A neural network can process a set of input data for patients whose prostate recurrence is at least partially known to produce output. The particular characteristics selected for the input of the neural network can be obtained by using the SVRc described above (for example, using Analyze by the support vector machine of the analysis tool 132) or use another appropriate feature selection method. The error module of the tool 1 3 2-24- 200538734 (21) can determine the output and the desired output corresponding to the input data Between the errors (eg, the difference between the predicted result and a patient's known result). The analysis tool 132 can then use the approximate objective function based on the CI to evaluate the performance of the neural network. Analysis Tool 1 3 2 The weighted connection of the neural network can be adjusted according to the result of the objective function (for example, the relative importance of the characteristics). The weighted connection of the adaptive neural network can be adjusted Additional details of the correlation between the characteristics and the predicted results are set out in [8] and [9]. # The harmony index can be expressed in the following form: _ ΙΩΙ where i (Uj) 4l: ti > tj} 5] 1〇 : Others are patient and can be compared based on the paired comparison between & and & obtained separately for patients i and j. In this example, it consists of all patient pairs {i, j} that meet the following conditions: :

•患者i和j兩者都歷經復發，且患者i的復發時間 ti比患者j的復發時間q較爲短；或 •只有患者i歷經復發且tj比患者j的追踪問診時間 tj較爲短。 CI的分子表經神經網路預測較早復發的患者實際上沒 有較早復發的次數。其分母爲符合預定狀況的患者配對之 總數。 通常，當CI增加時，較佳者經最大化，該模型更準 確。如此，經由較佳地將CI實質最大化，或CI的近似， -25- 200538734 (22) 可改善模型的效能。本發明一具體實例提供如下述之c i 近似：• Patients i and j both have relapsed, and the recurrence time ti of patient i is shorter than the recurrence time q of patient j; or • Only patient i has relapsed and tj is shorter than the follow-up interview time tj of patient j. The molecular table of the CI predicted by the neural network that patients with earlier relapses did not actually have earlier relapses. The denominator is the total number of patient pairs that meet the predetermined conditions. Generally, when CI increases, the better is maximized, and the model is more accurate. In this way, -25- 200538734 (22) can improve the performance of the model by maximizing CI substantially, or CI approximation. A specific example of the present invention provides a c i approximation as follows:

一〆 Λ /Ν 、一 〆 Λ / Ν,

且此處〇< r q且n> 1。邮，(·）可視爲/K•，乂·）的近似 〇 由本發明提出且經實驗地證明可達到改良結果之另一 種CI近似爲下示者：And here, < r q and n > 1. Post, (·) can be regarded as an approximation of / K •, 乂 ·) 〇 Another CI approximation proposed by the present invention and experimentally proven to achieve improved results is shown below:

乃=Σϋ 爲一正歸因子。此處每一 係由&與~之間的差 値予以加權。將c ω (或C )最小化的程序係企圖將Ω中 的每一樣品配對移動以滿足>7且因而使你…）=1。 當Ω中的配對的輸出之間的差値大於邊際r之時，此 樣品配對即停止對目標函數的貢獻。此機制可以在模型訓 練之中有效地克服資料的過度匹配（over-fitting )且使最 優化較佳地焦注於只將Ω中的更多樣品配對移動以滿足 >γ。訓練樣品的影響係根據訓練中的配對方式比較予 以適配地調整。要特別提及者，R中的正邊際7對於改良 的正規化效能係較佳者。換言之，神經網路的參數係在訓 200538734 (23) 練中經由在所有患者資料都已進入之後計算CI而調整。 然後，於使目標函數最小化及因而使CI最大化的目標下 使神經網路調整參數。如上面所用者，過度匹配通常係指 神經網路的複雜性。特定言之，若神經網路太複雜，則網 路會對”雜訊”資料起反應。過度匹配的風險在於其可能容 易地導致遠超過訓練資料的範圍之預測。 有關使用實質地根據C I的近似的目標函數來訓練神 經網路所用系統和方法之額外細節都載於上引在2005年 一月 25日提出申請，名稱爲’’Methods and Systems for Predicting Occurrence of an Event”的美國專利申請第 _ / _，_號及2004年二月27日提出申請的美國臨時專利申 請第 60/548,3 22號和 2004年六月 4日提出申請的 60/5 77,05 1 號之中。 圖3爲適當影像處理工具所具範例功能之流程圖。圖 3中的功能主要是有關組織影像的分割以分類影像中的病 理學物件（如，將物件分數成細胞質、腔、核、基質、背 景、矯作物、和紅血球）。於一例中，影像處理工具可包 括光學顯微鏡，使用 SPOT Insight QE Color Digital Camera (ΚΑΙ 2000 )以20x放大倍率捕捉組織影像且用 1 600x1 2 00像素產生影像。該影像可用TiH格式貯存或有 24位元每像素之影像。此等設備只爲範例且任何其他種適 當的影像捕捉設備都可以使用而不違離本發明的範圍。該 影像處理工具也可包括任何適當的硬體、軟體或彼等的組 合用以分割和分類所捕捉的影像中之物件’然後測量該物 -27- 200538734 (24) 件的形態學特性。於一具體實例中，該影像處理工具可包 括市售 Definiens Cellenger Developer Studio (ν·4·0) ’ 經調適以實施例如，上述各種病理學物件的分割和分類及 測量此等物件的各種形態學特性。有關 Definiens C e 11 e n g e r產品的其他細節載於[1 0 ]之中。影像處理工具可 測量諸物件的各種形態學特性，包括以光譜爲基礎的特性 (紅、綠、藍（RGB )道特性，例如平均値、標準偏差、 等），位置、尺寸、周長、形狀（不對稱性、緻密性、橢 圓配合，等）及對鄰近物件的關係（對比）。該影像處理 工具可針對影像中每一經鑑定的病理學物件的每一證例測 量此等特性且可輸出此等特性以，例如，用預測模型1 02 (圖1A )，檢驗套組122 (圖IB ) 1或分析工具132 (圖 1 C )予以評定。視需要地，該影像處理工具也可針對每一 測量的特性輸出該影像的整體統計學摘要。有關經分類的 病理學物件的形態特性之測量的其他細節都在下面與表1 和2相關處說明。下面爲圖3中所示影像處理工具的功能 之說明。 初始分割。於第一階段中，影像處理工具可將一影像 (例如H&E染色組織微陣列（TM A )影像或Η&Έ全組織 切片）分割成稱爲物件的連續像素小組。此等物件可經由 根據顏色類似性和形狀類似性找出連續區的區域成長規則 系統而獲得。物件的尺寸可經由調整一些參數而變異[1 1 ] 。於此系統中，一像素以外的物件典型地爲最小處理單位 。如此，所有形態學特性計算和操作可針對物件而實施。 -28- 200538734 (25) 例如’在對一影像施加底限時，即對物件的特性 限。其結果，物件內的所有像素都受指配到相同 於一具體實例中，可將物件的尺寸以最細層次控 1 0 — 2 0像素。根據此層次，經由從較低層次中的 形成更大物件可構建後續更高且更粗之層次。 背景提取。於初始分割之後，影像處理工具 度底限和凸包（convex hull)從背景（載片的透 割出影像組織核心。強度底限爲將影像像素分β 組織核心’’與”背景”之強度値。具有大於或等於 強度値之任何像素係經分數爲”組織核心，，像素， 素即被分類爲’’背景”像素。一幾何物件的凸包爲 件的最小凸集（convex set )。若 S內任何兩點 整個線段P Q也在S內，則集S爲凸者。 粗分割。於下一階段中，影像處理工具可 foregroumd )(如，TMA核）再分割或對應於核 的粗區。例如，H&E染色影像內的主特徵特性爲 餘病理學物件而言，彼等係經染色成藍色。所以 藍色道（R - B )強度値中的差値可用爲一項區別 別者，對於在初始分割步驟中得到的每一影像物 可以測定出在平均紅色與藍色像素強度値之間的 可以使用長度/寬度比來決定一物件是否應分類 。例如，落在（R - B )特性底限値和長/寬底限 物件即可分類爲核部位。類似者，可以使用綠色 將組織核中的物件分類爲白空間。組織基質係由 値施以底 的類別。 制到成爲 較小物件 可使用強 明區）分 g兩類：” 底限値的 否則該像 含有該物 所形成的 將前景（ 與白空間 相對於其 ，在紅和 特性。特 件而言， 差値。也 爲核部位 値之下的 道底限値 紅色所占 -29- 200538734 (26) 盡。影像物件的強度差値d，”紅色比”r = R ( R + G + B )和紅 色道標準偏差σ R可用來分類基質物件。 白空間分類。於粗分割階段中’白空間區可對應於影 像中的腔（病理學物件）與矯作物（破裂組織部位）兩者 。較小的白空間物件（面積小於1 〇〇像素者）常爲矯作物 。如此，影像處理工具可應用面積過濾器將彼等分類爲矯 作物。 # 核去·融合和分類。於粗分割階段中，所得核部位常 爲涵蓋數個實體核的連續融合區。再者，核區也可能包括 周圍的誤分類細胞質。因此，此等融合核部位可能需要核 去-融合以得到個別的核。 影像處理工具可以使用兩種不同的作法來去-融合該 核。第一種作法可根據將構成核部位的影像物件融合在形 狀約束（圓性）之下的區成長規則系統。當融合不嚴重時 ，此作法經定出可良好操作。 ^ 於嚴重融合的情況中，影像處理工具可以使用根據監 督學習的不同作法。此作法包括由一專家（病理學家）將 核部位予以人工標記出。屬於經標記的核之影像物件所具 特性可用來設計統計學分類器。 於一具體實例中，爲了減少特性空間維的數目，可以 使用兩種不同的分類器：the Bayesian分類器和k最近鄰 分類器，在訓練集上實施特性選擇[i 2 ]。可以使用漏-法（ levave-one_out method) [13]來交叉確認，且可以使用順 序前向搜尋規則系統來找出最佳特性。最後，可以設計出 -30- 200538734 (27) 特性數分別等於1與5的兩種Bayesian分類器。可以將類 別條件分布假設爲具有對角線協方差矩陣的高斯分布。 於某些具體實例中，輸入影像可包括不同類型的核： 上皮核、纖維母細胞、基礎核、內皮核、凋亡核和紅血球 由於在分等腫瘤程度中，上皮核數目典型地係經視爲一項 重要特性，所以將上皮核與他種區分出來可具重要性。影 像處理工具可經由將所偵檢的核分類成兩類別而完成此步 驟：上皮核與基於形狀（偏心率）和尺寸（面積）特性之 ’’餘者’’。 有關根據本發明的影像分割和分類之額外細節經載於 上引2004年十一月17日提出申請的美國專利申請第 10/991，847號，及2003年十一月17日提出申請的美國臨 時專利申請第60/520,815號和2004年三月12日提出申請 的第60/552,497號之中。 如上所述者，影像處理工具可在以工具分割和分類影 像中的物件之後測量各種形態學特性。此等形態學特性可 指示出一或更多性質及/或統計。該物件性質可包括光譜 性質（如，色道平均値，標準偏差和亮度）及構造/形狀 性質（如，面積、長度、寬度、緻密度、密度）。統計可 包括最小値、最大値、平均値和標準偏差且可對影像物件 的每一性質計算出。表1和2 (如附）顯示出可根據本發 明測量的各種形態學特性例子。此等表中的形態學特性係 使用可指示出此等特性所測量的各種性質及/或統計値之 標誌予以命名。表1和2中所示的特別命名標誌係從上述 -31 - 200538734 (28) 市售Definiens軟體產品調適而得，因而係諳於此技著所 知悉者。 要了解者，表1和2中所示的電腦產生之形態學特性 只爲範例且任何電腦產生的形態學特性即可以利用而不違 離本發明範圍。例如，表1和2包括不同的形態學性組合 。表2中縮減且修改過的特性組（亦即，與表1的特性相 比之下爲經縮減且修改過者）係在***癌復發和存活時 間領域中從實施表1的硏究之額外實驗所得者。特別者， 該額外實驗提供有關可以與結果更可能相關聯之特性類別 的額外洞察。本案發明人預期持續的實驗及/或其他合適 硬體、軟體、或彼等的組合之使用可產生可與此等和其他 醫療狀況相關聯之各種其他電腦產生特性組（如，表2中 的特性亞組）。 參看表1和2，特性”腔· StdDevAreaPxl”，”腔’’指出 一影像物件類型，”StdDev”表要使用經鑑定的腔的所有證 例計算出之統計値（標準偏差），且”AreaPxl”指要以該統 計値評定的物件證例之特性（許多像素之面積）。影像處 理工具可對先前在影像中分割及分類出的所有物件測量形 態學特性。例如，影像處理工具可對包括”背景”、”細胞 質”上皮核”腔”、"基質”、”基質核”和”紅血球”之物 件測量形態學特性。”背景”包括未被組織佔據的數位影像 部份。’’細胞質”係指一細胞的細胞質，其可爲一不定形部 位（如’在一例如H&E染色組織的影像中包圍上皮核的 粉紅色部位）。’’上皮核”指的是存在於腺體單位的上皮細 -32- 200538734 (29) 胞/腔和基礎細胞內之核，其顯現爲被細胞質包圍的’’圓形” 物件。”腔”指的中央腺體空間，於該處由上皮細胞沈積出 分泌物，顯現爲被上皮細胞包圍的密閉白色部位。有時候 ，腔可能被***液（此於H&E染色組織內典型地呈粉 紅色）或”碎屑”（如，巨噬細胞、死細胞、等）所塡充。 腔與上皮細胞質和該一起形成一腺體單位。”基質”指的是 維持***組織的體系結構，具有不同密度之結締組織形 • 式。基質組織係存在於諸腺體單位之間，且在H&E染色 組織中呈現爲紅至粉紅色。”基質核”爲伸長型細胞，不具 或具最少量的細胞質者（纖維母細胞）。此項類也可以包 括內皮細胞和發炎細胞，且若有癌症時，上皮核也經發現 散布在基質內。’’紅血球’’爲小的紅色圓形物體，常位於脈 管（動脈或靜脈），但也可能分散在整個組織之內。 下面表中的’’C2EN”爲核面積對細胞質的相對比例。上 皮細胞更贅生性/惡質性，核所佔據的面積愈多。’’EN2SN” B 爲數位組織影像中所含上皮細胞對基質細胞之百分比或相 對量。”L2Core"爲組織中所含腔的數目或面積。Gleason 等級愈高，癌更具侵入性，因而腔含量愈少。通常，此係 因爲癌症發生時上皮細胞係以未受控制方式複製，造成腔 被上皮細胞所塡充之故。’’C2L”爲細胞質對腔的相對量。 ” C EN 2 L ”爲細胞質上皮細胞對腔的相對量。 於本發明一方面中，提供系統和方法用以篩選一醫療 狀況（如疾病）的抑制劑化合物。圖4爲根據本發明一具 體實例篩選一抑制劑化合物中包括的範例階段流程圖。於 -33- 200538734 (30) 階段4 0 2處，可得到患者的第一資料組，其中包括一或多 項的臨床資料，形態學資料和分子資料。於階段404給患 者投予一試驗化合物。階段404之後，在階段406從患者 得到第二資料組。該第二資料組可或可不包括在第一資料 組中包括的相同資料類型（亦即，特性）。於階段408， 可將第二資料組與第一資料組相比較，此處在投予試驗化 合物後的第二資料組中之改變指出該試驗化合物爲一種抑 • 制劑化合物。比較資料組的階段408可包括，例如，本發 明預測模型所產生的對第一資料組的輸入有回應性之輸出 與由該預測模型所產生對該第二資料組的輸入有回應性之 輸出相比較。例如，該抑制劑化合物可爲一所給藥物且本 發明可定出該藥物對一醫療狀況是否爲一種有效的醫療治 療。 至此要說明本發明預測醫療狀況的諸具體實例之各種 範例應用。於第一例中，本發明具體實例係使用臨床和形 B 態學資料來預測***癌的復發。於第二實施例中，本發 明具體實例係使用臨床，形態學和分子資料來預測*** 癌復發及整體存活率。於第三實施例中，係使用本發明具 體實例來預測患者在***切斷術後侵害性疾病的發生。 於第四實施例中，使用本發明具體例來預測肝毒物學。 ***癌總覽乃 = Σϋ is a positive return factor. Each series here is weighted by the difference between & and ~. The procedure to minimize c ω (or C) is an attempt to pair move each sample in Ω to satisfy > 7 and thus make you ...) = 1. When the difference between the outputs of the pairings in Ω is greater than the marginal r, this sample pairing stops contributing to the objective function. This mechanism can effectively overcome the over-fitting of the data during model training and make the optimization better focus on only pairing and moving more samples in Ω to satisfy > γ. The influence of the training samples is adjusted according to the comparison of the pairing methods in training. In particular, the positive margin 7 in R is better for improved regularization performance. In other words, the parameters of the neural network are adjusted in training 200538734 (23) by calculating the CI after all patient data has been entered. Then, the neural network adjusts the parameters with the goal of minimizing the objective function and thus maximizing the CI. As used above, overmatching usually refers to the complexity of the neural network. In particular, if the neural network is too complex, the network will react to "noise" data. The risk of overmatching is that it can easily lead to predictions well beyond the scope of the training data. Additional details on the use of systems and methods for training neural networks using substantially objective functions based on CI are set out in the above application filed January 25, 2005, entitled `` Methods and Systems for Predicting Occurrence of an Event "US Patent Application No. _ / _, _ and US Provisional Patent Application No. 60 / 548,3 22 filed on February 27, 2004 and 60/5 77, filed on June 4, 2004, 05 in No. 1. Figure 3 is a flowchart of example functions of appropriate image processing tools. The functions in Figure 3 are mainly related to the segmentation of tissue images to classify pathological objects in the images (for example, to score objects into cytoplasm, (Lumen, nucleus, matrix, background, orthopedics, and red blood cells). In one example, image processing tools may include an optical microscope, using a SPOT Insight QE Color Digital Camera (ΚΑΙ 2000) to capture tissue images at 20x magnification and 1 600x1 2 00 pixels to produce images. The images can be stored in TiH format or have 24-bit per pixel images. These devices are just examples and any other suitable image capture Can be used without departing from the scope of the present invention. The image processing tool may also include any appropriate hardware, software, or a combination thereof to segment and classify objects in the captured image 'and then measure the object- 27- 200538734 (24) morphological characteristics of the piece. In a specific example, the image processing tool may include a commercially available Definiens Cellenger Developer Studio (ν · 4.0.0) 'adapted to implement, for example, the above-mentioned various pathological objects. Segment and classify and measure various morphological characteristics of these objects. Additional details about Definiens C e 11 enger products are contained in [1 0]. Image processing tools can measure various morphological characteristics of objects, including the use of spectra as Basic characteristics (red, green, blue (RGB) channel characteristics, such as average chirp, standard deviation, etc.), position, size, perimeter, shape (asymmetry, compactness, ellipse fit, etc.) and proximity to objects (Contrast). This image processing tool can measure and output these characteristics for each evidence of each identified pathological object in the image For example, the prediction model 10 02 (Fig. 1A), the test set 122 (Fig. IB) 1 or the analysis tool 132 (Fig. 1 C) are evaluated. If necessary, the image processing tool can also be used for each measured characteristic An overall statistical summary of the image is output. Additional details regarding the measurement of the morphological characteristics of the classified pathological objects are described below in relation to Tables 1 and 2. The following is a description of the functions of the image processing tool shown in FIG. Initial segmentation. In the first stage, an image processing tool can divide an image (such as a H & E stained tissue microarray (TM A) image or a Η & Έ whole tissue slice) into a continuous pixel group called an object. These objects can be obtained through a region growing rule system that finds continuous regions based on color similarity and shape similarity. The size of the object can be changed by adjusting some parameters [1 1]. In this system, objects other than one pixel are typically the smallest processing unit. In this way, all morphological characteristics calculations and operations can be implemented for objects. -28- 200538734 (25) For example, ‘When applying a floor to an image, it is a property limit of the object. As a result, all pixels in the object are assigned the same as in a specific example, and the size of the object can be controlled at the finest level from 10 to 20 pixels. According to this level, subsequent higher and coarser levels can be constructed by forming larger objects from the lower levels. Background extraction. After the initial segmentation, the image processing tool ’s degree limit and convex hull cut out the image tissue core from the background (transparency of the slide. The intensity limit is the intensity that divides the image pixels into β tissue cores and “background”値. Any pixel with an intensity greater than or equal to 値 is scored as “organic core,” and pixels are classified as “background” pixels. The convex hull of a geometric object is the smallest convex set of the pieces. If the entire line segment PQ of any two points in S is also in S, then the set S is convex. Rough segmentation. In the next stage, the image processing tool can foregroumd (for example, TMA kernel) and then segment or correspond to the coarse of the kernel. Area. For example, the main feature in H & E-stained images is that for pathological objects, they are stained blue. Therefore, the difference 値 in the intensity 値 of the blue channel (R-B) can be used as a distinction. For each image obtained in the initial segmentation step, the difference between the average red and blue pixel intensity 値 can be determined. The length / width ratio can be used to determine whether an object should be classified. For example, objects falling within the (R-B) characteristic threshold 値 and length / width threshold can be classified as nuclear sites. Similarly, you can use green to classify objects in the tissue core as white space. Tissue matrix is a type that is subdivided by 値. If you want to make it smaller, you can use the strong bright area.) It can be divided into two categories: "The bottom limit is too high. Otherwise, the image contains the object formed by the foreground (and white space relative to it, in red and characteristics. Special features. ， Risk. It is also the bottom of the road below the nuclear site. The red occupies -29- 200538734 (26). The intensity difference of the image object is d. The “red ratio” r = R (R + G + B) And red track standard deviation σ R can be used to classify matrix objects. White space classification. In the coarse segmentation stage, the 'white space area' can correspond to both the cavity (pathological object) and the corrected crop (fractured tissue site) in the image. Small white space objects (those with an area of less than 100 pixels) are often orthopedic crops. In this way, image processing tools can apply area filters to classify them as orthopedic crops. # Checking out, fusion and classification. In the coarse segmentation stage The obtained nuclear site is often a continuous fusion region covering several solid nuclei. Furthermore, the nuclear region may also include surrounding misclassified cytoplasm. Therefore, these fusion nuclear sites may require nuclear de-fusion to obtain individual nuclei. Images Processing tools To use two different methods to de-fusion the core. The first method can be based on the area growth rule system that fuses the image objects constituting the core part under the shape constraint (roundness). When the fusion is not serious, this method It has been determined to work well. ^ In the case of severe fusion, image processing tools can use different methods based on supervised learning. This method involves manually marking the nuclear site by an expert (pathologist). It is labeled The characteristics of kernel image objects can be used to design statistical classifiers. In a specific example, in order to reduce the number of feature space dimensions, two different classifiers can be used: the Bayesian classifier and k-nearest neighbor classifier. Feature selection [i 2] is implemented on the training set. Levage-one method (13) can be used for cross-validation, and sequential forward search rule system can be used to find the best feature. Finally, we can design -30- 200538734 (27) Two Bayesian classifiers with characteristic numbers equal to 1 and 5. The category conditional distribution can be assumed to have a diagonal covariance Gaussian distribution of the array. In some specific examples, the input image may include different types of nuclei: epithelial nuclei, fibroblasts, basal nuclei, endothelial nuclei, apoptotic nuclei, and red blood cells. Due to the number of graded tumors, the number of epithelial nuclei It is typically considered an important feature, so distinguishing epithelial nuclei from others can be important. Image processing tools can do this by classifying the nuclei detected into two categories: epithelial nuclei and shape-based (Eccentricity) and dimensional (area) characteristics. The additional details regarding image segmentation and classification according to the present invention are set out in U.S. Patent Application No. 10, filed November 17, 2004, cited above. / 991,847, and US provisional patent applications 60 / 520,815 filed on November 17, 2003 and 60 / 552,497 filed on March 12, 2004. As described above, the image processing tool can measure various morphological characteristics after the objects in the image are divided and classified by the tool. These morphological characteristics may indicate one or more properties and / or statistics. The object properties may include spectral properties (e.g. color channel average, standard deviation and brightness) and structure / shape properties (e.g., area, length, width, density, density). Statistics can include minimum, maximum, average, and standard deviation and can be calculated for each property of the image object. Tables 1 and 2 (if attached) show examples of various morphological characteristics that can be measured according to the present invention. The morphological properties in these tables are named using signs that indicate the various properties and / or statistical properties measured by these properties. The specially named signs shown in Tables 1 and 2 are derived from the above-mentioned -31-200538734 (28) commercially available Definiens software products, and are therefore known to those skilled in the art. It is to be understood that the computer-generated morphological characteristics shown in Tables 1 and 2 are merely examples and any computer-generated morphological characteristics may be utilized without departing from the scope of the present invention. For example, Tables 1 and 2 include different morphological combinations. The reduced and modified set of characteristics in Table 2 (ie, reduced and modified compared to the characteristics of Table 1) is an additional step from the implementation of Table 1 in the field of prostate cancer recurrence and survival time. Experimental winner. In particular, this additional experiment provides additional insights into the categories of properties that can be more likely to be associated with the results. The inventors of the present case anticipate that continuous experimentation and / or use of other suitable hardware, software, or a combination thereof may result in various other computer-generated feature sets that can be associated with these and other medical conditions (e.g., Table 2 Characteristics subgroup). Referring to Tables 1 and 2, the characteristics "Cavity · StdDevAreaPxl", "Cavity" indicates a type of image object, and the "StdDev" table uses the statistical 値 (standard deviation) calculated from all cases of the identified cavity, and "AreaPxl "Refers to the characteristics of the object evidence (area of many pixels) to be evaluated with that statistic. Image processing tools can measure morphological characteristics of all objects previously segmented and classified in the image. For example, image processing tools can "Background", "cytoplasm", epithelial nucleus, "lumen", "matrix", "stromal nucleus" and "red blood cells" were measured for morphological characteristics. "Background" includes portions of digital images that are not occupied by the tissue. "Cytoplasm" refers to the cytoplasm of a cell, which can be an irregular site (such as' a pink site surrounding the epithelial nucleus in an image of, for example, H & E stained tissue). "Epithelial nucleus" refers to the presence The epithelium of the glandular unit is thin -32- 200538734 (29) The nucleus of the cell / lumen and basal cell appears as a "round" object surrounded by the cytoplasm. The "lumen" refers to the central glandular space in which Secretions are deposited by epithelial cells everywhere, appearing as closed white areas surrounded by epithelial cells. Sometimes the cavity may be covered with prostate fluid (this is typically pink in H & E stained tissue) or "detritus" (such as , Macrophages, dead cells, etc.). The cavity and epithelial cytoplasm together form a glandular unit. "Matrix" refers to the structure that maintains prostate tissue, with connective tissue shapes of different densities. Stromal tissue lines exist between glandular units and appear red to pink in H & E stained tissues. "Stromal nucleus" is an elongated cell with no or minimal amount of cytoplasm (fibroblasts) ). This category can also include endothelial cells and inflammatory cells, and if there is cancer, the epithelial nucleus has also been found to be dispersed in the matrix. "Red blood cells" are small red round objects, often located in the vessels (arteries or Vein), but may also be scattered throughout the tissue. "C2EN" in the table below is the relative ratio of nuclear area to cytoplasm. Epithelial cells are more neoplastic / malignant and the more area the nuclei occupy. "'EN2SN" B is the percentage or relative amount of epithelial cells to stromal cells contained in the digital tissue image. "L2Core" is the number or area of cavities contained in the tissue. The higher the Gleason grade, the more invasive the cancer, and therefore the lower the cavity content. Usually, this is because the epithelial cell line replicates in an uncontrolled manner at the time of the cancer, causing the cavity to be filled with epithelial cells. "C2L" is the relative amount of cytoplasm to lumen. "C EN 2 L" is the relative amount of cytoplasmic epithelium to lumen. In one aspect of the invention, systems and methods are provided for screening a medical condition (such as a disease). Inhibitor compounds. Fig. 4 is a flowchart of an exemplary phase included in the screening of an inhibitor compound according to a specific example of the present invention. At -403-200538734 (30) Stage 4 02, a first data set of patients can be obtained, where Includes one or more clinical, morphological, and molecular data. A patient is administered a test compound at stage 404. After stage 404, a second data set is obtained from the patient at stage 406. The second data set may or may not include The same data type (ie, characteristics) included in the first data set. At stage 408, the second data set can be compared to the first data set, here in the second data set after the test compound is administered The change indicates that the test compound is an inhibitor compound. Phase 408 of the comparison data set may include, for example, input to the first data set generated by the prediction model of the present invention The responsive output is compared to an output responsive to the input of the second data set produced by the predictive model. For example, the inhibitor compound can be a given drug and the invention can determine the drug to a medical Whether the condition is an effective medical treatment. At this point, various exemplary applications of specific examples of predicting medical conditions of the present invention will be explained. In the first example, the specific examples of the present invention use clinical and morphological data to predict prostate cancer. Recurrence. In the second embodiment, the specific example of the present invention uses clinical, morphological and molecular data to predict the recurrence and overall survival rate of prostate cancer. In the third embodiment, the specific example of the present invention is used to predict the patient in the prostate. The occurrence of invasive diseases after amputation. In the fourth embodiment, specific examples of the present invention are used to predict liver toxicology. Prostate Cancer Overview

***癌爲美國男人死亡的前導肇因，在2004年預 期有230,000個新診斷例及近乎30,000個死亡例。以PSA -34-Prostate cancer is the leading cause of death in American men, with 230,000 new diagnoses and nearly 30,000 deaths expected in 2004. Taking PSA -34-

200538734 (31) 進行以血清爲基礎的篩選之擴大使用提供醫生:ί (亦即Tla-c，Τ2 )偵檢出***癌之能力，{ 在***癌或區域散布者，而只有小百分比在幸 檢到。經報導的早期偵檢和診斷之益處對患者牙 師兩造都在選擇療程上施加鉅大壓力。在選擇ί 介入時，正確豫後的需要係具關鍵性者，因爲） 瘤都是無痛者且要求最低的介入（亦即”警戒性 而他者則較具侵害性且建議給予早期介入（亦I 法/激素/佐劑系統性療法/臨床實驗安置）。再3 斷根***切除術的警戒性等符之隨機化實驗[ 只能導致最輕微的益處（***切除術後只有 死率減低），可推測出需要更好的患者分層措拥 別化的患者看護[14]。 PCa的自然史再度強調在患者診斷時面對碧 [1 5 ]。即使早期***癌可用局部療法予以治瘦 25 — 40%的男人會發展出PSA/生物化學性復發 使事態甚至更複雜者，已有復發過的*** PSA/BCR後約8年仍有可會g發生轉移（平均8与 中間時間爲5年），可猜測出在此組患者的治渥 早的鑑定（包括預測彼等的BCR時間以及彼等 移之傾向兩者）對於彼等的整體存活係最重要幸 ，既有的預測模型在準確性上有限且對於特定達 彼等的腫瘤病理學而個體化。雖然有多種遺傳， 命形態變化已涉及在PCa的疾病發生之中，不站 ΐ較早階段 L括向部化 義移階段偵 ]泌尿科醫 J始治療性 :部份的腫 等待”）， 〗，放射療 ί，在比較 1，從手術 6.6 %的致 ί來導引個 者的挑戰 〖，仍有約 (BCR)。 癌男人在 :;BCR 後 ί法訣中愈 2發展出轉 〖。可惜地 (者會針對 環境和生 I在目前尙 -35- 200538734 (32) 無單一的生化途徑，基因突變或臨床生物標誌可預測-所 給患者結果。斷根***切除術2 1年改變得再度普及且 PSA廣泛使用後15年，泌尿學家仍然不能告訴患者，對 於局部化疾病有何種治療可導致最佳的臨床去病或整體存 活。 只根據臨床特性資料的豫後圖解表事實上確實能提供 有用的臨床狀態和結果之預測，但在準確性和通用性上需 要改良[1 6]。本發明具體實例提供一種”系統性病理學”作 法以成功地改良預測模型對PSA/BCR後-***切除術 之準確性。此代表有腫瘤樣品的患者之”個體化”觀，包括 細胞和微解剖形態學特性，臨床型態及分子標誌物的定量 性評估以創造出極爲準確且整體性的預測模型。經由利用 域特論（domain expertise )，發展出用於預測P S A復發 的極準確模型。此等成果驗證了系統病理學在產生預測性 和豫後性模型上的用處。再者，分析證實一有限組的臨床 變數、分子生物標誌物、和組織形態學特性可以導出及包 括在泌尿學者/病理學者所用的預測檢驗中以根據指定的 臨床結果構造出最優患者治療計畫。所選出的與PSA復發 相關之分子特性推測出生長因子信令機制（透過雄激素受 體（後文"AR”，將於下文說明之）與細胞偶合血管化（透 過CD 34 )之收歛角色。CD 34爲一種透膜性糖蛋白，存 在於將人體內的血管連線之內皮細胞上。進一步的硏究正 進行中以期更佳地了解此等觀察及對預測***癌進展的 潛在影響。此外也要提及者爲所選出的影像分割和形態學 -36- 200538734 (33) 特性，係除了對模型開發和準確性具重 織描述器之外，部份地代表極準確， Gleason Score。經定義的與 Gleason S 之形態學特性部份地包括腺體構造的整 的形狀和尺寸（細胞質組成），上皮細 單一上皮細胞之展示。 雄激素受體蛋白（AR )在天然發生 φ 睪酮及其5、α-還原代謝物、二氫睪200538734 (31) Expanded use of serum-based screening to provide doctors: ί (aka Tla-c, T2) the ability to detect prostate cancer, {spread in prostate cancer or regional spread, and only a small percentage are lucky Detected. The reported benefits of early detection and diagnosis put tremendous pressure on both patients' dental practitioners to choose a course of treatment. When choosing intervention, the need for correct aftermath is critical, because) tumors are painless and require minimal intervention (that is, "vigilant" while others are more invasive and it is recommended that early intervention (also I method / hormonal / adjuvant system therapy / clinical experimental placement). 3 randomized trials of cautious equivalence of decapitated prostatectomy [can only lead to the slightest benefit (only reduced mortality after prostatectomy), It can be inferred that better patient stratification measures and individualized patient care are needed. [14]. The natural history of PCa once again emphasizes facing patients during diagnosis [1 5]. Even early prostate cancer can be treated with local therapies. 25 — 40% of men will develop PSA / biochemical relapses to make matters even more complicated. Metastasis may still occur in the prostate about 8 years after PSA / BCR has relapsed (mean 8 and 5 years in between) ), It can be guessed that the early identification of patients in this group (both predicting their BCR time and their tendency to migrate) is the most important for their overall survival. The existing prediction models are accurate Sexually limited Individualized for specific tumor pathology. Although there are multiple inheritances, life-form changes have been involved in the occurrence of PCa disease, and do not stand in the early stages. J Therapeutic: Partial swelling is waiting ")," Radiation therapy, in comparison 1, the challenge of guiding the individual from 6.6% of the surgery, there is still about (BCR). The cancer man developed a turn in: 2 after the BCR. It is a pity that (for the environment and health I at present 尙 -35- 200538734 (32) there is no single biochemical pathway, genetic mutations or clinical biomarkers can be predicted-the results given to patients. Rootectomy prostatectomy 21 changes again in 1 year Fifteen years after the popularity of PSA and its widespread use, urologists still cannot tell patients what treatments for localized disease can lead to the best clinical removal or overall survival. Heyu diagrams based on clinical characteristics data are indeed true It can provide useful predictions of clinical status and results, but needs to be improved in accuracy and versatility. [16] The specific examples of the present invention provide a "systematic pathology" approach to successfully improve the prediction model for PSA / BCR- Accuracy of prostatectomy. This represents the "individualized" view of patients with tumor samples, including quantitative evaluation of cellular and microanatomical morphological characteristics, clinical patterns and molecular markers to create extremely accurate and holistic Prediction model. By using domain expertise, a highly accurate model for predicting PSA recurrence has been developed. These results have verified The usefulness of traditional pathology in generating predictive and postmenopausal models. Furthermore, analysis confirms that a limited set of clinical variables, molecular biomarkers, and histomorphological characteristics can be derived and included in urologists / pathologists In the prediction test, the optimal patient treatment plan is constructed based on the specified clinical results. The selected molecular characteristics related to PSA recurrence are used to infer the growth factor signaling mechanism (through the androgen receptor (hereinafter " AR ", (Explained below) Couples the convergent role of vascularization with cells (through CD 34). CD 34 is a transmembrane glycoprotein that exists on endothelial cells that connect blood vessels in the body. Further research is ongoing With a view to better understanding these observations and their potential impact on predicting the progression of prostate cancer. In addition, the characteristics of the selected image segmentation and morphology-36- 200538734 (33) should also be mentioned, in addition to the model development and accuracy In addition to the re-weaving descriptor, partly represents extremely accurate, Gleason Score. The defined and morphological characteristics of Gleason S partially include the integration of glandular structure The shape and dimensions (cytoplasmic composition), epithelial cells of epithelial single display androgen receptor protein (AR) in the naturally occurring testosterone and φ 5, α- reducing metabolite, dihydro epididymis

Ley dig細胞合成之後，即接收此等激素 之後，此等激素會循環通過整個身體且 受體AR發生作用的雄激素會刺激胎兒 輔助性腺體之發育，春機發動期男性的 ，及維持成年男性特徵和生育功能。雄 類固醇激素受體一起地，構成一族轉一 白，可透過與特定基因序列的交互作用 • 對AR針對***癌的硏究業已推 激素受體的存在與彼等對雄激素性激素 關性之間可能存在著正面相關聯。例如 專利第6,472,4 1 5號提出早期的*** 驅動者且可，至少暫時地，經由雄激 French et al.美國專利第 6,821，767 號 量AR，以期讓醫生使用雄激素受體檢 斷評定中。不過，這些硏究都沒有提出 本文所揭示的可預測***癌發生之自 要性的數種新穎組 無主觀性且量性之 coring System 相關 體外觀，上皮細胞 胞核及基質內摻混 .的雄激素系激素（ 酮）於雄性睪九的 ;。特別者，在合成 _結合到 AR。透過 ,體內的男生殖器和 男性特徵化和生長 激素受體、與其他 作用性轉錄調節蛋 而控制基因轉錄。 :測出在癌細胞內雄 刺激用以生長之相 ，Sovak et al.美國 癌生長係由雄激素 素撤除予以停止。 :提出多種方法來測 定於***癌的診 使用 AR測量配合 動化模型。 -37- 200538734 (34) 實施例1 :***癌復發的預測臨床和形態學資料 使用根據Definiens Cellenger軟體的MAGIC組 像分析系統從每一***組織像提取出初始大到500 許多原形態學特性。將整組原特性以不可知論方式選 避免忽視掉潛在有用的特性。不過，此等形態學特性 全部都具同等資訊性，且根據完全特性組構建的預測 會因爲•’廣延阻咒’’（curse of dimensionality)而可能 不良的預測性能[1 3 ]。所以要施以廣延性減低程序， 最後選出一有8個形態學特性的組。 從一群有進行過斷根***切除術的***癌患 個1 5 3個患者亞組進行硏究。使用在手術後可測量的 腺特異性抗原（PSA )來定義***癌復發（也稱爲 化學復發（BCR ))。患者係在手術後追踪。將彼等 等的最後一次看診之復發狀態，以及彼等的追踪時間 錄下來，產生一組就檢資料（right-censored data)。 在手術前從活組織檢體及在手術後使用切出的*** 者都測量Gleason計分。在此硏究中考慮到的四種特 床測量，或特性，爲（1 )活體組織Gleason等級， 活體組織G1 e a s ο η計分，（3 )手術後G1 e a s ο η等級， 4)手術後Gleason計分。 從臨床導出的Gleason計分特性分開地分析形態 性以同時預測PSA/BCR復發的機率和時間。然後將 與Gleason計分（特生）組合以建立復發及到復發時 織影 個的 出以 並非 模型 具有 且於 者選 前列 生物 在彼 ，記 對於 體兩 疋臨 (2 ) 與（ 學特 影像 的時 -38- 200538734 (35) 間預測。由此連合特性組所達到的改良G 1 e a s ο η準確度指 出影像特性確實提供額外的資訊因而改良復發預測率及整 體預測模型。 因爲此群患者具有就檢結果資料，所以必須構建存活 分析模型以預測復發。爲了對不同類型資料產生潛在的規 則系統偏差，乃使用兩種存活分析規則系統：1 ) Cox回 歸模型[17];和2 )上述SVRc，如應用於支援向量機者。 採用以5倍交叉確認予以估計的和諧指數來測量模型的預 測準確度[13][18]。 兩種規則系統都應用於三資料組：（1 )單獨的 Gleason計分臨床特性；（2 )單獨的所選形態學特性；和 (3 )形態學特性和Gleason計分臨床特性的組合。實驗 結果列於表3之中。 於此實施例中選定的臨床特性爲BXGGTOT，BXGG1 ，GGTOT，和GG1，且所選的形態學特性係有關上皮核（ 上皮、核、MaxCompactness)，背景（背景、After the synthesis of Ley dig cells, that is, after receiving these hormones, these hormones will circulate through the entire body and the androgen that acts on the receptor AR will stimulate the development of auxiliary fetal glands in the fetus. Male characteristics and fertility. Together with androgen hormone receptors, they form a family that can be turned into white through interactions with specific gene sequences. Studies of AR against prostate cancer have pushed the existence of hormone receptors and their relationship to androgen hormones. There may be a positive correlation. For example, Patent No. 6,472,4 15 proposes an early prostate driver and may, at least temporarily, via androgen French et al. US Patent No. 6,821,767 to AR, in order to allow doctors to use androgen receptor testing . However, none of these studies has proposed the novel subjective and quantitative coring System related bodies, epithelial cell nuclei, and intrastromal admixtures disclosed in this article that can predict the essentiality of prostate cancer. Androgens are hormones (ketones) in males; In particular, in the synthesis _ combined to AR. Through the male genitalia and male characterization and growth hormone receptors in the body, and other functional transcription regulating eggs, gene transcription is controlled. : Measured in male cancer cells to stimulate the growth phase, Sovak et al. American cancer growth system was stopped by androgen removal. : Propose multiple methods to determine the diagnosis of prostate cancer. Use AR measurement combined with an activation model. -37- 200538734 (34) Example 1: Predicted clinical and morphological data of prostate cancer recurrence Using the MAGIC group image analysis system based on Definiens Cellenger software, many original morphological characteristics of up to 500 were extracted from each prostate tissue image. Select the entire set of original features in an agnostic manner to avoid ignoring potentially useful features. However, all of these morphological characteristics are equally informative, and predictions constructed based on the complete characteristic group may have poor predictive performance due to the “curse of dimensionality” [1 3]. Therefore, it is necessary to apply the procedure of reducing the spread, and finally select a group with 8 morphological characteristics. A subgroup of 153 patients with prostate cancer who underwent radical prostatectomy was investigated. A prostate specific antigen (PSA) that is measurable after surgery is used to define prostate cancer recurrence (also known as chemical recurrence (BCR)). The patient was followed after the operation. The recurrence status of their last visit and their follow-up time were recorded to generate a set of right-censored data. Gleason scores were measured from biopsies before surgery and those who used the excised prostate after surgery. The four special bed measurements or characteristics considered in this study were (1) Gleason grade of living tissue, G1 eas ο η score of living tissue, (3) G1 eas ο η grade after surgery, 4) after surgery. Gleason scores. Morphology was analyzed separately from clinically derived Gleason scoring characteristics to simultaneously predict the probability and time of PSA / BCR relapse. Then it will be combined with Gleason score (special student) to establish recurrence and the occurrence of the shadow weaving at the time of recurrence is not the model and the leading creature is selected in the other. Time-38- 200538734 (35) predictions. The improved G 1 eas ο accuracy achieved by this conjoint feature group indicates that image characteristics do provide additional information and thus improve recurrence prediction rates and overall prediction models. Because this group of patients With test results data, a survival analysis model must be constructed to predict recurrence. In order to generate potential rule system deviations for different types of data, two survival analysis rule systems are used: 1) Cox regression model [17]; and 2) above SVRc, as applied to support vector machines. The prediction accuracy of the model was measured using the harmony index estimated with 5x cross-validation [13] [18]. Both rule systems are applied to three data sets: (1) individual Gleason score clinical characteristics; (2) individual selected morphological characteristics; and (3) a combination of morphological characteristics and Gleason clinical characteristics. The experimental results are listed in Table 3. The clinical characteristics selected in this example are BXGGTOT, BXGG1, GGTOT, and GG1, and the selected morphological characteristics are related to the epithelial nucleus (epithelial, nuclear, MaxCompactness), background (background, background,

StdDevAreaPxl )，和腔（腔、MaxBorderLengthPxl，腔、 MinRadiusofsmallestencolosinge，腔、StdDevBorder LengthPxl，腔。SumBorderlengthPxl，腔、StdDevAreaPxl ，和腔、M i n C 〇 m p a c t n e s s )。更特別者，於此實施例中， 有關腔的面積、邊界長度、和形狀（緊密度）都經測定以 與疾病進展相關聯。腔愈小且愈緊密，癌可能更進展。事 實上，對於更具侵蝕性的癌（Gleason等級4和5 )，可 以預測腔會從組織幾乎或完全地消失。此外也測定與癌進 -39- 200538734 (36) 展相關聯的上皮核緊密度之形態學特性，此處緊密度係由 Definiens Cellenger軟體計算的，其爲上皮核的長寬乘積 對上皮核面積之比例。此可能是因爲隨著癌的進展，上皮 核對基質的侵入會增加之故（亦即，具有進展的癌之組織 典型地會包括豐盛的上皮核）。在本實施例中經測定以與 結果相關聯之以背景爲基底的形態學特性係測量分析中所 用組織核心的實際尺寸。 表3 -預測準確性之比較StdDevAreaPxl), and cavity (lumen, MaxBorderLengthPxl, cavity, MinRadiusofsmallestencolosinge, cavity, StdDevBorder LengthPxl, cavity. SumBorderlengthPxl, cavity, StdDevAreaPxl, and cavity, Mi n C m p a c t n s. More specifically, in this embodiment, the area of the cavity, the length of the border, and the shape (tightness) are all determined to correlate with disease progression. The smaller and tighter the cavity, the more likely the cancer is to progress. In fact, for more aggressive cancers (Gleason grades 4 and 5), the cavity can be predicted to disappear almost or completely from the tissue. In addition, the morphological characteristics of the tightness of the epithelial nucleus associated with cancer progression-39- 200538734 (36) were also determined. Here the tightness is calculated by Definiens Cellenger software, which is the length-width product of the epithelial nucleus versus the area of the epithelial nucleus. Ratio. This may be because the invasion of the epithelial nucleus into the stroma increases as the cancer progresses (i.e., tissues with progressive cancer typically include abundant epithelial nuclei). The background-based morphological characteristics determined in this example to correlate with the results are the actual dimensions of the tissue core used in the analysis. Table 3-Comparison of prediction accuracy

Gleason 影像 Gleason +影像 Cox 0.6952 0.6373 0.7261 S VRc 0.6907 0.7269 0.7871 根據表3，形態學特性的預測性能可與Gleason計分 相比，且形態學特性和Gleason計分的組合則達到更高的 預測率，此點可確認由組織影像分析系統所提取的形態學 特性確實可提供Gleason計分以外的額外資訊。所以，形 態學測量的使用可增強整體復發預測。 實施例2 :***癌復發和整體存活的預測臨床，形態學 和分子資料 進行兩硏究，分別以88%和87%的預測準確率成功地 預測***異性抗原（P S A )復發。經由將臨床、分子和 形態學特性組合著機器學習，可創建一堅強的平台，其在 -40- 200538734 (37) 患者診斷，治療管理和豫後中具有廣泛應 個硏究以預測***癌患者的整體存活率 爲因任何肇因所致死亡。 對一群5 3 9個業經斷根***切除術 ’包括則列腺切除術樣品構成的局密 TMAs)。使用蘇木素（hematoxylin)和! H&E )染色的組織切片實施形態學硏究且 (IHC )評估分子生物學決定因素。透過 選定的一組特性導出用於PSA復發和整體 模型。對於具有在每一域中的完全無誤失 用經開發用來處置已檢資料的支援向量機 歸評定。以所產生用來界定風險組的計分 CI )估測該模型的預測性能。 使用一群1 3 2個患者，有4 1項特性 特性，14項分子特性，和1 〇項形態學特 有8 8%的準確率來預測PSA復發。於一群 1 〇項特性（3項臨床特性，1項分子特性 特性）經測得可用8 7 %準確率預測P S A 1 4項特性（2項臨床、1項分子、和1 1項 用80%準確率預測整體存活率。使用對數 險組之間（Ρ<0·000 1 )觀察到腫瘤復發與 異。 本硏究揭露出透過組合著臨床變數， 組織學，以機器學習分析的新系統作法， 用。也進行第三 ，此處目標結果 的患者進行硏究 度組織微陣列（ 曙紅（e 〇 s i η )( 用免疫組織化學 監督多元學習從 存活率者之預測 資料之患者，使 (SVRc)進行回 使用和諧指數（ (包括1 7項臨床 性）經選出，可 268個患者，有 ，和6項形態學 復發；此外，有 形態學）經查可 等級檢驗，於風 死亡中的明顯差 分子標6志物，和 有改良的*** -41 - 200538734 (38) 癌復發預測之遞增傾向。 患者臨床特性。對一群5 3 9個進行過斷根***切除 術的患者進行硏究。使用包括患者年齡、手術前P S A、和 G 1 e a s ο η等級之除—鑑別（d e - i d e n t i f i e d )患者資訊追溯地 收集1 7項臨床特性（顯示於下面表4之中）。Gleason image Gleason + image Cox 0.6952 0.6373 0.7261 S VRc 0.6907 0.7269 0.7871 According to Table 3, the prediction performance of morphological characteristics can be compared with Gleason score, and the combination of morphological characteristics and Gleason score can achieve a higher prediction rate. This confirms that the morphological characteristics extracted by the tissue imaging analysis system can indeed provide additional information beyond the Gleason score. Therefore, the use of morphological measurements can enhance overall recurrence prediction. Example 2: Prediction of prostate cancer recurrence and overall survival Clinical, morphological, and molecular data Two studies were performed to successfully predict the recurrence of prostate heterosexual antigen (PSA) with 88% and 87% prediction accuracy, respectively. By combining clinical, molecular, and morphological characteristics with machine learning, a strong platform can be created that has extensive research in predicting patients with prostate cancer in -40- 200538734 (37) patient diagnosis, treatment management, and post-mortem. The overall survival rate is death due to any cause. To a group of 5 3 9 patients who underwent radical prostatectomy, including TMAs (prosthetic thyroidectomy samples). Hematoxylin and! H & E) stained tissue sections were used to perform morphological studies and (IHC) to assess molecular biology determinants. Derived from a selected set of characteristics for PSA recurrence and overall models. For complete error-free evaluation in each domain, a support vector machine has been developed to handle the checked data. The score (CI) generated to define the risk group is used to estimate the predictive performance of the model. Using a cohort of 132 patients, 41 traits, 14 molecular traits, and 10 morphology-specific 88% accuracy were used to predict PSA recurrence. From a group of 10 properties (3 clinical properties, 1 molecular property), PSA 1 4 properties (2 clinical, 1 molecular, and 11 with 80% accuracy) can be predicted with 87.7% accuracy. The overall survival rate is predicted with certainty. Tumor recurrence and difference are observed between logarithmic risk groups (P < 0.00 1). This study revealed a new systematic approach by combining clinical variables, histology, and machine learning analysis, The third step is to perform intensive tissue microarrays (eosin (e)) on patients with targeted outcomes (using immunohistochemistry to monitor multivariate learning from predictors of survivors to make (SVRc ) The harmonization index (including 17 clinical items) was selected, and 268 patients were included, and there were 6 morphological recurrences. In addition, there were morphological findings. Grade examination was found to be obvious in wind death. Differentially labeled 6-magazines, and have an improved tendency to predict prostate cancer recurrence -41-200538734 (38). Clinical characteristics of patients. A group of 5 3 9 patients who underwent radical prostatectomy were investigated. Use includes Age, before surgery P S A, and G 1 e a s ο addition η hierarchy - Identification (d e - i d e n t i f i e d) a patient information retroactively clinical characteristics of 17 item collector (shown in the following Table 4 in).

表4·收集到的臨床特性 特性 說明 age 年齡（以年計） race 種族 prepsa ***特異性抗原（毫微克/分升） t nm TNM臨床階段 u i c c UIC C臨床階段 dr e 數位直腸檢驗可觸診到 In 淋巴結狀態 s v i 精囊侵入 margins +/-外科邊際 e c e 位於囊外的腫瘤 bxggl 通性活體組織G1 e a s ο η等級 bxggtot 活性組織Gleason計分 ggl 顯性手術後Gleason等級 ggtot 手術後Gleason計分 pr sited 二倍體，四倍體，非整倍體 pp—sphas S期中倍性細胞百分比 pp fr ac 倍性增生分數 -42- 200538734 (39) 從所選***切除樣品推構建組織微陣列（TMAs ) 。將每一樣品所得直徑0 · 6毫米的組織核心以三重複隨機 地排列於每一接受石塊（Beecher Instruments，Silver Spring, MD )。將此等TMA塊的切片（5微米）放置於荷 電的聚離胺酸塗覆載片上，且用於形態學和免疫組織化學 (IHC )分析中（參閱下文）。 臨床特性的缺失値則分配以包括全部特性的彈性加性 回歸模型以在不參照結果之下估計缺失特性之値，且只有 彼等有完全臨床（於分配之後），形態學和分子資料，以 及無缺失結果資訊之患者，才繼續接受硏究。硏究1 (槪 念證明）的有效樣品大小係包括1 3 2個患者。主要標的分 類爲患者在***癌手術之後是否復發。在兩次連續觀測 到PSA>0.2毫微克/毫升的增高之患者即視爲有復發的前 列腺癌。若患者在最後一次看診沒有復發，或患者在其最 近看診結果爲未知（亦即，因漏失追踪所致），則患者的 結果即視爲已檢者（encored )。到復發的時間係經定義 爲從斷根***切除術到PSA (生化）復發爲止之時間（ 以月計算）。 硏究2係使用從原有5 3 9個患者群選出的26 8患者， 包括硏究1的1 3 2個患者中之1 2 9個，進行的。取代來自 TMA核心的H&E影像者，係對來自斷根***切除術的 整個切片進行分析。硏究3探討相同的268 -患者群，但 是係預測整體存活率。此處的目標結果爲因任何肇因所致 死亡。 -43- 200538734 (40) 影像分析與形態學硏究。將擷取自每一患者的原有腫 瘤組織之代表性部位，不論是組織核心或整體切片，都予 以數位化且使用H&E染色載片分析。影像係用光學顯微 鏡於 20X 放大倍率以 SPOT Insight QE Color Digital Camera ( KA 1 2 000 )攝取的。只選擇包含大於80%腫瘤的 部位進行最優影像分割與定量分析。 分子分析·將每一組 12種生物標誌物，包括 Cytokeratin 18 (腔細胞）、Cytokeratin 14 (基底細胞） 、CD45 (淋巴細胞）、CD34 (內皮細胞）、CD68 (巨噬 細胞）、Ki67 (增殖）、PSA ( hK - 3，激肽釋放酶）、 PSMA (生長受體）、胞轉蛋白 D1 (細胞周期）、p27 ( 細胞周期）、雄激素受體（內分泌）和Her — 2/neu (信令 )以標準色原免疫組織化學應用於所有7個TMA塊。雄 激素擷取係用〇·〇1 Μ檸檬酸緩衝液（pH6)在壓力鍋內對 所有抗體處理3 0分鐘。有關此一程序的範例方法和系統 經載於上引2003年七用21日提出申請且名稱爲"Methods and compositions for the preparation and use of fixed-treated cell-lines and tissue in fluorescence in situ hybridization”的美國專利申請第1 0/624,23 3號之中。原 抗體（表5中所示者）皆經稀釋在Tris-緩衝食鹽水（含 0.1%Tween)之中且在41：施用16小時，接著以1:1〇〇〇 稀釋率施用經生物素化二次抗體（Vector ) —小時。 -44 - (41) (41)200538734 表5.抗體表 生物標誌物 株系 Ki-67 株系 i-67(DAKO) Cy tokeratin 1 8 株系 DC-lO(Novocastra) CD45 株系 XI 6/99 CD68 株系 5 1 4H2(Novocastra UK) CD34 株系 QBEnd lOl(DAKO) AR 株系 AR27(Novocastra) Cytokeratin 14 株系 LL002(Novocastra) 胞轉蛋白D 1 株系 P2D1 1F11 PSA 株系 PA05 (Neomarkers) PSMA 株系 ZMD.80(Zymed)p p27 株系 DCS72(Oncogene) Her-2/neu KIT DAKOp 多株型，其他都是單株型 陰性對照載片係接受正常小鼠血清（DAKO )作爲原 抗體。將諸載片使用Harris蘇木素予以複染色且由兩個獨 立的病理學家予以細察並由第三位病理學家解析所有矛盾 之處。所記錄的來自所有5 3 9個患者及彼等的個別三重複 核心之IHC資料包括經針對特別的硏究抗原染色過的細胞 之百分比和強度（0 - 3+ )。於可用時，將此兩種量度組 合以造出對該特別生物標誌物的染色指數（下面的表6， 顯示出範例分子特性表）。對AR (雄激素受體）CK 1 4 ( -45- 200538734 (42) C y t o k e r a t i η 1 4 )、胞轉蛋白 D 1、P S A (***特異性抗 原）、PSMA (***特異性膜抗原）、p27和Her-2/neu 計算出染色指數，而其餘標誌物（亦即，Ki67、CK18 ( Cytokeratin 1 8 ) 、CD45、CD68 )都是根據具有所給強度 的陽性細胞之百分比予以計算。此等生物標誌物要在下文 進一步說明。染色指數的範圍〇 — 3 00，且係按下述計算： 1 * (對一生物標誌物具有1 +強度的陽性染色細胞之百分 比）+2* (對該生物標誌物具有2 +強度的陽性染色細胞之 百分比）+ 3 * (對該生物標誌物具有3 +強度的陽性染色細 胞之百分比），此處陽性染色細胞之百分比指的是每1〇〇 計數細胞中經鑑定爲陽性的細胞之數目。有關此種染色指 數的額外細節經於[1 9]之中。此種染色指數只具範例性且 測量分子特性所用的任何其他適當方式都可以使用而不違 離本發明範圍。 於上述生物標誌物的討論中，p2 7屬於稱爲胞轉蛋白 -依賴性激酶抑制劑的細胞周期調節劑族，其可結合到胞 轉蛋白一 CDK複合物且造成細胞周期停止在G1期。生物 標誌物P27經推測會促進細胞凋亡且在某些組織的終端分 化中具有其作用。利用免疫組織化學，核P27表現的喪失 會體隨著更侵害性的表現型。Her - 2/neu爲受體酪胺酸激 酶EGFR族之一員且在某些人類癌症的病生成中起重要作 用。由免疫組織化學證明Her2/neu在細胞膜上的過度表 現與更侵害性乳癌類型相關聯。Ki67爲許多種增生標誌物 中之一者，可用不同強度染著核且可用來評估目標腫瘤樣 -46- 200538734 (43) 品所具細胞活性之增殖指數或量度。C D 4 5爲一種細胞表 面抗原，可用來鑑定最後要變爲免疫細胞例如淋巴細胞（ T細胞、B -細胞、NK細胞等）之細胞。強度經認爲不像 其分布/存在及與其他組織學元件的結合等一樣重要。 C D 6 8爲與彳谷體密切結合的一種細胞質抗原。其係在整 個單核細胞分化級聯中表現但通常在巨噬細胞中比在單核 細胞中更密集。Table 4 · Clinical characteristics collected collected age age (in years) race prepsa prostate specific antigen (nanograms / deciliter) t nm TNM clinical stage UICC UIC C clinical stage dr e Digital rectal examination can be palpated to In lymph node status svi seminal vesicle invasion margins +/- surgical margin ece tumors located outside the capsule bxggl general living tissue G1 eas ο η grade bxggtot active tissue Gleason score ggl after dominant surgery Gleason grade ggtot post surgery Gleason score pr sited 2 Ploid, tetraploid, aneuploid, pp-sphas, percentage of ploidy cells in S phase, pp fr ac, ploidy hyperplasia score -42- 200538734 (39) Tissue microarrays (TMAs) are constructed from selected prostate resection samples. Tissue cores with a diameter of 0.6 mm obtained from each sample were randomly arranged in triplicate on each receiving stone (Beecher Instruments, Silver Spring, MD). Sections (5 micrometers) of these TMA blocks were placed on charged polylysine coated slides and used for morphological and immunohistochemical (IHC) analysis (see below). Missing clinical characteristics are assigned to an elastic additive regression model that includes all characteristics to estimate the missing characteristics without reference to the results, and only they have full clinical (after distribution), morphological and molecular data, and Patients without missing outcome information will continue to be investigated. The valid sample size for Study 1 (proof of proof) included 1 3 2 patients. The main target classification is whether patients have relapsed after prostate cancer surgery. Patients with an increase in PSA > 0.2 ng / ml between two consecutive observations are considered to have recurrent prostate cancer. If the patient did not relapse at the last visit, or if the patient's most recent visit result was unknown (ie, due to missed follow-up), the patient's result is considered encored. The time to relapse is defined as the time (in months) from the radical prostatectomy to the recurrence of PSA (biochemical). Study 2 was performed using 268 patients selected from the original 539 patient population, including 1 2 of the 132 patients in Study 1. Instead of H & E images from the core of the TMA, the entire section from a radical prostatectomy was analyzed. Study 3 explores the same 268-patient population, but predicts overall survival. The target result here is death from any cause. -43- 200538734 (40) Image analysis and morphological research. Representative parts of the original tumor tissue, whether from the core or the whole section, taken from each patient were digitized and analyzed using H & E stained slides. Images were taken with an optical microscope at 20X magnification with a SPOT Insight QE Color Digital Camera (KA 1 2 000). Only sites with more than 80% tumors were selected for optimal image segmentation and quantitative analysis. Molecular analysis: each group of 12 biomarkers, including Cytokeratin 18 (lumen cells), Cytokeratin 14 (basal cells), CD45 (lymphocytes), CD34 (endothelial cells), CD68 (macrophages), Ki67 (proliferation ), PSA (hK-3, kallikrein), PSMA (growth receptor), cytosolic D1 (cell cycle), p27 (cell cycle), androgen receptor (endocrine), and Her — 2 / neu ( Signaling) was applied to all 7 TMA blocks with standard chromogen immunohistochemistry. Androgen extraction was performed with all the antibodies in a pressure cooker for 30 minutes in a pressure cooker with 0.001 M citric acid buffer (pH 6). The example method and system for this procedure are set out in the above-mentioned application on July 21, 2003 and the name is "Methods and compositions for the preparation and use of fixed-treated cell-lines and tissue in fluorescence in situ hybridization" U.S. Patent Application No. 10 / 624,23 3. The original antibodies (shown in Table 5) were diluted in Tris-buffered saline (containing 0.1% Tween) and administered at 41: 16 hours Then, the biotinylated secondary antibody (Vector) was administered at a 1: 1000 dilution rate for one hour. -44-(41) (41) 200538734 Table 5. Antibody table biomarker strain Ki-67 strain i-67 (DAKO) Cy tokeratin 1 8 strain DC-lO (Novocastra) CD45 strain XI 6/99 CD68 strain 5 1 4H2 (Novocastra UK) CD34 strain QBEnd 10 (DAKO) AR strain AR27 (Novocastra) Cytokeratin 14 strain LL002 (Novocastra) Cytoprotein D 1 strain P2D1 1F11 PSA strain PA05 (Neomarkers) PSMA strain ZMD.80 (Zymed) p p27 strain DCS72 (Oncogene) Her-2 / neu KIT DAKOp multiple strains Type, others are single-type negative control slides that receive normal mouse serum (DAKO) as the primary antibody. The slides were counterstained with Harris hematoxylin and carefully examined by two independent pathologists and all contradictions resolved by a third pathologist. Records from all 5 3 The IHC data for 9 patients and their individual triplicate cores includes the percentage and intensity (0-3+) of cells stained for specific research antigens. When available, combine these two measures to create Staining index for this particular biomarker (Table 6 below shows a table of exemplary molecular properties). For AR (androgen receptor) CK 1 4 (-45- 200538734 (42) Cytokerati η 1 4), cell Transfection D 1, PSA (Prostate Specific Antigen), PSMA (Prostate Specific Membrane Antigen), p27 and Her-2 / neu were used to calculate the staining index, while the remaining markers (ie, Ki67, CK18 (Cytokeratin 1 8)) , CD45, CD68) are calculated based on the percentage of positive cells with the given strength. These biomarkers are described further below. The staining index ranges from 0 to 3 00 and is calculated as follows: 1 * (percentage of positively stained cells with 1 + intensity to a biomarker) + 2 * (positive to 2 + intensity to the biomarker Percentage of stained cells) + 3 * (Percentage of positively stained cells with 3 + intensity to the biomarker), where the percentage of positively stained cells refers to the number of cells identified as positive per 100 counts of cells number. Additional details on this dyeing index are given in [19]. Such a staining index is merely exemplary and any other suitable means for measuring molecular properties can be used without departing from the scope of the invention. In the discussion of the above biomarkers, p2 7 belongs to a family of cell cycle regulators called cytosolic protein-dependent kinase inhibitors, which can bind to the cytosolic protein-CDK complex and cause the cell cycle to stop at the G1 phase. The biomarker P27 is speculated to promote apoptosis and have its role in the terminal differentiation of certain tissues. Using immunohistochemistry, the loss of nuclear P27 expression is followed by a more aggressive phenotype. Her-2 / neu is a member of the EGFR family of receptor tyrosine kinases and plays an important role in the pathogenesis of certain human cancers. It was demonstrated by immunohistochemistry that the overexpression of Her2 / neu on the cell membrane was associated with a more aggressive type of breast cancer. Ki67 is one of many proliferative markers that can stain the nucleus with different intensities and can be used to evaluate the proliferation index or measure of cell viability of the target tumor-like -46- 200538734 (43) product. CD 45 is a cell surface antigen that can be used to identify cells that will eventually become immune cells such as lymphocytes (T cells, B cells, NK cells, etc.). Strength is not considered as important as its distribution / presence and integration with other histological elements. C D 6 8 is a cytoplasmic antigen that binds closely to glutenous bodies. It is expressed throughout the monocyte differentiation cascade but is usually more dense in macrophages than in monocytes.

-47- 200538734 (44) 表6.分子特性 特性 說明 atki67t1 atki 6 7 12 atki67t3 at k i 6 7 p 1 a t k i 6 7 p 2 a t k i 6 7 p 3 atki67al atk i 6 7 a2 atki 6 7 a3 a t c 1 8 13 atcd45t3 atcd68t3 at c d 3 4 p at c d 3 4 s at c d 3 41 atcd3 4tp atcd34ts atcd3 4ps a t c 1 8 p 3 atcd45p3 atc18 a3 atcd45a3 ar s i c 1 4 s i c d 1 s i psasi p am a s i p 2 7 s i her2 si ar p s i c 1 4p s i c d 1 p s i p s ap s i p s m ap s i p 2 7 p s i her2psi ar a s i c14 as i cd1asi psaasi psmaasi p 2 7 a s i her2asi Ki-67,於強度部位1中（腫瘤） Ki-67,於強度部位2中（腫瘤） Ki-67,於強度部位3中（腫瘤） Ki-67,於強度部位1中（PIN) Ki-67,於強度部位2中（PIN) Ki-67,於強度部位3中（PIN) Ki-67,於強度部位1中（腺體） Ki-6 7,於強度部位2中（腺體） Ki-67,於強度部位3中（腺體） Cytokeratin 18(腫瘤） CD45(腫瘤） CD68(腫瘤） CD34(PIN) CD34(基質） CD34(腫瘤） CD34(腫瘤 /PIN) CD34(腫瘤/基質） CD34(PIN/基質） Cytokeratin 1 8 (PIN) CD45(PIN) Cytokeratinl8(腺體） CD45(腺體） AR(腫瘤）染色指數 Cytokeratinl4(腫瘤）染色指數 胞轉蛋白D1 (腫瘤）染色指數 PSA(腫瘤）染色指數 PSMA(腫瘤）染色指數 p27(腫瘤）染色指數 Her-2/neu(腫瘤）染色指數 AR(PIN)染色指數 Cytokeratin 14(PIN)染色指數 胞轉蛋白Dl(PIN)染色指數 PSA(PIN)染色指數 PSMA(PIN)染色指數 p27(PIN)染色指數 Her-2/neu(PIN)染色指數 AR(腺體）染色指數 Cytokeratinl4(腺體）染色指數 胞轉蛋白D1 (腺體）染色指數 PSA(腺體）染色指數 PSMA(腺體）染色指數 p27(腺體）染色指數 Her-2/neu(腺體）染色指數 -48- 200538734 (45) 分析與統計硏究·進行三項硏究：用1 32個患者的初 始槪念分析證明（硏究1 )及用2 6 8個患者的擴充硏究（ 硏究2和硏究3 )。於硏究1和硏究2兩者之中’分析包 括兩步驟··鑑別可預測PSA復發的特性及根據此等特性發 展一模型，最終目標爲使用該模型於未來斷根***切除 患者中預測生化（PSA )復發。硏究3的目標爲鑑別特性 及發展出一模型以預測後-***切除之整體存活率。使 P 用上述類型的用於已檢資料之支援向量回歸（SVRc )以在 每一此等硏究中的發展出所得模型。 模型的預測準確性係使用和諧指數（CI )予以評估。 在處理已檢結果之中，此常爲所選量度。和諧指數係根據 兩隨機選定的符合下列任一準則的患者所得豫後計分之間 的配對比較：兩個患者都經歷該事件且第一患者的事件時 間比第二患者較爲短或者只有第一患者經歷該事件且其事 件時間比第二患者的追踪時間較爲短。CI係估計具有從模 B 型所得較高豫後計分的患者在比具有較低計分的患者更短 的時間內經歷該事件之機率且與R0C曲線下的面積（ AUC )相關聯。也可以使用其他量度來測量預測模型之能 力。例如，可以使用敏感度和特異性來評定診斷。如作爲 另一例子者，可以使用”p —値”，其代表促成，例如，觀 察到的在諸層之間的差異之單獨機會所具機率（例如，參 閱圖8，1 0和1 2 )。所以，P —値愈低，愈可能與結果有 真正統計關聯。典型地，其標準爲低於或等於〇 · 〇 5的p -値爲統計顯著者。 -49- 200538734 (46) 硏究1 .於此分析中，依序應用上述SVRc於臨床， 分子，和形態學資料，首先使用臨床特性作爲通過SVRc 對分子資料運作的’’貪進”（greedy-forward )特別選擇（ ’’ F S ”）規則系統的錨（a n c h 〇 r )。於此步驟之後，對形態 學資料運作第二SVRc貪進特性選擇規則系統，使用臨床 特性與所選分子特性的組合作爲錨。最後步驟包括對臨床 特性，所選分子特性和所選形態學特性的組合運作貪退選 擇規則系統以導出最後模型。於特性選擇之中，決定何種 特性進入（或保持）在模型中之準則係根據該特性有（或 無）增加和諧指數，亦即，增加預測性資訊而定。 經由使用內部和外部兩種確證評定模型的預測準確度 。內部確證係使用5 -倍交叉確證而進行。爲了實施外部確 證，乃從患者群創造出一系列試驗患者組且透過和諧指數 將預測結果與此等患者的實際結果相比較。於施用比二層 次確證設計中，從完整的患者記錄組隨機地選出一患者亞 組且只使用剩餘的患者採用才所述程序構建預測模型。然 後使用拒絕的記錄應用到經訓練的模型以得到預測準確性 。將此二步驟重複B次而得到B個預測率，而最後預測率 爲平均値。選用於最後模型的特性爲在B個所剝離散模型 中出現足量次數者。 使用所選特性組，經由將和諧指數直接最大化而發展 出一神經網路模型。特別者，使用上述類型的神經網路（ NNci)，其中使用實質根據和諧指數的近以之目標函數來 -50- 200538734 (47) 訓練網路。此最後模型的輸出即用來估計個別未來患^的 P S A復發風險。 硏究2 ·此硏究的目標與硏究1相同；不過’使用不 同的特性選擇和確證程序。取代錨定作法者，將所有特性 以彼等與PSA復發時間的關聯進行評等（以和諧指數測量 )且選用通過一某一預定底限値（C U 0 · 6 0 )之彼等特性 φ 。此係在我們的領域專家將影像特性數目減少，且然後在 一系列η -特性模型（如1 一特性、2 -特性、3 -特性， 等）之中評估此等特性之後完成的。使用一前向特性選擇 程序，將可使每一 η -特性模型的和諧指數最大化之特性 用於下一個η+ 1 -特性模型中。於CI不能再由預定的底 限予以改進之時，此程序即結束。然後使用後個特性選擇 程序，移除特性以期增加CI。當任何特性的移除不能改善 CI時，此程序即終止。 # 使用一種單純的啓動程式技術來選擇特性。於此作法 中，利用替換來採樣患者且用爲訓練組’且對沒有選到者 以模型評定之。作爲比較者’只有使用包括在Kattan手術 後圖解表中的彼等特性來運作此種特性選擇規則系統，如 在Kattan et al.美國專利第6,409,664號中所述者，該專 利以全文經引用方式倂於本文。使用最後模型的輸出來估 測個別未來患者的P S A復發風險。 硏究3 ·此硏究的目標爲使用硏究2中分析的相同君羊 -51 - 200538734 (48) 和特性組以及相同的特性選擇規則系統來鑑 存活率之特性。使用最後模型的輸出來估測 因任何肇因所致死亡之風險。 結果 通用作法爲應用系統病理學（形態學分 和患者臨床輪廓）對一群後***切除術的 狀態發展出用於P S A復發和整體存活率之預 要地提及者，在使用單獨地得自硏究1的臨 於標準Cox模型分析中之時，預測PS a復 有5 9 %。只有在將形態學特性和分子特性與 後，預測準測率層次才增加到8 8 %。下面諸 改進係如何達到者。 硏究1 ·對於此群中的1 3 2個患者，診 齡爲63歲（最小：40，最大：81)，且中f 爲8.2毫微克/分升（最低：1.1，最高81.9 腺切除術樣品，3 2 %具有低於少之G1 e a s ο η Gleason 7且其餘 8%爲大於 7。69個患 PT2N0M0，40 個患者（30%)爲 pT3aN0M0, 者（18% )爲 pT3bN0M0 或 ρΤ1-3Ν+。（表 究所用臨床特性之摘要表單）。 別可預測整體 個別未來患者 析、分子標記 ***癌患者 測模型。要重 床病理學特性 發因準確率只 SVRc整合之 段係說明此種 斷時的中間年 間手術後PSA )。根據前列 計分，60%爲 者（52% )爲 其餘23個患 7包含三個硏 -52- (49) 200538734 表7 .臨床資訊_ — 硏究1 硏究2和i N 132 268-47- 200538734 (44) Table 6. Description of molecular characteristics atki67t1 atki 6 7 12 atki67t3 at ki 6 7 p 1 atki 6 7 p 2 atki 6 7 p 3 atki67al atk i 6 7 a2 atki 6 7 a3 atc 1 8 13 atcd45t3 atcd68t3 at cd 3 4 p at cd 3 4 s at cd 3 41 atcd3 4tp atcd34ts atcd3 4ps atc 1 8 p 3 atcd45p3 atc18 a3 atcd45a3 ar sic 1 4 sicd 1 si psasi p am asip 2 7 si her2 si pic sicd 1 psips ap sipsm ap sip 2 7 psi her2psi ar asi c14 as i cd1asi psaasi psmaasi p 2 7 asi her2asi Ki-67, in intensity part 1 (tumor) Ki-67, in intensity part 2 (tumor) Ki- 67, in intensity part 3 (tumor) Ki-67, in intensity part 1 (PIN) Ki-67, in intensity part 2 (PIN) Ki-67, in intensity part 3 (PIN) Ki-67, In intensity part 1 (gland) Ki-6 7, in intensity part 2 (gland) Ki-67, in intensity part 3 (gland) Cytokeratin 18 (tumor) CD45 (tumor) CD68 (tumor) CD34 (PIN) CD34 (matrix) CD34 (tumor) CD34 (tumor / PIN) CD34 (tumor / matrix) CD34 (PIN / matrix) Cytokeratin 1 8 (PIN) CD45 (PIN) Cytokeratinl8 (gland) CD45 (gland) AR (tumor) staining index Cytokeratinl4 (tumor) staining index Cytokeratin D1 (tumor) staining index PSA (tumor) staining index PSMA (tumor) staining index p27 (tumor) staining index Her-2 / neu (tumor) staining index AR (PIN) staining index Cytokeratin 14 (PIN) staining index cytosine Dl (PIN) staining index PSA (PIN) staining index PSMA (PIN) staining index p27 (PIN) staining index Her-2 / neu (PIN) staining index AR (gland) staining index Cytokeratinl4 (gland) staining index Cytokeratin D1 (gland) staining index PSA (gland) staining index PSMA (gland Body) staining index p27 (gland) staining index Her-2 / neu (gland) staining index -48- 200538734 (45) Analysis and statistical studies · Three studies were performed: the initial thoughts of 1 32 patients The analysis proved (Research 1) and the expanded study with 268 patients (Research 2 and Research 3). In both Study 1 and Study 2 'The analysis consists of two steps ... Identify the characteristics that can predict the recurrence of PSA and develop a model based on these characteristics. The ultimate goal is to use this model to predict biochemistry in future patients with radical prostatectomy. (PSA) relapse. The goals of Study 3 were to identify characteristics and develop a model to predict overall survival after post-prostate resection. Use P to support vector regression (SVRc) of the type described above to develop the resulting models in each of these studies. The prediction accuracy of the model was evaluated using the Harmony Index (CI). In processing the checked results, this is often the selected metric. The Harmony Index is based on a pairwise comparison of post-yuuyu scores obtained from two randomly selected patients who meet any of the following criteria: Both patients experienced the event and the event time of the first patient was shorter than that of the second patient or only the first One patient experiences the event and the event time is shorter than the follow-up time of the second patient. The CI Department estimates that patients with higher post-hepatic scores from Model B have a greater chance of experiencing the event in a shorter time than patients with lower scores and are associated with the area under the ROC curve (AUC). Other metrics can also be used to measure the power of predictive models. For example, sensitivity and specificity can be used to assess a diagnosis. As another example, "p-値" can be used, which represents the probability of a separate opportunity contributing to, for example, the observed differences between the layers (see, for example, Figures 8, 10 and 12) . Therefore, the lower P — 値 is, the more likely it is to have a true statistical correlation with the results. Typically, p- 値 whose criterion is less than or equal to 0.05 is statistically significant. -49- 200538734 (46) Study 1. In this analysis, the above-mentioned SVRc was applied sequentially to clinical, molecular, and morphological data, and clinical characteristics were first used as the "greedy" (greedy) operation of molecular data through SVRc. -forward) Specially select ("FS") the anchor of the rule system. After this step, a second SVRc greed characteristic selection rule system is operated on the morphological data, using a combination of clinical characteristics and selected molecular characteristics as anchors. The final step involves operating a combination of clinical selection, selected molecular properties, and selected morphological properties to select a rule system to derive the final model. In the selection of characteristics, the criterion for determining which characteristics to enter (or maintain) in the model is based on the presence or absence of the characteristic to increase the harmony index, that is, to add predictive information. The predictive accuracy of the model is assessed through the use of both internal and external confirmations. Internal validation is performed using 5-fold cross-validation. To implement external validation, a series of trial patient groups were created from the patient population and the predicted results were compared with the actual results of these patients through the Harmony Index. In the two-tiered confirmation design of administration ratio, a subgroup of patients was randomly selected from the complete group of patient records and only the remaining patients were used to construct a predictive model using the procedure described above. The rejected records are then applied to the trained model to obtain prediction accuracy. These two steps are repeated B times to obtain B prediction rates, and the final prediction rate is average 値. The characteristics selected for the final model are those that appear a sufficient number of times in the B stripped scattered models. Using the selected set of features, a neural network model is developed by directly maximizing the harmony index. In particular, a neural network (NNci) of the type described above is used, in which the network is trained using a near objective function based on the harmony index -50- 200538734 (47). The output of this final model is used to estimate the risk of PSA recurrence in individual future patients. Study 2 • The purpose of this study is the same as Study 1; however, ’uses different feature selection and verification procedures. Instead of the anchoring method, all characteristics are evaluated by their correlation with PSA recurrence time (measured by the harmony index) and their characteristics φ which pass a certain predetermined threshold 値 (C U 0 · 6 0) are selected. This is done after our domain experts reduced the number of image features and then evaluated these features in a series of η-characteristic models (such as 1-characteristic, 2-characteristic, 3-characteristic, etc.). Using a forward feature selection procedure, the feature that maximizes the harmony index of each η-characteristic model is used in the next η + 1-characteristic model. This process ends when CI can no longer be improved by a predetermined threshold. Then use the latter feature selection procedure to remove features in order to increase CI. This process is terminated when the removal of any feature does not improve CI. # Use a simple launcher technique to select features. In this approach, substitutions are used to sample patients and used as a training group 'and to evaluate those who have not been selected using a model. As a 'comparator' only such characteristics selection rule system can be operated using their characteristics included in the Kattan post-operative diagram, as described in Kattan et al. US Patent No. 6,409,664, which is incorporated by reference in its entirety Suffocate in this article. The output of the last model was used to estimate the risk of PSA recurrence in individual future patients. Study 3 · The goal of this study is to identify survival characteristics using the same monarch -51-200538734 (48) and characteristic group and the same characteristic selection rule system analyzed in Study 2. Use the output of the final model to estimate the risk of death from any cause. The results are commonly used as a method for applying systematic pathology (morphology points and clinical outlines of patients) to the development of a group of postprostatectomy states as a prerequisite reference for PSA recurrence and overall survival. At the time of the analysis of the standard Cox model, the prediction of PS a was 59%. Only after the morphological and molecular characteristics are combined can the prediction accuracy level increase to 88%. How the following improvements are achieved. Study 1 · For 132 patients in this group, the age of diagnosis was 63 years (minimum: 40, maximum: 81), and the middle f was 8.2 ng / dL (minimum: 1.1, maximum 81.9 adenectomy) Of the samples, 32% had less than G1 eas ο η Gleason 7 and the remaining 8% were greater than 7. 69 patients had PT2N0M0, 40 patients (30%) were pT3aN0M0, and those (18%) were pT3bN0M0 or ρΤ1-3Ν +. (Summary of the clinical characteristics used in the study.) Do not predict the overall analysis of individual future patients, molecularly-labeled prostate cancer patient test models. It is important to understand the accuracy of the pathological characteristics, and only the SVRc integration segment indicates such a fault. PSA after mid-year surgery). According to the forefront score, 60% (52%) of the remaining 23 patients 7 contains three 硏 -52- (49) 200538734 Table 7. Clinical Information _ — Study 1 Study 2 and i N 132 268

年齡（歲） 平均 中間 範圍 種族 高加索族 西班牙 非洲美洲 未知 手術後PSA(毫微克/分升） 平均 中間 範圍 TNM階段 pT2N0 pT3aN0 pT3bN0 pTl -3N + UICC階段 Tla<5% T 1 b > 5 %Age (years) Average Middle Range Race Caucasian Spain Africa America Unknown PSA (nanograms / deciliter) Average Middle Range TNM stage pT2N0 pT3aN0 pT3bN0 pTl -3N + UICC stage Tla < 5% T 1 b > 5%

Tic不能觸診或看出 T2a< 1/2 Μ T2b< 1 葉 T2c兩葉Tic cannot palpate or see T2a < 1/2 Μ T2b < 1 leaf T2c

T3 a單側ECE T3c SV + DRE結果 不能觸診 可觸診 淋巴結涉及 陰性 陽性 精囊涉及 Μ j \ \\ 有 手術邊際 陰性 陽性 62 62 63 63 40-8 1 40-8 1 1 2 0(90.9%) 2 4 1 (89.9%) 8(6.1%) 12(4.5%) 2(1.5%) 9(3.4%) 2(1.5%) 6(2.2%) 12.2 8.2 10.8 7.8 1.1-81.9 0.9-81.9 6 9(52.3 %) 1 5 7(5 8.6%) 4 0(3 0.3 %) 7 2(26.9%) 13(9.8%) 22(8.2%) 10(7.6%) 17(6.3%) 0(0.0%) 1(0.3%) 0(0.0%) 1(0.3%) 4 9(3 7.1 %) 112(41.8%) 2 3(17.4%) 5 8(21.7%) 2 7(20.5%) 4 5(16.8%) 2 3(17.4%) 3 4(12.7%) 8(6.1%) 15(5.6%) 2(1.5%) 2(0.8%) 5 6(42.4%) 118(44.0%) 7 6(5 7.6%) 1 5 0(5 6.0%) 12 1(91.7%) 2 5 0(93.3 %) 11(8.3%) 18(6.7%) 113(85.6%) 2 3 6(8 8.0%) 19(14.4%) 3 2(12.0%) 10 8(81.8%) 2 17(81.0%) 24(18.2%) 5 1(19.0%) -53- (50)200538734 硏究 硏究2和3 織 及組 涉活 外無有性 囊顯T3 a unilateral ECE T3c SV + DRE results can not be palpated palpable lymph nodes involving negative positive seminal vesicles involved M j \ \\ have positive margin of surgery 62 62 63 63 40-8 1 40-8 1 1 2 0 ) 2 4 1 (89.9%) 8 (6.1%) 12 (4.5%) 2 (1.5%) 9 (3.4%) 2 (1.5%) 6 (2.2%) 12.2 8.2 10.8 7.8 1.1-81.9 0.9-81.9 6 9 (52.3%) 1 5 7 (5 8.6%) 4 0 (3 0.3%) 7 2 (26.9%) 13 (9.8%) 22 (8.2%) 10 (7.6%) 17 (6.3%) 0 (0.0%) 1 (0.3%) 0 (0.0%) 1 (0.3%) 4 9 (3 7.1%) 112 (41.8%) 2 3 (17.4%) 5 8 (21.7%) 2 7 (20.5%) 4 5 (16.8% ) 2 3 (17.4%) 3 4 (12.7%) 8 (6.1%) 15 (5.6%) 2 (1.5%) 2 (0.8%) 5 6 (42.4%) 118 (44.0%) 7 6 (5 7.6% ) 1 5 0 (5 6.0%) 12 1 (91.7%) 2 5 0 (93.3%) 11 (8.3%) 18 (6.7%) 113 (85.6%) 2 3 6 (8 8.0%) 19 (14.4%) 3 2 (12.0%) 10 8 (81.8%) 2 17 (81.0%) 24 (18.2%) 5 1 (19.0%) -53- (50) 200538734 Research 2 and 3 Sexual cystic manifestation

Gleason Φ 級 2 3 4 5 活組織G1 e a s ο η等級 2 3 4 5 6 7 8 9 顯性手術後Gleason等級 2 3 4 手術後Gleason等級 5 6 7 8 倍倍整中均間圍部均間 性二四非相平中範性平中圍 倍 S 倍範 比 分 百 體性 體體倍倍 比 分 百 份 7 0(5 3.0%) 1 5 9(5 9.3 %) 62(47.0%) 1 0 9(40.7%) 0(0.0%) 1(0.4%) 2 4(18.2%) 4 3(16.0%) 85(64.4%) 1 8 4(68.7%) 22(16.7%) 38(14.2%) 1(0.7%) 2(0.8%) 0(0.0%) 1(0.4%) 0(0.0%) 0(0.0%) 6(4.6%) 7(2.6%) 27(20.5%) 5 6(20.9%) 4 1(31.1%) 97(36.2%) 4 8(3 6.4%) 90(3 3.6%) 7(5.3%) 13(4.9%) 3(2.3%) 4(1.5%) 3(2.3%) 20(7.5%) 9 8(74.2%) 2 0 1 (75.0%) 3 1(23.5%) 47(17.5%) 6(4.6%) 21(7.8%) 3 6(27.3 %) 8 6(3 2.1 %) 7 9(5 9.9%) 1 4 8(5 5.2%) 10(7.6%) 12(4.5%) 1(0.8%) 4(0.4%) 7 4(5 6.1 %) 1 4 5 (54.1 %) 54(40.9%) 1 15(42.9%) 4(3.0%) 8(3.0%) 2.3 1.1 2.4 1 . 1 0.0-63.8 0.1-66.4 3.4 2.6 3.5 2.4 0.0-20.0 0.0-20.0 -54- 200538734 (51) 2 0個（1 5 % )患者經歷P S A復發，而其餘患者（8 5 % )皆爲已檢者。對於已檢患者，中間追踪時間爲6 0.8個 月、或剛好超過5年。都未到達整體中間P S A復發時間。 所有1 7種臨床特性都選用來預測PS A復發，最具資訊性 者評註如下（臨床病理學特性與模型所選用的次數）：活 體組織Gleason等級（112)、種族（112) 、UICC臨床 階段（1 1 0 )、倍性（1 1 0 )、和D RE結果（1 0 9 )。 # 影像分析和形態學硏究·圖5a和5b分別示出健康和 異常***組織的數位化影像，係根據本發明分割和分類 之後得到者。於組織中有標記出各種病理學物件供示範說 明所用。總共有496項形態學特性（上面表1所示）由影 像分析軟體所產生。 於496項形態學特性中，選出圖6中所示10項形態 學特性來預測PSA復發。所選形態學特性係關聯於下列病 理學物件，此處在特性後括號內的數字係指示出該等物件 ® 經選出以在最後模型的產生中與結果相關聯之次數：紅血 球、上皮核、腔、基質、細胞質、和組織背景（紅血球最 小像素長度（20 )、上皮核最大緊密度（1 7 )、最小包的 腔最小半徑（1 4 )、上皮核最小像素寬度（1 1 )、基質最 大密度（1 〇 )、腔最大邊際像素長度（1 0 )、上皮核最小 標準偏差通道2 ( 1 〇 )，最小包的上皮核最大半徑（1 〇 ) 、細胞質邊界像素長度的標準偏差（1 0 )、及背景像素面 積的標準偏差（i 〇 ))。更特別者，於此實施例中，紅血 球長度’腔的最小包半徑和邊界長度，細胞質邊界長度， -55- 200538734 (52) 基質密度（如，基質所覆蓋的面積之方根除以其半徑）、 及背景面積等形態學特性都經測定以與結果相關聯。上皮 核的緊密度、寬度、綠道値、和最小包的半徑等形態學特 性（例如，造出與物有相同面積的橢圓，然後放大到其完 全包住上皮核爲止，且計算出最小包圍橢圓的半徑對原橢 圓的半徑之比例）也都測定以與結果相關聯。 對於至少某些此等相關聯的數種可能原因都在上面與 P 實施例1相關處說明過。例如，上皮核緊密度之形態學特 性可能爲上皮核在周圍樣式中的”背對背”（back to back) 本質之反映，此可推測出腺體與腔形成/分化之損失，因 而與較高Gleason等級（亦即，較高的疾病進展）一致。 此外，腔最小包半徑的形態學特性係關聯於腔的整體尺寸 ，此爲隨著Gleason等級的增加而驟減和縮小。 此外，於此硏究中定出的相互關係可由下述假說至少 部份地解釋：隨著上皮核侵入基質內，上皮核典型地會變 B 得在形狀（如，具有較小變異而更圓）和尺寸（例如，面 積和邊界長度）上較不多樣化且具有較低的顏色變異。基 質的侵入也可以解釋爲何要測定基質形態學特性來與疾病 進展相關聯。特別地，癌性影像典型地係由少量基質予以 鑑定，係因爲隨著癌的進展，基質面積會被上皮細胞的細 胞質所取代之故。此會促使基質的密度質變得更高，係因 爲基質緊密度減低且在形狀上變得更非整數性（隨著物件 變形且變得更薄，物件半徑增加多於面積）。於此硏究中 測定的相關係之額外推論可爲移動經過組織的紅血球之豐 -56- 200538734 (53) 盛度可反映血管生成或新血管形成之某些量度，此可作爲 細胞離開***且在外接種的手段-因而影響PSA/BCR 復發的臨床結果-而可能關聯於疾病進展。 如上所述，可理解者，由本文所提供的講述測定之特 別形態學特性中至少有某些可與結果相關聯者可能取決於 ，例如，本發明計算形態學特性所用的特別硬體、軟體、 或彼等的組合。本文述及的Definiens Cellenger軟體和由 該軟體所測量的特別形態學特性只是範例且可以使用任何 其他的硬體，軟體或彼此的組合而不違離本發明範圍。 分子分析.由IHC評估的1 2種生物標誌物中，記錄 出總共43項獨立特性（下面的表8a，8b和8c顯示出所 觀察的生物標誌物之摘要-分子特性）。Gleason Φ grade 2 3 4 5 biopsy G1 eas ο grade 2 3 4 5 6 7 8 9 Gleason grade after dominant surgery 2 3 4 Gleason grade 5 6 7 8 times after surgery Two to four non-leveling, normality, normality, siege, percentage, body, body, body, fold, percentage, 10 0 (5 3.0%) 1 5 9 (5 9.3%) 62 (47.0%) 1 0 9 ( 40.7%) 0 (0.0%) 1 (0.4%) 2 4 (18.2%) 4 3 (16.0%) 85 (64.4%) 1 8 4 (68.7%) 22 (16.7%) 38 (14.2%) 1 (0.7 %) 2 (0.8%) 0 (0.0%) 1 (0.4%) 0 (0.0%) 0 (0.0%) 6 (4.6%) 7 (2.6%) 27 (20.5%) 5 6 (20.9%) 4 1 (31.1%) 97 (36.2%) 4 8 (3 6.4%) 90 (3 3.6%) 7 (5.3%) 13 (4.9%) 3 (2.3%) 4 (1.5%) 3 (2.3%) 20 (7.5 %) 9 8 (74.2%) 2 0 1 (75.0%) 3 1 (23.5%) 47 (17.5%) 6 (4.6%) 21 (7.8%) 3 6 (27.3%) 8 6 (3 2.1%) 7 9 (5 9.9%) 1 4 8 (5 5.2%) 10 (7.6%) 12 (4.5%) 1 (0.8%) 4 (0.4%) 7 4 (5 6.1%) 1 4 5 (54.1%) 54 ( 40.9%) 1 15 (42.9%) 4 (3.0%) 8 (3.0%) 2.3 1.1 2.4 1. .1 0.0-63.8 0.1-66.4 3.4 2.6 3.5 2.4 0.0-20.0 0.0-20.0 -54- 200538734 (51) 2 0 One (15%) patients experienced recurrence of PSA, while the remaining patients (85%) were already examined. For patients examined, the median follow-up time was 6 0.8 months, or just over 5 years. None of them reached the overall intermediate PSA recurrence time. All 17 clinical characteristics were selected to predict the recurrence of PS A. The most informative comments are as follows (clinical pathological characteristics and the number of times the model is selected): Gleason grade (112), ethnicity (112), clinical stage of UICC (1 1 0), ploidy (1 1 0), and D RE results (1 0 9). # Image analysis and morphological studies. Figures 5a and 5b show digitized images of healthy and abnormal prostate tissue, respectively, obtained after segmentation and classification according to the present invention. Various pathological objects are marked in the tissues for demonstration purposes. A total of 496 morphological characteristics (shown in Table 1 above) were generated by image analysis software. Of the 496 morphological characteristics, 10 morphological characteristics shown in Figure 6 were selected to predict PSA recurrence. The selected morphological characteristics are associated with the following pathological objects, where the numbers in parentheses after the characteristics indicate the number of times these objects have been selected to correlate with the results in the production of the final model: red blood cells, epithelium, Cavity, stroma, cytoplasm, and tissue background (minimum pixel length of red blood cells (20), maximum tightness of the epithelial nucleus (17), minimum radius of the smallest luminal cavity (14), minimum pixel width of the epithelial nucleus (1 1), matrix Maximum density (10), maximum marginal pixel length of the cavity (1 0), minimum standard deviation of epithelial nucleus channel 2 (10), maximum radius of minimum epithelial nucleus (10), standard deviation of pixel length of cytoplasmic boundary (1 0), and the standard deviation of the background pixel area (i 0)). More specifically, in this embodiment, the length of the red blood cell 'minimum packet radius and boundary length of the cavity, cytoplasmic boundary length, -55- 200538734 (52) matrix density (eg, the square of the area covered by the matrix is divided by its radius) Morphological characteristics such as,, and background area were determined to correlate with results. Morphological characteristics of the epithelial nucleus such as tightness, width, green tract, and radius of the smallest packet (for example, an ellipse with the same area as the object is created, and then it is enlarged until it completely covers the epithelial nucleus, and the minimum envelope is calculated The ratio of the radius of the ellipse to the radius of the original ellipse) was also determined to correlate with the results. Several possible causes for at least some of these are explained above in connection with P embodiment 1. For example, the morphological characteristics of tightness of the epithelial nucleus may be a reflection of the "back to back" nature of the epithelial nucleus in the surrounding pattern, which can be inferred from the loss of glandular and cavity formation / differentiation, and is therefore associated with higher Gleason Grades (ie, higher disease progression) are consistent. In addition, the morphological characteristics of the minimum package radius of the cavity are related to the overall size of the cavity, which decreases and shrinks as the Gleason level increases. In addition, the interrelationships identified in this study can be explained at least in part by the hypothesis that as the epithelial nucleus invades into the matrix, the epithelial nucleus typically becomes B in shape (eg, with smaller variations and more rounded ) And size (for example, area and border length) are less diverse and have lower color variation. Matrix invasion may also explain why matrix morphology is measured to correlate with disease progression. In particular, cancerous images are typically identified by a small amount of stroma, because the stroma area is replaced by the cytoplasm of epithelial cells as the cancer progresses. This will result in a higher density of the matrix due to the reduced compactness of the matrix and its non-integral shape (as the object deforms and becomes thinner, the radius of the object increases more than the area). An additional corollary of the relationship determined in this study may be the abundance of red blood cells moving through the tissue -56- 200538734 (53) Abundance may reflect some measure of angiogenesis or neovascularization, which can be used as the cell leaves the prostate and The means of external vaccination-thus affecting the clinical outcome of PSA / BCR relapse-may be associated with disease progression. As mentioned above, it is understandable that at least some of the special morphological characteristics of the measurement provided by this article that can be related to the results may depend on, for example, the special hardware and software used to calculate the morphological characteristics of the present invention. , Or a combination of them. The Definiens Cellenger software described herein and the special morphological characteristics measured by the software are just examples and any other hardware, software or combination of each other can be used without departing from the scope of the invention. Molecular analysis. Of the 12 biomarkers evaluated by IHC, a total of 43 independent characteristics were recorded (Tables 8a, 8b, and 8c below show the summary of the observed biomarkers-molecular characteristics).

-57- 200538734-57- 200538734

(I 保：tfe)侧缌銶(％)鏗$(+)«]涨Φ链»i:3r:1 叾8術 m i^tn 盤 0.1±0.63 0.0 0.0-6.3 100.0 土 0.00 100.0 100.0-100.0 O.OiO.OO 0.0 0.0-0.0 ΝΑ ΝΑ ΝΑ A 0.0±0.36 0.0 0.0-4.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ i 1.8±9.96 ! 0.0 0.0-96.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ δ λ 2.6±3.29 0.0 0.0-39.5 100.0±0.00 100.0 100.0-100.0 0·0±0·01 0.0 0.0-0.1 ΝΑ ΝΑ ΝΑ A 10.3121.51 0.0 0.0-100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ i 25.3±32.50 4.8 0.0-100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ 锲 Omi1 謂 λ 2.4±4.64 0.0 0.0-26.3 100.0±0.00 100.0 100.0- 100.0 0·0±0.04 0.0 0.0-0.4 0·0±0·01 0.0 0.0-0.1 A 9.8121.32 0.0 0.0-100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ i 23.9±31.38 4.7 0.0-100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ Ki-67 平均値±SD 中間値 範圍 CK18 平均値±SD 中間値 範圍 Q 2 Μ u CD68 平均値土SD 中間値 範圍 -58 - 200538734 (55)(I guarantee: tfe) side 缌 銶 (%) 铿 $ (+) «] up Φ chain» i: 3r: 1 叾 8 operation mi ^ tn disk 0.1 ± 0.63 0.0 0.0-6.3 100.0 soil 0.00 100.0 100.0-100.0 O .OiO.OO 0.0 0.0-0.0 ΝΑ ΝΑ ΝΑ A 0.0 ± 0.36 0.0 0.0-4.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ i 1.8 ± 9.96! 0.0 0.0-96.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ δ λ 2.6 ± 3.29 0.0 0.0-39.5 100.0 ± 0.00 100.0 100.0-100.0 0 · 0 ± 0 · 01 0.0 0.0-0.1 ΝΑ ΝΑ ΝΑ A 10.3121.51 0.0 0.0-100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ i 25.3 ± 32.50 4.8 0.0- 100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ 锲 Omi1 is called λ 2.4 ± 4.64 0.0 0.0-26.3 100.0 ± 0.00 100.0 100.0- 100.0 0 · 0.0 ± 0.0 0.0 0.0-0.4 0 · 0 ± 0 · 01 0.0 0.0-0.1 A 9.8121 .32 0.0 0.0-100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ i 23.9 ± 31.38 4.7 0.0-100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ Ki-67 average 値 ± SD middle 値 range CK18 average 値 SD Q 2 Μ u CD68 Average soil SD Intermediate soil range -58-200538734 (55)

(一弒窿)(％ )鏗粜寸eQus ( + )如涨靈键ΚΙ .。8嗽 PIN/基質 0·0土0.05 0.0 0.0-0.3 腫瘤/基質 0·0±0·08 0.0 0.0-0.4 腫瘤/PIN 0·0±0.06 0.0 0.0-0.5 | 腫瘤 0· 1 土0·2 1 0.0 0.0-0.9 基質 0·0±0·03 0.0 0.0-0.2 1 PIN 0·0±0·05 0.0 0.0-0.4 平均値土SD 中間値 範圍 200538734 (56) 表8b.以組織學成分的染色指數（ 0-3 00 )(硏究1)(1) (%) 铿 粜 inch eQus (+) such as the rising bond KI. 8 PIN / matrix 0 · 0 soil 0.05 0.0 0.0-0.3 tumor / matrix 0 · 0 ± 0 · 08 0.0 0.0-0.4 tumor / PIN 0 · 0 ± 0.06 0.0 0.0-0.5 | tumor 0 · 1 soil 0 · 2 1 0.0 0.0-0.9 Matrix 0 · 0 ± 0 · 03 0.0 0.0-0.2 1 PIN 0 · 0 ± 0 · 05 0.0 0.0-0.4 Mean soil SD Intermediate soil range 200538734 (56) Table 8b. Staining index by histological composition (0-3 00) (Research 1)

標誌物 腫瘤 PIN 腺體 AR 平均値± S D 179.8土71.4 64.3±75. 1 0 22.6土56.86 中間値 200 36.5 0 範圍 0-300 0-300 0-300 CK14 平均値± S D 2·6±5·83 3 1 ·2±57·35 4·7±20.42 中間値 0 0 0 範圍 0-42 0-285 0-150 胞轉蛋白D 1 平均値± S D 1 .5±5.15 (Κ0±0·27 0·0 土 0·0 中間値 0 0 0 範圍 0-33 0-3 0-0 PS A 平均値± S D 128.0土68.85 135·7±97·88 13·9±41·32 中間値 100 111 0 範圍 0-300 0-300 0-201 PSMA 平均値± S D 0·5±2·97 9·5±26.93 2.5±15.00 中間値 0 0 0 範圍 0-2 1 0-154 0-99 p27 平均値± S D 4·3±9.61 7.0±1 9.49 2·1±12·03 中間値 0 0 0 範圍 0-80 0-140 0-120 Her-2/neu 平均値± S D 4.1±1 8.50 0·1土1 ·00 0·0±0·00 中間値 0 0 0 範圍 0-146 0-100 0-0 -60- 200538734 (57) 從此等1 2種抗體，選出涵蓋1 4項特異性分子特性的 8種生物標誌物爲與PSA復發相關聯者。經更高度選出的 分子特性之某些例子經評註如下（生物標誌物- #模型選 出次數）且包括：AR染色指數—腫瘤（93 ) ，AR染色指 數—萎縮腺體（54 ) ，CD34 —腫瘤相關/PIN ( 22 ) ，Ki — 67 —腫瘤（18)與CD45— PIN相關（17)，此處PIN爲 ***上皮內贅瘤之縮寫。圖7a和7b分別示出AR和 CD34的代表性領域示表現型態。對於AR和CD34經高度 選出且十分異質的表現樣式有顯著的生物標誌物表現型態 。此等標誌物及彼等對腫瘤，萎縮腺體（對AR )和腫瘤 /PIN (對CD34)的關係可推測出影響PSA復發的臨床結 果之生物學和功能性意義。所選的選定標誌物組包括Ki -67和CD45，兩者在與AR和CD34相比時具有明顯但整 體而言較低的選出頻率。 分析和臨床硏究.使用上述領域專門知識和領域特異 性特性選擇程序，造出120無規***用以訓練（N= 100 ) 及檢驗（N = 32 )模型，最後特性組經縮減4 1種總特性， 其中有1 7種臨床特性，1 〇項形態學特性，和1 4項分子特 性。圖5顯示出所選特性的完全表單。該1 0種形態學特 性爲上述者。臨床和分子特性則在下文敘述。 臨床特性 1.活組織Gleason等級：病理學家所接受的經指配 合多重 Needle Biopsy Tissue Samples 之摘要 Gleason 等級 -61 - 200538734 (58) (福性和次級者）。發展出G 1 e a s ο η計分系統以倉U 準化’十分主觀性，可由組織學呈現出***癌體 ，產生個別等級之手段。根據腺體單位和上皮細胞 程度而將等級劃分爲1 - 5。將顯性（主要）和次顯 級）樣式加在一*起造出—Gleason Summary。此外 體分級系統中偶合也考慮整體基質緊密度，上皮細 和核特性等特性。 # 2.種族（如，非洲美洲，高加索，等） 3· UICC 階段：使用 International Union Cancer TNM編段系統以界定癌症的臨床階段，此f 腫瘤尺寸，’’N”表淋巴結涉入且”Μ”表轉移到遠端部 4 ·倍性結果：DN Α含量，其反映出在*** 細胞內的整體DNA含量。良性細胞及表現良好的 胞會以有序方式生長和***。於停止狀態中，彼等 完全組的染色體（此爲二倍體狀況）。此完全染色 ® 括來自Ma的23染色體（或N)及來自Pa的23( N )染色體（總共等於2N )。一細胞在其可***之 將其染色體的數目變爲二倍，造出兩個完全染色體 即4N，或四倍體狀態）。於***完成之後，每一 細胞接一半遺傳物質，因而再次變爲二倍體（2N) 對一組此等細胞實施PNA倍性分析，可以看到大 胞都是二倍體且其中有小部份（準備要***者）爲 。此外，在測量與製出每一細胞內的遺傳物質之量 可以看到顯性二倍體峰値及小四倍體峰値。一細 造經標 系結構 的分化 性（次 ，在整 胞尺寸 against 蠡丨’T”表 位。 癌上皮 腫瘤細 包含一 體組包 再度爲 前必須 組（亦 個新的 。若要 部份細 四倍體 圖中， 胞內的 -62- 200538734 (59) DNA量可以經由使用可結合到遺傳物質的染料將其染色而 測量。此染料的（Fuel gen染色）的濃度和分布可用影像 分析顯微術予以測量。 當腫瘤變壞時，彼等不會以彼等以往的有序方式*** 。取代靜止期所具的完全染色組者，其靜止期可能只具有 一組加半組。此等細胞所具DNA含量既不是二倍體也不 是四倍體，而是在兩者之間者。將此點細胞標繪在上述圖 i 表上，會產生在其他兩峰値之間的中途之非整倍體峰値。 硏究證實具有明顯非整倍體峰値的腫瘤表現得與不具有此 峰値者一樣地好。此係不令人訝異者，因爲在倍性狀態與 核等級之間存在著一種強烈關係之故。核等級可由對前列 腺癌有足夠經驗的任何病理學家予以評估。DNA倍性分析 所添加的値爲一種客觀性測量，可以在實驗室之間使用標 準技術予以比較且可用來對Gleason計分的近似準確度實 施快速核驗。例如，具有非整倍峰値的任何G1 e a s ο η計分 I 2 + 2 = 4或2 + 3 = 5腫瘤應該潛在地再詳估對計分的可能調整 〇 5 · DRE結果：得自數位直腸檢驗的結果（例如，陰性 或陽性），用來判定在***內以及***外擴充經由觸 診所得疾病程度。 6 ·淋巴結涉入：淋巴結含有腫瘤細胞（例如，前列 腺癌上皮細胞）的程度之量度，或經由臨床/外科檢查或 在***切除術時予以評定。 7 ·顯性活組織 G1 e a s ο η等級：參閱上面活組織 •63- 200538734 (60) G 1 e a s ο η計分之說明。此可以反映出在活組織或***切 除樣品上面看到的顯性G 1 e a s ο η分等樣式。 8 · S期中的倍性百分比：代表在細胞周期的增生期或 S期中的細胞含量部份，且反映腫瘤的生長潛勢。 9. 手術後Gleason計分：在手術後從***切除樣 品的各區採取的組織之計分。 10. TNM階段：根據***切除術後的UICC準則且 根據組織樣品的病理學檢驗之腫瘤，結和轉移。 11. 顯性手術後Gleason等級：伐***切除樣品中 所含最主要的組織學特性之顯性Gleason等級。 12. 年齢 1 3 .精囊涉入：精囊被腫瘤侵入。 14.手術前PSA :手術之前觀察到的PSA含量。 1 5 .倍性部份百分比：參閱上面對倍性結果之說明。 1 6.手術邊際涉入：手術邊際被腫瘤涉及，反映出手 術時移除腫瘤/***所在的床包含腫瘤細胞之程度。 1 7 .囊外涉入：腫瘤擴展超過***囊。 分子特性 1. AR-腫瘤：腫瘤的雄激素受體（AR)染色指數， 此爲AR染色陽性的細胞之百分比和強度的量度。針對前 列腺癌而言，染色指數可能代表在所評估的***樣品中 上皮細胞核內所偵檢到的棕色反應產物程度。 2. AR-腺體：腫瘤的AR染色指數，其係存在於腺體 -64- 200538734 (61) 結構之內者。 3. CD34—腫瘤/PIN:相對於和腫瘤/PIN相關聯的血 管內皮細胞之CD3局部化。 4· Ki67-腫瘤2:在腫瘤上皮細胞核內Ki67陽性核 之鑑別。 5· CD4 5— PIN3 :與PIN相關的CD45陽性淋巴細胞 之鑑定。 φ 6. CD34—腫瘤/基質：與腫瘤相關的CD34血管之面 部化。 7. Ki — 67-腫瘤3 ··參閱上文。 8. p27 —腫瘤：腫瘤上皮細胞核內的p27之鑑定。 9. C14 — PIN :腺體單位（上皮）基底細胞內的 cytokeratin 14 之鑑定。 10. CD34-腫瘤：相對於腫瘤相關血管的CD34局部 化。 # 11. PSA—腺體：針對腺體單位腔上皮細胞的PS A鑑 定。 12. PSMA - PIN :相對於經鑑定爲PIN的區之腺體/腔 細胞之PSMA鑑定。 13. CD34—PIN/基質：CD34對PIN相關血管之局部 化。 14. CD45 —腫瘤3 :與腫瘤相關聯的CD45陽性淋巴 細胞之鑑定。 由於在此種使用SVRc的程序中都有分析每一資料域 -65- 200538734 (62) ’因此模型的預測準確率得以增加。使用內部確證之下， 於單獨看臨床資料時，其和諧指數爲0.79。經由添加得自 分子領域的特性之下，其和諧指數增加到〇 . 8 1。經由添加 形態學特性所形成的最後模型則達到〇 . 8 4之指諧指數。 每一此等經內部確證的模型也經外部確證（如在上面”材 料和方法”中所述者），都具有所提及之傾向。對最後選 定的特性組使用NNci時，和諧指數達到〇. 8 8。 • NNci和SVRc模型所得輸出可解釋爲對一個別患者的 P S A復發相對風險估計。使用此計分的四分位數2 5 %， >25%-75%，>75% )，創造出患者風險組；表8呈現出根 據NNci模型對每一風險組所得Kaplan_Meier復發估計値 。該等組別在PSA復發時間上顯示出統計上顯著的差異（ 對數等級檢驗’ p —値<0.000 1 ) 。p —値代表促成基質之 間所觀察到的差異之單獨機會（於此等實施例中爲風險組 )的槪率。所以，p 一値愈低，更可能看到真正的統計結 © 合。通常，低於或等於0.05的任何p-値都是統計上顯著 者。 硏究2·對於此群中的268個患者（其中包含在硏究 1中分析過的1 3 2個患者中之1 2 9個），診斷時的中間年 齡爲6 3歲（最小：3 8，最大：8 1 )，且在斷根***切 除術之前的中間PS A爲7.8毫微克/分升（最低：〇.9，最 筒：8 1 · 9 )。根據***切除術樣品，4 0 %的腫瘤具有低 於7的Gleason計分，而有55%的***切除樣品具有 Gleason 7。其餘5%的***切除樣品具有大於7的 -66 - 200538734 (63)Marker Tumor PIN Gland AR Mean 値 ± SD 179.8 ± 71.4 64.3 ± 75.1 1 2 22.6 ± 56.86 Intermediate 値 200 36.5 0 Range 0-300 0-300 0-300 CK14 Mean 値 ± SD 2 · 6 ± 5 · 83 3 1 · 2 ± 57 · 35 4 · 7 ± 20.42 Intermediate 値 0 0 0 Range 0-42 0-285 0-150 Cytoprotein D 1 Mean 値 ± SD 1.5 .5 ± 5.15 (K0 ± 0.27 27 0 · 0 Soil 0 · 0 Middle 値 0 0 0 Range 0-33 0-3 0-0 PS A Average 値 ± SD 128.0 Soil 68.85 135 · 7 ± 97 · 88 13 · 9 ± 41 · 32 Middle 値 100 111 0 Range 0 -300 0-300 0-201 PSMA average 値 ± SD 0.5 · ± 2 · 97 9 · 5 ± 26.93 2.5 ± 15.00 middle 値 0 0 0 range 0-2 1 0-154 0-99 p27 average 値 ± SD 4 3 ± 9.61 7.0 ± 1 9.49 2 · 1 ± 12 · 03 Middle 値 0 0 0 Range 0-80 0-140 0-120 Her-2 / neu Mean 値 ± SD 4.1 ± 1 8.50 0 · 1 soil 1 · 00 0 · 0 ± 0 · 00 Intermediate 0 0 0 Range 0-116 0-100 0-0 -60- 200538734 (57) From these 1 or 2 antibodies, select 8 biomarkers covering 14 specific molecular characteristics It is associated with PSA recurrence. Some examples of more highly selected molecular properties are commented below (生生Markers-# model selection times) and include: AR staining index-tumor (93), AR staining index-atrophy gland (54), CD34-tumor associated / PIN (22), Ki-67-tumor (18) and CD45—PIN related (17), where PIN is an abbreviation for prostate intraepithelial neoplasia. Figures 7a and 7b show the representative fields of AR and CD34, respectively, showing the phenotype. AR and CD34 are highly selected and very heterogeneous. The expression pattern has significant biomarker phenotypes. These markers and their relationship to tumors, atrophic glands (for AR), and tumors / PIN (for CD34) can be inferred to affect the clinical outcome of PSA recurrence. And functional significance. The selected selected marker groups include Ki-67 and CD45, both of which have significant but overall lower selection frequencies when compared to AR and CD34. Analysis and clinical research. Using the above-mentioned domain expertise and domain-specific feature selection procedures, 120 random divisions were created for training (N = 100) and testing (N = 32) models. Finally, the feature set was reduced by 41 types The total characteristics include 17 clinical characteristics, 10 morphological characteristics, and 14 molecular characteristics. Figure 5 shows the full form of the selected feature. The 10 morphological characteristics are the above. Clinical and molecular characteristics are described below. Clinical characteristics 1. Gleason grade of living tissue: Summary of multi-Needle Biopsy Tissue Samples accepted by pathologists Gleason grade -61-200538734 (58) (blessed and secondary). Developed a G 1 e a s ο η scoring system based on bin U standardization, which is very subjective. It can show the prostate cancer body by histology and generate individual grades. Grades 1 to 5 are based on glandular units and the extent of epithelial cells. Adding the dominant (primary) and minor explicit styles together creates a Gleason Summary. In addition, the coupling in the body classification system also takes into account the characteristics of overall matrix compactness, epithelium fineness and nuclear characteristics. # 2. Race (eg, African America, Caucasus, etc.) 3. UICC stage: Use the International Union Cancer TNM segmentation system to define the clinical stage of the cancer, this f tumor size, "N" epidemic lymph node involvement and "M The table is transferred to the distal part 4. Ploidy results: DN A content, which reflects the overall DNA content in prostate cells. Benign cells and well-performing cells will grow and divide in an orderly manner. In the stopped state, Their complete set of chromosomes (this is a diploid condition). This fully stained® includes chromosome 23 (or N) from Ma and chromosome 23 (N) from Pa (a total of 2N). A cell can divide in it (It doubles the number of chromosomes to create two complete chromosomes, 4N, or tetraploid.) After the division is completed, each cell receives half of the genetic material, so it becomes diploid again (2N). PNA ploidy analysis was performed on a group of these cells. It can be seen that the large cells are diploid and a small part of them are ready to divide. In addition, the genetic material in each cell is measured and produced. Amount can To dominant diploid peaks and small tetraploid peaks. A fine differentiation of the standard structure of the meridian system (second, the whole cell size against '丨 T epitopes. Cancer epithelium tumors include a whole package Once again the former must be grouped (also a new one. For partial fine tetraploid plots, the intracellular -62- 200538734 (59) DNA amount can be measured by staining it with a dye that binds to genetic material. The concentration and distribution of this dye (Fuel gen staining) can be measured with image analysis microscopy. When the tumors deteriorate, they will not divide in their orderly manner in the past. It replaces the complete staining group in the stationary phase. Or, the quiescent period may have only one group plus half group. The DNA content of these cells is neither diploid nor tetraploid, but somewhere in between. The cells at this point are plotted in the above figure On the table, aneuploidy peaks appear halfway between the other two peaks. Studies have shown that tumors with significant aneuploidy peaks behave as well as those without such peaks. This Department is not surprising, because in ploidy and nuclear There is a strong relationship between the levels. The nuclear level can be evaluated by any pathologist with sufficient experience in prostate cancer. The plutonium added by DNA ploidy analysis is an objective measure and standards can be used between laboratories Techniques are compared and can be used to perform a quick check on the approximate accuracy of the Gleason score. For example, any G1 eas ο η score with an aneuploidy 値 score I 2 + 2 = 4 or 2 + 3 = 5 tumors should potentially Re-evaluate possible adjustments to scoring. DRE results: Results from a digital rectal exam (eg, negative or positive) to determine the extent of disease in the prostate as well as outside the prostate through contact with the clinic. 6. Lymph node involvement: A measure of the extent to which a lymph node contains tumor cells (for example, prostate cancer epithelial cells), or assessed through clinical / surgical examination or during prostatectomy. 7 · Dominant living tissue G1 e a s ο η grade: refer to the above description of living tissue • 63- 200538734 (60) G 1 e a s ο η scoring description. This can reflect the dominant G 1 e a s ο η grading pattern seen on biopsy or prostate resection samples. 8 · Ploidy percentage in S phase: represents the cell content fraction in the proliferative phase or S phase of the cell cycle, and reflects the growth potential of the tumor. 9. Gleason score after surgery: Score of tissue taken from various regions of the prostatectomy sample after surgery. 10. TNM stage: tumors, nodules and metastases according to UICC guidelines after prostatectomy and pathological examination of tissue samples. 11. Gleason Grade after Dominant Surgery: The dominant Gleason grade of the most important histological characteristics contained in the prostatectomy sample. 12. Nianxiu 1 3. Seminal Vesicle Involvement: The seminal vesicle was invaded by the tumor. 14. PSA before surgery: PSA content observed before surgery. 1 5. Percentage of ploidy: See the description of ploidy results above. 1 6. Margin of surgical involvement: The margin of surgery is involved by the tumor, reflecting the extent to which tumor cells are contained in the bed where the tumor / prostate was removed during surgery. 17. Extracapsular involvement: Tumor expansion beyond the prostate capsule. Molecular characteristics 1. AR-tumor: Androgen receptor (AR) staining index of the tumor. This is a measure of the percentage and intensity of AR-positive cells. For prostate cancer, the staining index may represent the extent of brown response products detected in the epithelial nucleus in the prostate samples being evaluated. 2. AR-gland: AR staining index of the tumor, which exists within the gland -64- 200538734 (61) structure. 3. CD34-Tumor / PIN: CD3 localization relative to vascular endothelial cells associated with tumor / PIN. 4. · Ki67-Tumor 2: Identification of Ki67-positive nuclei in tumor epithelial nuclei. 5. · CD4 5—PIN3: Identification of PIN-related CD45 positive lymphocytes. φ 6. CD34—tumor / stroma: Facialization of tumor-associated CD34 vessels. 7. Ki — 67-tumor 3 · See above. 8. p27-tumor: identification of p27 in the nucleus of tumor epithelial cells. 9. C14 — PIN: Identification of cytokeratin 14 in basal cells of the glandular unit (epithelium). 10. CD34-tumor: CD34 localization relative to tumor-associated blood vessels. # 11. PSA—Gland: PS A assay for glandular unit cavity epithelial cells. 12. PSMA-PIN: PSMA identification of glandular / lumen cells relative to the area identified as PIN. 13. CD34-PIN / matrix: CD34 localizes PIN-related vessels. 14. CD45-Tumor 3: Identification of CD45-positive lymphocytes associated with tumors. Since each data field is analyzed in this program using SVRc -65- 200538734 (62) ′, the prediction accuracy of the model is increased. Using internal confirmation, when looking at clinical data alone, the harmony index is 0.79. By adding properties derived from the molecular realm, its harmony index increased to 0.81. The final model formed by adding morphological characteristics reaches a finger index of 0.84. Each of these internally validated models is also externally validated (as described in "Materials and Methods" above), and has the tendencies mentioned. When NNci is used for the finally selected feature group, the harmony index reaches 0.88. • The output from the NNci and SVRc models can be interpreted as an estimate of the relative risk of PSA recurrence in another patient. Using this scoring quartile of 25%, > 25% -75%, > 75%) to create a patient risk group; Table 8 presents Kaplan_Meier recurrence estimates for each risk group based on the NNci model. . These groups showed statistically significant differences in PSA recurrence time (log-rank test 'p- 値 < 0.000 1). p-値 represents the probability of a separate opportunity (in these examples, the risk group) contributing to the observed differences between the matrices. Therefore, the lower p is, the more likely it is to see a true statistical combination. In general, any p- 値 below or equal to 0.05 is statistically significant. Study 2 · For the 268 patients in this cohort (including 1 2 of 1 2 2 patients analyzed in Study 1), the median age at diagnosis was 6 3 years (minimum: 3 8 , Max: 81), and the median PS A before the radical prostatectomy was 7.8 nanograms per deciliter (minimum: 0.9, most barrel: 81 · 9). According to prostatectomy samples, 40% of tumors had a Gleason score of less than 7 and 55% of prostatectomy samples had a Gleason 7. The remaining 5% prostatectomy samples have greater than 7 -66-200538734 (63)

Gleason計分。有 157個患者（59% )經診斷爲具有 PT2N0M0疾病，72個患者（27%)爲pT3aN0M0，且其餘 39個患者（14%)爲pT3bN0M0或pTl— 3N+。（參看上 面表5有關此群所用的所有經分析的臨床病理學特性之詳 節）。38個患者（14% )經歷PSA復發，而其餘患者（ 8 6% )都受檢過。對於已檢患者，中間追踪時間爲58.7個 月，或剛好5年以下。整體中間PSA復發時間沒有記錄。 選榀三種臨床特性來預測PSA復發：TNM臨床階段、手 術邊際、和淋巴結。 影像分析和形態學硏究·使用更新版的影像分析軟體 但分析相同的H&E染色載片，產生總共3 5 0項形態學特 性（於上面表2中顯示出）。 圖9顯示出，於3 5 0項特性中，選出6項形態學特性 來預測PSA復發，此處，此等形態學特性都關聯於上皮核 、基質、細胞質、紅白球、和腔的病理學物件（亦即，上 皮核 MinCompactne0251，基質 MaxStddevChannel30569， 細胞質 StddevMaxDiff0148，紅白球 MeanAreaPxl 03 86， 紅血球 StddevAreaPxl0388，和腔 MinAsymmetry0295)。 更特別者，於此硏究中，選出上皮核緊密度、基質的藍道 値，細胞質最大差異（例如，從細胞質的所有顏色値之最 大値減去屬於細胞質的最小平均値，再將結果除以物件亮 度），紅血球的面積、與腔的不對稱性等形態學特性來與 結果相關聯。 對於至少某些此等相關聯的各種可能理由都在上面與 -67- 200538734 (64) 實施例1及/或硏究1相關處說明過。例如，包括上皮細 胞緊密度’基質因滲入的上皮細胞而變異和破壞、及縮減 的腔尺寸之證據等形態學特性全部都提供較高Gleason等 級的組織學跡象（亦即，更高的疾病進展）。更高的 Gleason等級可推得支持轉移之更侵害性的***腫瘤及/ 或支持手術後PSA復發的腫瘤蔓延。此外，在多種結構中 的紅血之鑑別可推測血管的豐盛性。加添的血管之證據會 創造出使讓上皮細胞從***出來之可能途徑而分布到會 產生PSA的外面部位中。 圖9顯示出其於下面列出在硏究2中所選的臨床和分 子特性。此等臨床和分子特性的說明都在上面提供過。 臨床特性 1 . TNM階段 2.手術邊際涉入 • 3 .淋巴結涉入 分子特性 1. AR染色指數（腫瘤） 圖9中的每一數代表根據相應的特性及圖9中具有數 小數値的所有其他特性之預測模型和諧指數。例如，根據 TNM臨床階段、手術邊際、上皮核MinCompactness0215 、淋巴結、和基質MaxStddevChannel 3 05 69等特性的模型 之CI爲0.84 83。根據相同的5項特性加上AR染色指數 -68- 200538734 (65) (腫瘤）的模型之c I値爲〇 · 8 5 2 8。換言之’於模型中加 入AR染色指數分子特性可增加該模型的預測力。 义子分析.不需額外的免疫組織化學硏究。使用如在 β料和方法”段中所述原先收集的資料（參看附錄，表9 a 、 R u 、和9 c，有關分子特性的完整摘述）。Gleason scores. 157 patients (59%) were diagnosed with PT2N0M0 disease, 72 patients (27%) were pT3aN0M0, and the remaining 39 patients (14%) were pT3bN0M0 or pTl-3N +. (See Table 5 above for details of all the analyzed clinicopathological characteristics used in this group). Thirty-eight patients (14%) experienced recurrence of PSA, while the remaining patients (86%) had been tested. For patients examined, the median follow-up time was 58.7 months, or just under 5 years. Overall median PSA recurrence time was not recorded. Three clinical characteristics were selected to predict PSA recurrence: TNM clinical stage, surgical margin, and lymph nodes. Image analysis and morphology study • Using the updated version of image analysis software but analyzing the same H & E stained slides yielded a total of 350 morphological characteristics (shown in Table 2 above). Figure 9 shows that out of 350 characteristics, 6 morphological characteristics were selected to predict PSA recurrence. Here, these morphological characteristics are related to the pathology of the epithelial nucleus, stroma, cytoplasm, red and white balls, and cavity Objects (ie, epithelial nucleus MinCompactne0251, stroma MaxStddevChannel30569, cytoplasm StddevMaxDiff0148, red and white spheres MeanAreaPxl 03 86, red blood cells StddevAreaPxl0388, and cavity MinAsymmetry0295). More specifically, in this study, the tightness of the epithelial nucleus and the blue channel of the stroma were selected. Morphological characteristics such as the brightness of the object), the area of the red blood cells, and the asymmetry of the cavity are correlated with the results. Various possible reasons for at least some of these are explained above in connection with -67- 200538734 (64) Example 1 and / or Study 1. For example, morphological characteristics including evidence of epithelial cell tightness' matrix mutated and destroyed by infiltrating epithelial cells and evidence of reduced lumen size all provide histological signs of higher Gleason grades (i.e., higher disease progression ). Higher Gleason grades may support more aggressive prostate tumors that support metastasis and / or tumor spread that supports PSA recurrence after surgery. In addition, the identification of red blood in a variety of structures may presume the abundance of blood vessels. Evidence of added blood vessels will create a possible way for epithelial cells to exit the prostate to be distributed to the outer parts where PSA is produced. Figure 9 shows the clinical and molecular characteristics selected below in Study 2 listed below. Descriptions of these clinical and molecular properties are provided above. Clinical characteristics 1. TNM stage 2. Surgical marginal involvement • 3. Molecular features of lymph node involvement 1. AR staining index (tumor) Each number in FIG. 9 represents all according to the corresponding characteristic and the number of decimal places in FIG. 9 Prediction model harmony index of other characteristics. For example, the model based on the characteristics of TNM clinical stage, margin of surgery, epithelial nucleus MinCompactness0215, lymph nodes, and stroma MaxStddevChannel 3 05 69 has a CI of 0.84 83. Based on the same five characteristics plus AR staining index -68- 200538734 (65) (tumor), the model C I〇 was 0.85 28. In other words, adding the AR staining index molecular characteristics to the model can increase the predictive power of the model. Scent analysis. No additional immunohistochemical investigations are required. Use the data originally collected as described in the "Beta Materials and Methods" section (see Appendix, Tables 9a, Ru, and 9c for a complete excerpt of molecular properties).

-69- 200538734-69- 200538734

(£保座屍^保窿)倒缌銶(％)罌粜(+)却琛汆链#礙3^心6嗽 腺體 0.2±1.55 0.0 0.0-13.0 100.0±0.00 100.0 100.0-100.0 o.oto.oo 0.0 0.0-0.0 ΝΑ ΝΑ ΝΑ A 1.219.78 0.0 0.0-96.5 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ i 1·3±7·96 0.0 0.0-96.0 ΝΑ ΝΑ ΝΑ 1 i ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ 1 2.0±4.46 0.0 0.0-39.5 100.0 土 0·04 100.0 0.5-100.0 0·0±0·01 0.0 0.0-0.1 ΝΑ ΝΑ ΝΑ A 7.9±18.16 0.0 0.0-100.0 ΝΑ ΝΑ ΝΑ 1 ΝΑ 1 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ i 23.2±31.36 1.0 0.0-100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ 腫瘤 1.9±4.01 0.0 0.0-26.3 100·0±0.00 100.0 100.0- 100.0 0.0±0.04 0.0 0.0-0.4 0.0±0.01 0.0 0.0-0.1 A 7·3±17·04 0.0 0.0-100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ i 22.1±30.30 1.3 0.0 -100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ 標誌物 Q 1^- ¢- is 2 Q ¥1 〇〇齑鋥醒 口疮爝 U CD45 平均値±SD 中間値 範圍 CD68 平均値±SD 中間値 範圍 -70- 200538734(£ 保 座 尸 ^ 保 窿) Inverted) (%) 粜 (+) 汆 琛 汆 Chain # 33 ^ 心 6 嗽 腺 0.20.21.55 0.0 0.0-13.0 100.0 ± 0.00 100.0 100.0-100.0 o.oto .oo 0.0 0.0-0.0 ΝΑ ΝΑ ΝΑ A 1.219.78 0.0 0.0-96.5 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ i 1 · 3 ± 7 · 96 0.0 0.0-96.0 ΝΑ ΝΑ ΝΑ 1 i ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ 1 2.0 ± 4.46 0.0 0.0-39.5 100.0 Soil 0 · 04 100.0 0.5-100.0 0 · 0 ± 0 · 01 0.0 0.0-0.1 ΝΑ ΝΑ ΝΑ A 7.9 ± 18.16 0.0 0.0-100.0 ΝΑ ΝΑ 1 ΝΑ 1 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ i 23.2 ± 31.36 1.0 0.0-100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ Tumor 1.9 ± 4.01 0.0 0.0-26.3 100 · 0 ± 0.00 100.0 100.0- 100.0 0.0 ± 0.04 0.0 0.0-0.4 0.0 ± 0.01 0.0 0.0-0.1 A 7 · 3 ± 17 · 04 0.0 0.0-100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ i 22.1 ± 30.30 1.3 0.0 -100.0 ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ ΝΑ Marker Q 1 ^-¢-is 2 Q ¥ 1 〇〇齑 鋥 Aphthous ulcers U CD45 Average ± SD Intermediate range CD68 Average ± SD Intermediate range -70- 200538734

(e 奴ife^(N保窿)(％ )罌爱寸 eaus (+ )却斑 Φ链··枭 ΚΙ ·06 嗽 PIN/基質 寸 Ο 〇 ? Ο 〇 ^ +1 〇 〇 · • 〇 〇 腫瘤/基質 0·0±0·08 0.0 0.0-0.4 腫瘤/PIN 00 Ο Ο ^ Ο 〇 ° +1 ό 〇 · 〇 ° 腫瘤 0·1±0·18 0.0 0.0-0.9 基質 0·0±0.1 1 0.0 0.0-1 .7 PIN 0.0±0·04 0.0 0.0-0.4 平均値±SD 中間値 範圍 200538734 (68) 表9b. 以組織學成分的染色指數（0-3 00 )(硏究2和(e slave ife ^ (N 保 窿) (%) 罂 love inch eaus (+) but spot Φ chain ·· 枭 ΚΙ · 06 PIN / matrix inch 〇 〇 〇 〇 〇 ^ +1 〇〇 · 〇〇 tumor / Matrix 0 · 0 ± 0 · 08 0.0 0.0-0.4 tumor / PIN 00 〇 〇 ^ 〇 〇 ° +1 〇 〇 ° tumor 0 · 1 ± 0 · 18 0.0 0.0-0.9 matrix 0 · 0 ± 0.1 1 0.0 0.0-1 .7 PIN 0.0 ± 0 · 04 0.0 0.0-0.4 Mean 値 ± SD Intermediate 値 range 200538734 (68) Table 9b. Staining index (0-3 00) by histological component (Research 2 and

硏究3 ) 標誌物 腫瘤 PIN 腺體 AR 平均値± S D 172.1±75.3 79.6±82·74 28·9±67·25 中間値 200 66.0 0 範圍 0-300 0-300 0-300 CK14 平均値± S D 2·1±6·32 34.4±6 1.46 8.5±32.62 中間値 0 0 0 範圍 0-69 0-300 0-300 胞轉蛋白D 1 平均値± S D 1 ·4±6·99 0.0±0·21 0·0±0.0 中間値 0 0 0 範圍 0-90 0-3 0-0 PS A 平均値± S D 1 18·3±71·10 1 39·4±97·1 6 22·8±55.14 中間値 100 134 0 範圍 0-300 0-300 0-300 PSMA 平均値± S D 0·2±2·09 6·4±2 1.02 2·9±22·94 中間値 0 0 0 範圍 0-2 1 0-154 0-300 p27 平均値± S D 3·9±8.20 6·4±1 8.83 1 ·3±8.65 中間値 0 0 0 範圍 0-48 0-140 0-120 Her-2/neu 平均値± S D 3 ·4±16.69 0·2±1.12 0·0±0·00 中間値 0 0 0 範圍 0-150 0-10 0-0 -72- 200538734 (69) 選擇單一分子性用來預測PSA復發：AR染色指數-腫瘤。 分析與統計硏究.使用領域專業知識與啓動程式，規 則系統找到1 〇項特性之亞組（3項臨床病理學，6項形態 學，及1項分子特性），其具有〇·87之和諧指數（CI) (上面表9顯示出所選持性的完全名單）。所得SVRc模 型的輸出也可解釋爲對一個別患者的PSA復發之相對風險 P 估計。使用此計分的四分位數（<25%，<25% — 75%， >75% )，可創出患者風險組；由SVRc模型預測出的每一 風險組之Kaplan-Meier復發估計都呈現在圖10中。該等 組在PSA復發時間上顯示出統計明顯的差異（對數等級檢 驗，p —値 < 0 · 0 0 0 1 )。 硏究3 .此硏究使用與硏究2相同的群使得患者的臨 床形態學特性都相同。就結果而論，有1 9個患者（7% ) 因任何原因而死亡，而剩餘患者（93% )於彼等最後一次 I 看診與檢後仍存活。沒有達到整體中間死亡時間。選用兩 種臨床特性來預測因任何肇因之死亡：TNM臨床階段和患 者年齡。 影像分析和形態學硏究.於此硏究中使用來自硏究2 的相同組3 5 0種形態學特性。圖1 1顯示出，在3 5 0種特 性中，選用1 1種形態學特性來預測因任何肇因所致死亡 ，此處，此等特性係有關基質、紅血球、和上皮核之病理 學物件（亦即，基質 MinMeanChannel0 5 3 5、紅血球 MeanStddevChann30474、基質 MinMeanChannel20539、紅 -73- 200538734 (70) 血球 MinMeanChannel20443、紅血球Study 3) Marker tumor PIN gland AR mean ± SD 172.1 ± 75.3 79.6 ± 82 · 74 28 · 9 ± 67 · 25 middle 値 200 66.0 0 range 0-300 0-300 0-300 CK14 mean 値 ± SD 2 · 1 ± 6 · 32 34.4 ± 6 1.46 8.5 ± 32.62 Intermediate 値 0 0 0 Range 0-69 0-300 0-300 Transcellular D 1 Mean 値 ± SD 1 · 4 ± 6 · 99 0.0 ± 0 · 21 0 · 0 ± 0.0 Middle 値 0 0 0 Range 0-90 0-3 0-0 PS A Average 値 ± SD 1 18 · 3 ± 71 · 10 1 39 · 4 ± 97 · 1 6 22 · 8 ± 55.14 Middle 値100 134 0 Range 0-300 0-300 0-300 PSMA Average 値 ± SD 0 · 2 ± 2 · 09 6 · 4 ± 2 1.02 2 · 9 ± 22 · 94 Middle 値 0 0 0 Range 0-2 1 0- 154 0-300 p27 Mean 値 ± SD 3 · 9 ± 8.20 6 · 4 ± 1 8.83 1 · 3 ± 8.65 Middle 値 0 0 0 Range 0-48 0-140 0-120 Her-2 / neu Mean 値 ± SD 3 · 4 ± 16.69 0 · 2 ± 1.12 0 · 0 ± 0 · 00 Intermediate 値 0 0 0 Range 0-150 0-10 0-0 -72- 200538734 (69) Select single molecule to predict PSA recurrence: AR staining Index-tumor. Analysis and statistical research. Using domain expertise and startup programs, the rule system finds a subgroup of 10 characteristics (3 clinical pathologies, 6 morphology, and 1 molecular characteristic), which has a harmony of 0.87 Index (CI) (Table 9 above shows the full list of selected holdings). The output of the resulting SVRc model can also be interpreted as an estimate of the relative risk of PSA recurrence in another patient, P. Using this quartile of the score (< 25%, < 25%-75%, > 75%), patient risk groups can be created; Kaplan-Meier recurrence for each risk group predicted by the SVRc model Estimates are presented in Figure 10. These groups showed statistically significant differences in PSA recurrence time (log-level test, p — 値 < 0 · 0 0 0 1). Study 3. This study used the same population as Study 2 to make the patient's clinical morphological characteristics the same. In terms of results, 19 patients (7%) died for any reason, while the remaining patients (93%) survived their last I visit and examination. The overall intermediate death time was not reached. Two clinical characteristics were selected to predict death from any cause: the clinical stage of TNM and the age of the patient. Image analysis and morphological studies. The same set of 3 50 morphological characteristics from Study 2 were used in this study. Figure 11 shows that among 350 characteristics, 11 morphological characteristics were selected to predict death from any cause. Here, these characteristics are pathological objects related to the stroma, red blood cells, and epithelial nucleus. (That is, matrix MinMeanChannel0 5 3 5, red blood cells MeanStddevChann30474, matrix MinMeanChannel20539, red-73- 200538734 (70) blood cells MinMeanChannel20443, red blood cells

StddeStddeChann20472、基質 MaxMaxDiff05 2 9、上皮核 MeanBordeLengtPx10206 ' 上皮核 MeanAreaPxl0194 、上 皮核 StddevElliptFit022 8、紅血球 StddeChann3 0476、與 紅血球 StddevEllipitiFit0420，此處”channel”（波道）指 的是影像的紅（R )、綠（G )、和藍（B )色道）。更特 別者，於此硏究中，測定基質的紅色道平均値、藍色道平 均値及最大差値之形態學特性以與結果相關聯。此外也測 定紅血球的紅色道平均値和標準偏差、綠色道平均値和標 準偏差、及橢圓匹配等形態特性以與結果相關聯。爲了測 定橢圓匹配的形態學特性，乃造出具有與紅血球相同面積 的橢圓，將超出該橢圓的紅血球之面積與不能塡滿紅血球 的橢圓內部之面積相比較，且對於不能吻合的者指定〇之 値，而對完全相合的物件指定1之値。此外係測定上皮核 的邊界長度，面積和橢圓匹配等形態學特性。 對於此等相關聯中至少某一些之各種可能原因係在上 面對實施例1及/或硏究1相關聯處說明過者。例如，上 皮核的整體形狀反映出較高Gleason等級的組織學外觀。 此外，於此硏究中，針對基質的相互關聯可經由隨著癌的 進展’因與上皮細胞的間斷導致基質展現出減低的對比（ 由最大差異形態學特性所測得者）之理解予以解釋。 分子分析·於此硏究中使用得自硏究2的相同分子特 性組。選出單一特性以預測因任何肇因所致死亡：PSA染 色指數-委縮腺體。 -74- 200538734 (71) 分析與統計硏究·於此群中，選出總計1 4種特性（2 種臨床病理學特性，1 1種形態學特性，和1種分子學特性 )。最後模型具有〇 · 8 0的和諧指數（CI )。所選特性的 完全名單經顯示於圖1 1中且列於下面。所選臨床和分子 特性經列於下面。臨床特性的說明都在上面提供過。 臨床特性 • 1. TNM階段 2.年齢 分子特性 1 . psapsi :指的是在***上皮內螫瘤（PIN )中的 ***特異性抗原（PSA )之染色指數。 圖1 1中的每一數値代表根據相應特性及圖Π中具有 較小數値的所有其他特性之預測模型所具和諧指數。例如 ，根據基質 MinMeanChannell0535的模型所具 CI爲 0.0804 而根據基質 MinMeanChannell 0 5 3 5 和 TNM 時的模 型之CI爲0.73 62。 所得SVRc模型的輸出也可解釋的個別患者的死亡相 對風險估計。使用此計分的四分位數（<25%，>25%— 75% ，>75%)，造出病人風險組；圖12列出以SVRc模型預 測出的每一風險組之Kaptan-Meier復發估計値。使用對數 等級檢驗，觀測出諸風險組之間的顯著存在率差異（ ρ<0·000 1 ) 〇 -75- 200538734 (72) 結果討論（實施例2 ) 所選特性組合觀察到的從硏究 1 ( 4 1 )到硏究2 ( 1 0 )之減少而仍保持模型的預測準確性強調可透過不同機器 學習規則系統所達到的精確性與過濾屬性。於268-患者群 中發展出的模型所具和諧指數爲〇 · 8 7 ;比較之下’在對此 群應用Kattan圖解表[20]之時，達到0.78之和諧指數。 也許更驚人者爲如在硏究2中討論過的上述模型以80%敏 感度正確分類早期PSA復發患者（5年內者）之能力。比 較之下，Kattan圖解表只能以54%敏感度作出相同的預測 。此進一步強調了此等預測檢驗的作用可作爲早期介入之 決定判定。最後，所呈現的模型輸出可用來估計隨時間的 患者復發可能性，此有異於在一所給年數內而沒有指明在 該時間框內何時發生患者復發之單一槪率估計。 於硏究3中，目標在於利用既有的源自硏究2之領域 知識及開發出用於整體存活率的預測模型。成功的最終結 果爲以總共1 4項組合領域特性用80%準確度預測個體的 整體存活率和到死亡之時間。雖然受限於小事件數目（自 任何肇因之7%死亡率）及沒有可比數的已公開之圖解表 ，不過其結果確實進一步支持系統作法對於開發此等類型 的預測檢驗之用途。 另有硏究進行以將此種”整體存活率’’分析擴充到包括 利用回顧性多-組織群以獨立的外部確證硏究進行不良結 果（亦即，轉移及因***癌所致死亡）之臨床測量。此 -76 - 200538734 (73) 外，最近已起始”系統病理學”作法以詳檢診斷針對活性檢 驗以期對手術前的治療問題給予影響。 前述實施例證實已成功地開發出整合臨床特性，腫瘤 組織形態測定及分子分析之”系統病理學”平台。經由使用 領域專業知識及對已檢資料的支援向量回歸（S VRc )，從 三領域選出諸特性且用來發展出用於PSA復發與整體存活 率之預測模型。要了解者，此種新穎的”系統病理學”作法 9 在個人化醫學領域中具有廣泛應用，以其係關聯於腫瘤診 斷，患者豫後及作爲預測對特定治療的回應之工具等之故 實施例3 :***切除術之後侵害性疾病之預測 臨床和形態學資料 本硏究係進行來對已接受***切除術的患者預測其 後的侵害性疾病（亦即，由陽性骨掃描證實有呈現對骨的 輳移性***癌之臨床失敗）。於本發明之前，沒有準確 的分析工具可用來提供此種預測。如上面述及者，本發明 的系統病理學作法業已證明可準確地預測PSA復發。本硏 究證明本發明也可以用來準確地預測***切除術之後的 遠程骨轉移。 對已進行過斷根***切除術的一群1 1 9患者進行硏 究’包括從***切除術樣品構成的組織微陣列（TMAs )。使用經蘇木素和曙紅（H&E )染色的組織切片實施形 態學（亦即，影像分析）硏究，且利用一系列以彼等對前 -77- 200538734 (74) 列腺癌進展的潛在生物學相關性選出的生物標誌物以免疫 組織氏學（IHC )評定生物學決定因子。透過監督多元學 習從一選定的特性組導出一預測模型用於臨床失敗（亦即 ，陽性骨掃描）。使用經開發出以處理已檢資料的用於回 歸之支援向量機（SVRc )評定於每一領域具有完全無誤失 資料之患者（n= 11 6 )。以產生的用來界定風險組之計分 使用和諧指數（CI )估測該模型的預測效能。 從1 1 6個患者，根據彼等的臨床特性選出6 1個患者 之亞組，其中包括經骨轉移鑑定爲臨床失敗之20個體。 使用此群來造出一模型以預測在***切除術之後5年內 的陽性骨掃描可能性。選出圖1 3中所示七種特性（包括 四項臨床特性和3項形態學特性），其可用89%的準確率 及分別爲86%和85%的敏感度和特異性預測臨床失敗。所 選形態學特性係關聯於細胞質和腔的病理學物件。更特別 者，所選形態學特性爲細胞質面積除以總組織面積，腔面 積除以總組織面積，及細胞質平均紅色道的標準偏差。臨 床特性如下所列。 臨床特性 1. 囊外蔓延（ECE) 2. 精囊侵入（SVI) 3. 顯性***切除術Gleason等級（PGG1 ) 4·淋巴結侵入（LNI) 結論 -78- 200538734 (75) 將臨床特性與形態學特性整合導致第一，準確性診斷 檢驗，用於預測在***切除術後5年內的臨床失敗。如 所述者，該檢驗能以89%準確度預測那些患者最可能在前 列腺切除術後5年期間內及何時具有臨床失敗。將分子特 性加到臨床和形態學特性所得模型之結果目前正在解決中 實施例4 :肝毒物學 形態學資料 本硏究係進行來展示在毒物學領域中的影像分析與統 計模型化能力。特定言之，該硏究要求取得及分析大鼠肝 切片，整體目標爲將該等切片分類爲正常或異常。要能夠 將此方法自動化，同時達到高層次的分類準確性，可造出 高處理量的平台，用以在臨床前硏究中客觀地篩選毒性。 硏究係分成兩期。起始期係使用一組1 0 0片大鼠肝切 片作爲訓練組；80片正常肝切片和20片異常者。使用此 組切片發展出使用上述組織影像分析系統的影像分析應用 及實施特性和模型選擇以分類該等切片。然後將所建立的 影像分析程序於硏究的第二期中應用於未經標記的1 00片 大鼠肝切片組，於其中檢驗在訓練期中設計的統計模型。 分割準確度 由一病理學家的評定所得對所有物件的總分割準確度 爲 8 0 % — 9 0 %。 -79- 200538734 (76) 統計學 該硏究的統計學成分包括二步驟。第一步驟包括從切 片影像分析產生的影像資料選出特性。將分類所用特性的 數目縮減可改良切片分類的堅強性與可靠性。第二步驟包 括使用所選特性組和每一切片的標記（異常，正常）訓練 一'模型’然後經由預測一獨立組標記未知的大鼠肝切片之 分類來檢驗該模型。 特性選擇 對每一上述物件產生的統計學測量爲： 一物件數目 一相對面積（％，相對於影像總面積） -最小尺寸（以像素計） 一最大尺寸（以像素計） -平均尺寸（以像素計） -尺寸的標準偏差 由於每切片所分析的多重影像之故，此等測量本身要 經對個別大鼠肝切片的所有影像予以平均。原特性的總數 爲 3 7 8。 特性選擇也包括兩步驟。第一步驟係利用領域專業知 識。一病理學家從切片的影像分析產生的原來特性名單中 選出特性。要包括或排除特性的決定係根據對肝病理學及 可能碰到的潛在異常性/毒性之了解。從原來3 7 8項特性 -80- 200538734 (77) 的組合，使用領域知識選出9 0項特性。 然後使用逐步判別分析檢驗此等特性以減少分類所用 特性數目。構成每一類別的特性組經假設爲具有共同協方 差矩陣之多元正態。選擇特性以根據從協方差分析的F -檢驗顯著性水平進入或離開模型，此係以已選出的特性作 爲協方差且所考慮的特性爲應變數。使用0.1 5的顯著性 水平。 # -於模型中無特性之下開始逐步選擇。於每一步驟， 檢查該模型。 —若以 Wilks’ Lambda (似然準則）測量所得模型中 對該模型的判別力貢獻最小之特性不能符合留用準則，則 將該特性移除掉。 -否則，將對模型的判別力貢獻最大而不在模型內之 特性加到模型中。 一當模型中的所有特性都符合留用準則且沒有其他特 ® 性符合加入準則之時，逐步選擇程序即停止。 分類/模型訓練 然後將所選擇性進入線性判別分析（LDA )，將每一 肝切片分類爲異常或正常。透過交叉確證校正模型輸出的 潛在偏差。 也利用神經網路作爲分類器。使用所選特性作爲對神 經網路的輸入，此爲具有零隱藏單位且在輸入層與輸出層 之間有直接連接之標準多層識別（MLP )結構。經由嘗試 -81 - 200538734 (78) 互接使對RO C曲線下面積的近似最大化來訓練模型’此 要在下面解釋。經發現以此種準則訓練過的MLP模型可 達到比經由典型準則，例如，均方誤差和交叉熵（cross entropy )訓練過的MLP模型更高的準確度。 使用兩種模型的輸出經由選用不同的模型輸入値作爲 割點（cut point )，計算對每一割點的敏感度和特異性’ 及將此等標繪在二維標繪圖（y 一軸爲敏感度且X -軸爲特 異度）而造出接受者操作特性（R0C )曲線。R〇C曲線下 面積（AUC )係使用兩種量度來評估每一模型的準確性且 可解釋爲模型將肝切片正確分類爲異常或正常之能力°典 型地，敏感度和特異度係分別以真實正率與真實負率予以 描述。例如，在本硏究的範疇中，異常類別係經視爲’’正’’ 結果，而正常類別則視爲”負”結果。所以，敏感度爲真實 正率，亦即經正確分類爲異常的肝切片比例；特定度，於 另一方面，爲真實負率，亦即，經正確分類爲正常的肝切 片比例。 從R0C曲線，選自訓練組的敏感度和特異度都在下 面結果段中提出。 模型檢驗 於發展出之後，即將線型判別函數和神經網路兩者的 參數鎖定。在從試驗組大鼠肝影像接收總計學測量之後， 應用該兩分類器分別使用每一模型輸出的交互確證結果所 佔定的個別割點。該等割點係對應於未來工業級應用之 -82- 200538734 (79) 100%敏感度及90%特異度（兩者都以交叉驗證爲基礎）。 對尽此外部確證肝組的起始評估，係由對肝切片的真正分 類非外行之另一方實施模型準確性評估。之後，此另一方 也提供檢驗關鍵來驗證結果。 結果 兩模型的ROC曲線下面積都非常接近1，表示在異常 與正常肝切片之間有幾乎完美的判別。使用LDA導出的 函數具有0.99的 AUC ;使用神經網路導出的函數具有 0.98 之 AUC。 此外也在ROC曲線中觀察者爲每一模型的敏感度和 特異度，取決於施加到模型輸出的割點來分類肝切片爲異 常者或正常者。表10摘要出敏感度-特異度配對之選擇 Φ 表10 LDA N N 特異度 敏感度 特異度 敏感度 1 0 0 % 65% 1 0 0 % 65% 99% 75% 9 9% 7 0% 9 8% 1 0 0 % 9 8% 85% 檢驗 將檢驗關鍵標記與該線性判別函數的預測分類及神經 -83- 200538734 (80) 網路的預測分類相比較。根據關鍵値’將結果摘列於表 1 1 a和1 1 b中如下述： 表1 1 a 檢驗關鍵標記 異常 正常 42(TP) 1 9(FP) 7(FN) 32(TN) 4 9 5 1 LDA 標記 異常 42(TP) 19(FP) φ 正常 7(FN) 32(TN) 49 51 100 敏感度=tp/(tp + fn)x100 = 42/(42 + 7)x100 = (42/49)x100 = 86% 特異度=TN/(FP + TN)xl00 = 32/(19 + 32)xl00 = (32/51)xl00 = 63% 表1 1 b 檢驗關鍵標記 異常 正常 36(TP) 1 9(FP) 1 3(FN) 32(TN) 49 5 1StddeStddeChann20472, matrix MaxMaxDiff05 2 9, epithelial nucleus MeanBordeLengtPx10206 'epithelial nucleus MeanAreaPxl0194, epithelial nucleus StddevElliptFit022 8, red blood cell StddeChann3 0476, and red blood cell StddevEllipitiFit0420 (here, the “green” is the “channel” of the red channel) ), And blue (B) color channels). More specifically, in this study, the morphological characteristics of the average red channel average, blue channel average, and maximum differential of the matrix were determined to correlate with the results. In addition, the morphological characteristics of the red channel mean channel and standard deviation, the green channel average channel and standard deviation, and ellipse matching were also measured to correlate with the results. In order to determine the morphological characteristics of the ellipse matching, an ellipse with the same area as the red blood cells was created. The area of the red blood cells beyond the ellipse was compared with the area of the interior of the ellipse that could not fill the red blood cells.値, and specify 1 of 値 for perfectly matched objects. In addition, the morphological characteristics of the epithelial nucleus such as boundary length, area, and ellipse matching were measured. The various possible causes for at least some of these associations are explained above in connection with Embodiment 1 and / or Study 1. For example, the overall shape of the epithelial nucleus reflects the histological appearance of a higher Gleason grade. In addition, in this study, the correlation to the stroma can be explained by the understanding that as the cancer progresses' the stroma exhibits reduced contrast due to discontinuities with epithelial cells (measured by the greatest difference in morphological characteristics) . Molecular analysis. The same molecular property set from Study 2 was used in this study. A single characteristic was selected to predict death from any cause: PSA staining index-contractile glands. -74- 200538734 (71) Analysis and statistical research · In this group, a total of 14 characteristics (2 clinicopathological characteristics, 11 morphological characteristics, and 1 molecular characteristic) were selected. The final model has a Harmony Index (CI) of 0.80. A complete list of selected features is shown in Figure 11 and listed below. Selected clinical and molecular characteristics are listed below. A description of the clinical characteristics is provided above. Clinical characteristics 1. TNM stage 2. Year-old molecular characteristics 1. psapsi: Refers to the staining index of prostate-specific antigen (PSA) in prostate intraepithelial tumor (PIN). Each number in Figure 11 represents the harmony index of the prediction model based on the corresponding characteristic and all other characteristics in the figure with smaller numbers. For example, the model with the matrix MinMeanChannell0535 has a CI of 0.0804 and the model with the matrix MinMeanChannell 0 5 3 5 and TNM has a CI of 0.73 62. The output of the resulting SVRc model can also explain the relative risk estimates of death for individual patients. Using this quartile of the scores (< 25%, > 25%-75%, > 75%), a patient risk group is created; Figure 12 lists the Kaptan-Meier recurrence estimates are rampant. Using log-rank test, a significant difference in the presence of risk groups was observed (ρ < 0 · 000 1) 〇-75- 200538734 (72) Discussion of results (Example 2) The reduction from 1 (4 1) to study 2 (1 0) while still maintaining the prediction accuracy of the model emphasizes the accuracy and filtering properties that can be achieved through different machine learning rule systems. The model developed in the 268-patient group has a harmony index of 0.87; by comparison, when the Kattan chart [20] is applied to this group, a harmony index of 0.78 is reached. Perhaps more surprising is the ability of the above model, as discussed in Study 2, to correctly classify patients with early PSA relapse (within 5 years) with 80% sensitivity. In comparison, the Kattan chart can only make the same prediction with 54% sensitivity. This further emphasizes that the role of these predictive tests can be used as a decision for early intervention. Finally, the model output presented can be used to estimate the likelihood of a patient's relapse over time, which is different from a single rate estimate over a given number of years without specifying when a patient's relapse occurs within that time frame. In Study 3, the goal was to use the existing domain knowledge from Study 2 and develop a predictive model for overall survival. The end result of success is to predict the overall survival rate and time to death of individuals with 80% accuracy with a total of 14 combined domain characteristics. Although limited by the number of small events (7% mortality from any cause) and no comparable published diagrams, the results do further support the use of the system approach for the development of these types of predictive tests. Additional studies were conducted to extend this "overall survival" analysis to include the use of retrospective multi-organization groups to conduct independent external confirmation studies for adverse results (i.e., metastasis and death from prostate cancer). Clinical measurement. This -76-200538734 (73) In addition, the "system pathology" approach has recently been initiated to examine the diagnosis in detail and to test the activity test in order to influence the pre-operative treatment problems. The foregoing examples demonstrate that integration has been successfully developed "Systemic pathology" platform for clinical characteristics, tumor tissue morphology determination and molecular analysis. Through the use of field expertise and support vector regression (S VRc) of the examined data, the characteristics are selected from the three fields and used to develop A predictive model for PSA recurrence and overall survival. To understand, this novel "systemic pathology" approach 9 has been widely used in the field of personalized medicine. It is related to the diagnosis of tumors. Tools for response to specific treatments, etc. Example 3: Prediction of invasive disease after prostatectomy. Clinical and morphological data. It is performed to predict subsequent invasive disease in patients who have undergone prostatectomy (ie, clinical failure of a prostate cancer showing bone migration confirmed by a positive bone scan). Prior to the present invention, there was no accurate Analytical tools can be used to provide such predictions. As mentioned above, the systemic pathology practice of the present invention has proven to accurately predict PSA recurrence. This study proves that the present invention can also be used to accurately predict the long-term distance after prostatectomy. Bone metastases. Study of a group of 119 patients who had undergone a radical prostatectomy including tissue microarrays (TMAs) from prostatectomy samples. Stained with hematoxylin and eosin (H & E) Morphological (ie, image analysis) studies of tissue sections were performed, and a series of biomarkers selected based on their potential biological relevance for pre-77- 200538734 (74) prostate cancer progression were used to immunize tissue The biological determinants are evaluated by IHC. A supervised multivariate learning is used to derive a predictive model from a selected set of characteristics for clinical failure (ie, positive (Bone scan). Use Support Vector Machines (SVRc) developed for processing checked data to evaluate patients with complete error-free data in each area (n = 116). Use this to define risk The score of the group uses the Harmony Index (CI) to estimate the predictive power of the model. From 116 patients, a subgroup of 61 patients was selected based on their clinical characteristics, including those identified as clinically failed by bone metastasis. 20 individuals. Use this cohort to create a model to predict the likelihood of positive bone scans within 5 years after prostatectomy. Seven characteristics (including four clinical characteristics and three morphological characteristics) were selected as shown in Figure 13 ), Which can predict clinical failure with 89% accuracy and sensitivity and specificity of 86% and 85%, respectively. The selected morphological properties are pathological objects related to the cytoplasm and cavity. More specifically, the selected morphological characteristics are the cytoplasmic area divided by the total tissue area, the lumen area divided by the total tissue area, and the standard deviation of the average red tract of the cytoplasm. The clinical characteristics are listed below. Clinical characteristics 1. Extracapsular spread (ECE) 2. Seminal vesicle invasion (SVI) 3. Dominant prostatectomy Gleason grade (PGG1) 4. Lymph node invasion (LNI) Conclusion -78- 200538734 (75) The clinical characteristics and morphology The integration of features leads to the first, accurate diagnostic test for predicting clinical failure within 5 years after prostatectomy. As mentioned, this test can predict with 89% accuracy those patients who are most likely to have clinical failure within 5 years after a prostatectomy. The result of adding molecular characteristics to the model of clinical and morphological properties is currently being solved. Example 4: Liver Toxicology Morphological Data This research was conducted to demonstrate the capabilities of image analysis and statistical modeling in the field of toxicology. In particular, the study required the acquisition and analysis of rat liver sections, with the overall goal of classifying these sections as normal or abnormal. To be able to automate this method while achieving high-level classification accuracy, a high-throughput platform can be created to objectively screen for toxicity in preclinical investigations. The research department is divided into two periods. In the initial period, a group of 100 rat liver slices were used as the training group; 80 normal liver slices and 20 abnormal patients were used. Using this set of slices, image analysis applications and implementation features and model selection using the tissue imaging analysis system described above were developed to classify the slices. The established image analysis program was then applied to an unlabeled group of 100 rat liver slices in the second phase of the study, in which the statistical model designed during the training period was tested. Segmentation accuracy The total segmentation accuracy for all objects, as assessed by a pathologist, is 80%-90%. -79- 200538734 (76) Statistics The statistical component of this study consists of two steps. The first step involves selecting features from the image data generated from the slice image analysis. Reducing the number of features used for classification improves the robustness and reliability of slice classification. The second step involves training a 'model' using the selected set of characteristics and the markers (abnormal, normal) for each slice and then testing the model by predicting the classification of an independent group of labeled liver slices of unknown rats. The statistical measurement produced by the feature selection for each of the above objects is: one object number relative area (% relative to total image area)-minimum size (in pixels)-maximum size (in pixels)-average size (in (Pixel count)-The standard deviation of the size is due to the multiple images analyzed per slice, and these measurements are themselves averaged over all images of individual rat liver slices. The total number of original features is 3 7 8. Feature selection also includes two steps. The first step is to use domain expertise. A pathologist selects features from a list of original features resulting from image analysis of the slice. The decision to include or exclude characteristics is based on knowledge of liver pathology and potential abnormalities / toxicities that may be encountered. From the original combination of 3 7 8 features -80- 200538734 (77), using domain knowledge to select 90 features. These characteristics are then examined using stepwise discriminant analysis to reduce the number of characteristics used for classification. The sets of characteristics that make up each category are assumed to be multivariate normals with a common covariance matrix. The characteristics are selected to enter or leave the model based on the F-test significance level from the analysis of covariance, which uses the selected characteristics as the covariance and the characteristics considered as the strain number. A significance level of 0.1 5 was used. #-Start selecting gradually without features in the model. At each step, check the model. -If the characteristic that contributes the least to the discriminative power of the model measured by Wilks ’Lambda (likelihood criterion) does not meet the retention criterion, then remove the characteristic. -Otherwise, add to the model the characteristics that contribute the most to the discriminative power of the model and not inside the model. The step-by-step selection process stops as soon as all features in the model meet the retention criteria and no other features meet the joining criteria. Classification / model training The selected discriminant analysis (LDA) is then used to classify each liver slice as abnormal or normal. The potential bias of the model output is corrected by cross-validation. A neural network is also used as a classifier. Using the selected feature as input to the neural network, this is a standard multilayer recognition (MLP) structure with zero hidden units and a direct connection between the input and output layers. The model is trained by trying -81-200538734 (78) to maximize the approximation of the area under the RO C curve. This is explained below. It has been found that MLP models trained with such criteria can achieve higher accuracy than MLP models trained with typical criteria, such as mean square error and cross entropy. Use the output of the two models by selecting different model inputs 値 as the cut points, calculate the sensitivity and specificity for each cut point 'and plot these on a two-dimensional plot (the y-axis is sensitive) Degree and X-axis is specificity) to create a receiver operating characteristic (R0C) curve. The area under the ROC curve (AUC) uses two measures to evaluate the accuracy of each model and can be interpreted as the model's ability to correctly classify liver slices as abnormal or normal. Typically, the sensitivity and specificity are measured in terms of The true positive rate and true negative rate are described. For example, in the scope of this study, anomalous categories are treated as 'positive' results, while normal categories are treated as 'negative' results. Therefore, the sensitivity is the true positive rate, that is, the proportion of liver slices that are correctly classified as abnormal; the specificity, on the other hand, is the true negative rate, that is, the proportion of liver slices that are correctly classified as normal. From the ROC curve, the sensitivity and specificity selected from the training group are presented in the results section below. Model test After the development, the parameters of both the linear discriminant function and the neural network are locked. After receiving totalistic measurements from the liver images of the experimental group of rats, the two classifiers are used to separately confirm the individual cut points occupied by the results of the interactive verification of each model. These cut-off points correspond to -82- 200538734 (79) 100% sensitivity and 90% specificity (both based on cross-validation) for future industrial applications. The initial evaluation of this externally confirmed liver group was performed by a model other than the layman who truly classified the liver slices. After that, the other party also provides a test key to verify the results. Results The area under the ROC curves of both models were very close to 1, indicating that there was an almost perfect discrimination between abnormal and normal liver sections. Functions derived using LDA have an AUC of 0.99; functions derived using neural networks have an AUC of 0.98. In addition, the sensitivity and specificity of the observer for each model in the ROC curve depend on the cut points applied to the model output to classify liver slices as abnormal or normal. Table 10 summarizes the selection of sensitivity-specificity pairings. Table 10 LDA NN Specificity Sensitivity Specificity Sensitivity 100% 65% 1 0 0% 65% 99% 75% 9 9% 7 0% 9 8% The 100% 9 8% 85% test compares the test key signature with the predictive classification of the linear discriminant function and the predictive classification of the neural-83-200538734 (80) network. The results are summarized in Table 1 1 a and 1 1 b according to the key, as follows: Table 1 1 a Check that the key mark is abnormal 42 (TP) 1 9 (FP) 7 (FN) 32 (TN) 4 9 5 1 LDA mark abnormal 42 (TP) 19 (FP) φ normal 7 (FN) 32 (TN) 49 51 100 Sensitivity = tp / (tp + fn) x100 = 42 / (42 + 7) x100 = (42/49 ) x100 = 86% specificity = TN / (FP + TN) xl00 = 32 / (19 + 32) xl00 = (32/51) xl00 = 63% (FP) 1 3 (FN) 32 (TN) 49 5 1

NN 標記 異常 36(TP)__1 9(FP) 正常 I 13(FN) 32(TN) 49 51 100 敏感度= TP/(TP + FN)x 1 00 = 3 6/(3 6+1 3 )x 1 00 = (3 6/49)x 1 00 = 7 3% 特異度=TN/(FP + TN)xlOO = 32/(19 + 32)xlOO = (32/51)xlOO = 63°/〇 LDA分類器所用割點等於0.003 1 ; NN分類器所用割 點等於〇.〇〇〇2。兩者都對應於1〇〇%敏感度和90%特異度 之系統要求。 -84- 200538734 (81) 討論 根據每一分類器在應用於檢驗組後之敏感度和特異度 ，LDA的性能超出NN。LDA分類器達86%的敏感度’意 表此分類器可於8 6 %時間內將異常大鼠肝切片正確地標記 爲異常者，與此其異者爲達到73 %敏感度之神經網路分類 器。兩種分類器的特異度都是63 %。每一模型的敏感度和 特異度兩者都低於先前所觀察到者’但從任何分類器經推 廣到一外部組合常導致其準確度的下降來看，此係不令人 驚訝者。未硏究證實影像和統計模型化技術之成分應用。 其他具體實例 如此，可看到本發明提供方法和系統用以預測一醫療 狀況的發生。雖然在本文中已詳細揭示特別的具體實例， 但其無意對後附申請專利範圍所具範圍給予限制。特別者 ，本案發明人擬及可作出多種取代，變更和修改而不遠離 該申請專利範圍所界定的本發明旨意和範圍。其他方面， 優點與修飾都視爲在後附申請專利範圍的範圍之內。所呈 申請專利範圍爲本文所揭示的發明之代表。其他未經聲明 的發明也涵蓋在內。申請人保留將此等發明追加在以後的 申請專利範圍內之權利。 由於對於上述本發明具體實例都是，至少部分可使用 電統系統可實施者，因此要理解者，用來實施至少部份上 述方法及/或所述系統的電腦程式係可擬及爲本發明一方 -85- 200538734 (82) 面者。該電腦系統可爲任何適當的設備，系統或裝置。例 如，該電腦系統可爲一可編程資料處理設備，通用電腦， 數位信號處理器或微處理器。該電腦程式可編入源代碼且 進行編譯以在電腦上實行、或可編入，例如，物件代碼。 此外也可擬及者，歸於上述電腦程式或電腦系統的某 些或全部功能性可在硬體內執行，例如，利用一或多種應 用特異性積體電路者。 • 適當者’電腦程式可用電腦可使形式儲存在載體媒體 上，此亦擬爲本發明一方面。例如，該載體媒體可爲固態 記憶體，光學或磁-光學記憶體，例如可讀及/或可寫的光 碟例如光碟（C D )或數位多用光盤（D V A )、或磁記憶體 例如磁片或磁帶，且該電腦系統可利用程式予以配置供操 作所用。該電腦程式也可從遠方編在載體媒體中的來源予 以供給例如用電子信號，包括射頻載波或光學載波。 •參考資料 下面所引用的參考資料全部都以彼等的全文以引用方 式倂於本文： [1] Scherr D., et al., Urology. 61(2 Suppl 1):14-24,NN flag abnormal 36 (TP) __ 1 9 (FP) Normal I 13 (FN) 32 (TN) 49 51 100 Sensitivity = TP / (TP + FN) x 1 00 = 3 6 / (3 6 + 1 3) x 1 00 = (3 6/49) x 1 00 = 7 3% specificity = TN / (FP + TN) x 100 = 32 / (19 + 32) x 100 = (32/51) x 100 = 63 ° / 〇LDA classification The cut point used by the classifier is equal to 0.003 1; the cut point used by the NN classifier is equal to 0.002. Both correspond to the system requirements of 100% sensitivity and 90% specificity. -84- 200538734 (81) Discussion According to the sensitivity and specificity of each classifier after being applied to the test group, the performance of LDA exceeds NN. LDA classifier achieves sensitivity of 86% 'means that this classifier can correctly mark abnormal rat liver slices as abnormal in 86% of the time, and the difference is a neural network classification with 73% sensitivity Device. The specificity of both classifiers is 63%. The sensitivity and specificity of each model are both lower than those previously observed ', but from the perspective of any classifier being extended to an external combination that often leads to a decline in its accuracy, this system is not surprising. No research has confirmed the application of components in imaging and statistical modeling techniques. Other specific examples As such, it can be seen that the present invention provides methods and systems for predicting the occurrence of a medical condition. Although specific examples have been disclosed in detail herein, it is not intended to limit the scope of the appended patent application. In particular, the inventor of this case intends and may make various substitutions, changes and modifications without departing from the spirit and scope of the present invention as defined by the scope of the patent application. In other respects, advantages and modifications are considered to be within the scope of the appended patent application. The scope of patents presented is representative of the invention disclosed herein. Other unclaimed inventions are also covered. The applicant reserves the right to add these inventions to the scope of future patent applications. Since the above specific examples of the present invention are all those who can implement at least part of the electrical system, it should be understood that the computer program used to implement at least part of the above method and / or the system can be proposed as the present invention One-85-200538734 (82) person. The computer system may be any suitable equipment, system or device. For example, the computer system may be a programmable data processing device, a general-purpose computer, a digital signal processor, or a microprocessor. The computer program can be compiled into source code and compiled for execution on a computer, or it can be programmed into, for example, object code. It is also contemplated that some or all of the functionality attributed to the computer program or computer system described above may be implemented in hardware, for example, using one or more application-specific integrated circuits. • Where appropriate, a computer program can be stored on a carrier medium using a computer, which is also intended to be an aspect of the present invention. For example, the carrier medium may be solid state memory, optical or magneto-optical memory, such as a readable and / or writable optical disc such as a compact disc (CD) or a digital versatile disc (DVA), or a magnetic memory such as a magnetic disk or Tapes, and the computer system can be programmed for operation. The computer program can also be supplied, for example, with electronic signals, including radio frequency carriers or optical carriers, from remotely programmed sources in the carrier medium. • References The references cited below are all incorporated by reference in their entirety: [1] Scherr D., et al., Urology. 61 (2 Suppl 1): 14-24,

Feb· 2003，Swindle P.W·，et al.，Urologic Clinics of NorthFeb. 2003, Swindle P.W., et al., Urologic Clinics of North

America· 3 0(2):3 77-40 1，May 2003 〇 [2] Wahlby C., et al.，Analytical Cellular Pathology 24，101-1 1 1，2002 ° -86- 200538734 (83) [3] Street W.N·，’’Xcyt: A System for RemoteAmerica · 3 0 (2): 3 77-40 1, May 2003 〇 [2] Wahlby C., et al., Analytical Cellular Pathology 24, 101-1 1 1, 2002 ° -86- 200538734 (83) [3 ] Street WN ·, "Xcyt: A System for Remote

Cytological Diagnosis and Prognosis of Breast Cancer，’’ In Soft Computing Techniques in Breast Cancer Prognosis and Diagnosis, L.C. Jamn (ed.), CRC Press, 1 999 o [4] Gleason D.F.，’’The Veteran’s AdministrationCytological Diagnosis and Prognosis of Breast Cancer, ’’ In Soft Computing Techniques in Breast Cancer Prognosis and Diagnosis, L.C. Jamn (ed.), CRC Press, 1 999 o [4] Gleason D.F., ’’ The Veteran ’s Administration

Cooperative Urologic Research Group: Histologic Grading and Clinical Staging of Pro static Carcinoma，’’ In Urologic Pathology : The Prostate, T annenbaum M . (e d.), 171-198,Cooperative Urologic Research Group: Histologic Grading and Clinical Staging of Pro static Carcinoma, ’’ In Urologic Pathology: The Prostate, Tannenbaum M. (E d.), 171-198,

Lea and Febiger，Philadelphia, 1 977 o [5 ] C r i s t i anni e t al·，An Introduction to Support Vector Machines, C am b r i d g e Un i v e r s i t y P re s s (2 0 0 0) o [6] Hastie, The Elements of Statistical Learning, Springer (200 1 ) o [7] F.E. Harrell et al., H Evaluating the yield of medical tests，” J A M A，2 4 7 ( 1 8 ): 2 5 4 3 - 2 5 4 6，1 982 ° [8] Bishop，C·，Neural Networks for Pattern ecognition，Oxford University Press ( 1 995) o [9] Fausett, L·，Fundamentals of Neural Networks，Lea and Febiger, Philadelphia, 1 977 o [5] C risti anni et al ·, An Introduction to Support Vector Machines, C am bridge Un iversity P re ss (2 0 0 0) o [6] Hastie, The Elements of Statistical Learning, Springer (200 1) o [7] FE Harrell et al., H Evaluating the yield of medical tests, "JAMA, 2 4 7 (1 8): 2 5 4 3-2 5 4 6, 1 982 ° [ 8] Bishop, C., Neural Networks for Pattern ecognition, Oxford University Press (1 995) o [9] Fausett, L., Fundamentals of Neural Networks,

New York，Prentice Hall (1 994) 〇 [10] Definiens Cellenger Architecture: A Technical Review，April 2004 〇 [11] B aat z M . and S chape A” ” Multiresolution Segmentation-An Optimization Approach for High Quality Multi-scale Image Segmentation，’’ In Angewandte 87- 200538734 (84)New York, Prentice Hall (1 994) 〇 [10] Definiens Cellenger Architecture: A Technical Review, April 2004 〇 [11] B aat z M. And S chape A ”” Multiresolution Segmentation-An Optimization Approach for High Quality Multi-scale Image Segmentation, '' In Angewandte 87- 200538734 (84)

Geographische Informations verarbeitung XII， Strobl， J” Blaschke， T.， Griesebner， G. (eds·)， Wichmann- Verlag，Geographische Informations verarbeitung XII, Strobl, J ”Blaschke, T., Griesebner, G. (eds ·), Wichmann- Verlag,

Heidelberg，12-23，2000。 [12]Fukunaga K., Introduction to Statistical Pattern Recognition, 2nd Edition, Boston: Academic Press, 1990。 [1 3 ] Duda R. O . et al·，Pattern Classification, 2nd Edition，John Wiley & Sons Inc.，200 1 oHeidelberg, 12-23, 2000. [12] Fukunaga K., Introduction to Statistical Pattern Recognition, 2nd Edition, Boston: Academic Press, 1990. [1 3] Duda R. O. et al., Pattern Classification, 2nd Edition, John Wiley & Sons Inc., 200 1 o

[14] Holmberg L. et al” A randomized trial comparing radical prostatectomy with watchful waiting in early prostate cancer, N. Engl. M. Med·，347:781-789 (2002)。 [1 5 ] P o un d C R e t al.，Natural history of progression after PSA elevation following radical prostatectomy， JAMA 1 999, 28 1:1 59 1 - 1 597 〇 [16] Kumar- S inha C. et al” Molecular markers to identify patients at risk for recurrence after primary treatment for prostate cancer, Urology 2003; 62 Suppl. 201:19-35。 [17]Cox D.R.，” Regression Models and Life Tables，” Journal of the Royal Statistical Society, B 34,187-20,1972 [18] H arr e 11 F.E” Regression Modeling Strategies， Springer-Verlag 200 1 〇 [1 9] T uxhorn et al·，” Reactive Stroma in Human Prostate Cancer: Induction of Myofibroblast Phenotype 88· 200538734 (85) and Extracellular Matrix Remodeling” Clinical Cancer Research 29 12 Vol. 8，29 1 2-2923, S e p t e m b e r 2 0 0 2。 for for 17, [2 0] Kattan e t al·， M Postoperative Nomogram Disease Recurrence After Radical Prostatectomy Prostate Cancer,” Journal of Clinical Oncology，Vol. No. 5 (May)，1 999: pp 1 499- 1 507 °[14] Holmberg L. et al ”A randomized trial comparing radical prostatectomy with watchful waiting in early prostate cancer, N. Engl. M. Med., 347: 781-789 (2002). [1 5] Po un d CR et al., Natural history of progression after PSA elevation following radical prostatectomy, JAMA 1 999, 28 1: 1 59 1-1 597 〇 [16] Kumar- S inha C. et al ”Molecular markers to identify patients at risk for recurrence after primary treatment for prostate cancer, Urology 2003; 62 Suppl. 201: 19-35. [17] Cox DR, "Regression Models and Life Tables," Journal of the Royal Statistical Society, B 34,187-20,1972 [18] H arr e 11 FE "Regression Modeling Strategies, Springer-Verlag 200 1 〇 [1 9] Tuxhorn et al., "Reactive Stroma in Human Prostate Cancer: Induction of Myofibroblast Phenotype 88 · 200538734 (85) and Extracellular Matrix Remodeling" Clinical Cancer Research 29 12 Vol. 8, 29 1 2-2923, September 2 0 0 2 For for 17, [2 0] Kattan et al., M Postoperative Nomogram Disease Recurrence After Radical Prostatectomy Prostate Cancer, "Journal of Clinical Oncology, Vol. No. 5 (May), 1 999: pp 1 499- 1 507 °

-89- 200538734 (86) 表1.形態學特性-89- 200538734 (86) Table 1. Morphological characteristics

Script ν1·0 (496 項特性） 特性Script ν1 · 0 (496 features) Features

背景.MaxAreaPxl 背景.MeanAreaPxl 背景·MinAreaPxl 背景.StdDevAreaPxl 背景.SumAreaPxl 細胞質.Objects 細胞質.ObjectsPct 細胞質.Max AreaPxl 細胞質.MeanAreaPxl 細胞質.MinAreaPxl 細胞質.StdDevAreaPxl 細胞質.SumAreaPxl 細胞質.MaxAsymmetry 細胞質.MeanAsymmetry 細胞質.MinAsymmetry 細胞質.StdDevAsymmetry 細胞質.MaxBorderlengthPxl 細胞質.MeanBorderlengthPxl 細胞質·MinBorderlengthPxl 細胞質.StdDevBorderlengthPxl 細胞質.SumBorderlengthPxl 細胞質.MaxBrightness 細胞質.MeanBrightness 細胞質.MinBrightness 細胞質.StdDevBrightness 細胞質.MaxCompactness 細胞質.MeanCompactness 細胞質.MinCompactness 細胞質.StdDevCompactness 細胞質.MaxDensity 細胞質.MeanDensity 細胞質.MinDensity 細胞質.StdDevDensity 細胞質·MaxDiff.ofenclosing.enclo 細胞質·MeanDiff.ofenclosing.encl 細胞質.MinDiff.ofenclosing.endo 細胞質.StdDevDiff· ofenclosing· en -90 (87) 200538734 特性_ 細胞質·MaxEllipticFit 細胞質.MeanEllipticFit 細胞質.MinEllipticFit 細胞質· StdDe vEllipticFit 細胞質·MaxLengthPxl 細胞質·MeanLengthPxl 細胞質.MinLengthPxl 細胞質.StdDe vLengthPxl 細胞質.SumLengthPxl 細胞質.MaxMax.Diff.Background. MaxAreaPxl background. MeanAreaPxl background. MinAreaPxl background. StdDevAreaPxl background. SumAreaPxl cytoplasm. Objects cytoplasm. ObjectsPct cytoplasm. Max AreaPxl cytoplasm. MeanAreaPxl cytoplasm. MinAreaPxl cytoplasm. .MaxBorderlengthPxl cytoplasm .MeanBorderlengthPxl cytoplasm · MinBorderlengthPxl cytoplasm .StdDevBorderlengthPxl cytoplasm .SumBorderlengthPxl cytoplasm .MaxBrightness cytoplasm .MeanBrightness cytoplasm .MinBrightness cytoplasm .StdDevBrightness cytoplasm .MaxCompactness cytoplasm .MeanCompactness cytoplasm .MinCompactness cytoplasm .StdDevCompactness cytoplasm .MaxDensity cytoplasm .MeanDensity cytoplasm .MinDensity cytoplasm .StdDevDensity Cytoplasm MaxDiff.ofenclosing.enclo cytoplasm MeanDiff.ofenclosing.encl cytoplasm MinDiff.ofenclosing.endo cytoplasm StdDevDiff ofenclosing en -90 (87) 200538734 Characteristics _ cytoplasm MaxEllipticFit cytoplasm. MeanEl lipticFit cytoplasm.MinEllipticFit cytoplasm.StdDe vEllipticFit cytoplasm.MaxLengthPxl cytoplasm.MeanLengthPxl cytoplasm.MinLengthPxl cytoplasm.StdDe vLengthPxl cytoplasm.SumLengthPxl cytoplasm.MaxMax.Diff.

細胞質.MeanMax.Diff. 細胞質.MinMax.Diff· 細胞質.StdDevMax.Diff· 細胞質.MeanMeanChannell 細胞質.MinMeanChannell 細胞質.StdDevMeanChannell 細胞質·MaxMeanChannel2 細胞質.MeanMeanChannel2 細胞質.MinMeanChannel2 細胞質.StdDevMeanChannel2 細胞質.MaxMeanChannel3 細胞質.MeanMeanChannel3 細胞質.MinMeanChannel3 細胞質，StdDevMeanChannel3 細胞質.MaxRadiusoflargestenclose 細胞質.MeanRadiusoflargestenclos 細胞質.MinRadiusoflargestenclose 細胞質· StdDevRadiusoflargestencl 細胞質.MaxRadiusofsmallestenclos 細胞質.MeanRadiusofsmallestenclo 細胞質.MinRadiusofsmallestenclos 細胞質.StdDevRadiusofsmallestenc 細胞質·MaxStdevChannell 細胞質.MeanStdevChannell 細胞質.MinStdevChannell 細胞質.StdDevStdevChannell 細胞質·MaxStdevChannel2 細胞質.MeanStdevChannel2 細胞質.MinStdevChannel2 -91 (88) 200538734 特性_Cytoplasm.MeanMax.Diff. Cytoplasm.MinMax.Diff. Cytoplasm.StdDevMax.Diff.cytoplasm.MeanMeanChannell Cytoplasm. StdDevMeanChannel3 cytoplasm .MaxRadiusoflargestenclose cytoplasm .MeanRadiusoflargestenclos cytoplasm .MinRadiusoflargestenclose cytoplasm · StdDevRadiusoflargestencl cytoplasm .MaxRadiusofsmallestenclos cytoplasm .MeanRadiusofsmallestenclo cytoplasm .MinRadiusofsmallestenclos cytoplasm .StdDevRadiusofsmallestenc cytoplasm · MaxStdevChannell cytoplasm .MeanStdevChannell cytoplasm .MinStdevChannell cytoplasm .StdDevStdevChannell cytoplasm · MaxStdevChannel2 cytoplasm .MeanStdevChannel2 cytoplasm .MinStdevChannel2 -91 (88 200538734 Features_

細胞質.StdDevStdevChannel2 細胞質.MaxStdevChannel3 細胞質.MeanStdevChanneB 細胞質.MinStdevChanneB 細胞質.StdDevStdevChanneB 細胞質.MaxWidthPxl 細胞質.MeanWidthPxl 細胞質.MinWidthPxl 細胞質.StdDevWidthPxl 上皮核.Objects 上皮核.ObjectsPct 上皮核.MaxAreaPxl 上皮核·MeanAreaPxl 上皮核.MinAreaPxl 上皮核.StdDe v AreaPxl 上皮核.SumAreaPxl 上皮核.MaxAsymmetry 上皮核.MeanAsymmetry 上皮核.MinAsymmetry 上皮核.StdDevAsymmetry 上皮核.MaxBorderlengthPx 上皮核.MeanBorderlengthP 上皮核MinBorderlengthPx 上皮核.StdDevBorderlengt 上皮核.SumBorderlengthPx 上皮核·MaxBrightness 上皮核.MeanBrightness 上皮核.MinBrightness 上皮核.StdDevBrightness 上皮核.MaxCompactness 上皮核·MeanCompactness 上皮核.MinCompactness 上皮核.StdDevCompactness 上皮核.MaxDensity 上皮核.MeanDensity 上皮核.MinDensity 上皮核.StdDevDensity 上皮核·MaxDiff.ofenclosi 上皮核.MeanDiff.ofenclos 上皮核.MinDiff.ofenclosi -92- (89) 200538734 特性_Cytoplasm.StdDevStdevChannel2 cytoplasm.MaxStdevChannel3 cytoplasm.MeanStdevChanneB cytoplasm.MinStdevChanneB cytoplasm.StdDevStdevChanneB cytoplasm.MaxWidthPxl cytoplasm.MeanWidthPxl cytoplasm.MinWidthPxl.cytoplasm.StapDpitWidthPxl epithelial nucleus.ObjectP epithelium epithelium epithelium epithelium epithelium epithelium epithelium epithelium epithelium epithelium sac epithelium epithelium epithelium cortex StdDe v AreaPxl epithelium. SumAreaPxl epithelium. MaxAsymmetry epithelium. MeanAsymmetry epithelium. MinAsymmetry epithelium. StdDevAsymmetry epithelium. MaxBorderlengthPx epithelium. MaxeanBorderlengthP epithelium MinorderBorderlengthPx epithelium epithelium. Epithelial core. MinBrightness Epithelial core. StdDevBrightness Epithelial core. MaxCompactness Epithelial core. MeanCompactness Epithelial core. MinCompactness Epithelial core. StdDevCompactness Epithelial core. MaxDensity Epithelial core. MeanDensity Epithelial core. MinDensity Epithelial core. .ofenclos epithelial nucleus. MinDiff.ofenclosi -92- (89) 200538734 Features_

上皮核.StdDevDiff.ofencl 上皮核.MaxEllipticFit 上皮核·MeanEllipticFit 上皮核.MinEllipticFit 上皮核.StdDevEllipticFit 上皮核·MaxLengthPxl 上皮核.MeanLengthPxl 上皮核.MinLengthPxl 上皮核.StdDevLengthPxl 上皮核.SumLengtiiPxl 上皮核.MaxMax.Diff. 上皮核·MeanMax.Diff. 上皮核.MinMax.Diff. 上皮核.StdDevMax.Diff. 上皮核·MaxMeanChannell 上皮核.MeanMeanChannell 上皮核.MinMeanChannell 上皮核.StdDe vMeanChannel 上皮核·MaxMeanChannel2 上皮核.MeanMeanChannel2 上皮核.MinMeanChannel2 上皮核.StdDevMeanChannel 上皮核.MaxMeanChannel3 上皮核.MeanMeanChanneB 上皮核.MinMeanChanneB 上皮核.StdDevMeanChannel 上皮核MaxRadiusoflarges 上皮核.MeanRadiusoflarge 上皮核.MinRadiusoflarges 上皮核.StdDevRadiusoflar 上皮核.MaxRadiusofsmalle 上皮核MeanRadiusofsmall 上皮核.MinRadiusofsmalle 上皮核· StdDevRadiusofsma 上皮核.MaxStdevChannell 上皮核.MeanStdevChannell 上皮核.MinStdevChannell 上皮核.StdDevStdevChanne 上皮核.MaxStdevChannen 上皮核.MeanStdevChannel2 -93- (90) 200538734 特性_Epithelial core.StdDevDiff.ofencl Epithelial core.MaxEllipticFit Epithelial core. MeanEllipticFit Epithelial core. MinEllipticFit Epithelial core. StdDevEllipticFit Epithelial core. Nucleus MeanMax.Diff. Epithelial core.MinMax.Diff. Epithelial core.StdDevMax.Diff. Epithelial coreMaxMeanChannell Epithelial core. MeanMeanChannell Epithelial core. nuclear .StdDevMeanChannel epithelial nuclear .MaxMeanChannel3 epithelial nuclear .MeanMeanChanneB epithelial nuclear .MinMeanChanneB epithelial nuclear .StdDevMeanChannel epithelial nuclear MaxRadiusoflarges epithelial nuclear .MeanRadiusoflarge epithelial nuclear .MinRadiusoflarges epithelial nuclear .StdDevRadiusoflar epithelial nuclear .MaxRadiusofsmalle epithelial nuclear MeanRadiusofsmall epithelial nuclear .MinRadiusofsmalle epithelial epithelial nuclear · StdDevRadiusofsma Nuclear.MaxStdevChannell Epithelial Nucleus. MeanStdevChannell Epithelial Nucleus. MinStdevChannell Epithelial Nucleus. StdDevStdev Channe epithelium. MaxStdev Channen epithelium. MeanStdevChannel2 -93- (90) 200538734 Characteristics_

上皮核.MinStdevChannel2 上皮核· StdDevStdevChanne 上皮核·MaxStdevChanneB 上皮核.MeanStdevChanneB 上皮核.MinStdevChanneB 上皮核.StdDevStdevChanne 上皮核.MaxWidthPxl 上皮核.MeanWidthPxl 上皮核.MinWidthPxl 上皮核.StdDevWidthPxl 腔.Objects 腔.ObjectsPct 腔.MaxAreaPxl 腔·MeanAreaPxl 腔.MinAreaPxl 腔· StdDevAreaPxl 腔· SumAreaPxl 腔·MaxAsymmetry 腔.MeanAsymmetry 腔.MinAsymmetry 腔· StdDev Asymmetry 腔.MaxBorderlengthPxl 腔.MeanBorderlengthPxl 腔· MinBorderlengthPxl 腔· StdDevBorderlengthPxl 腔· SumBorderlengthPxl 腔.MaxBrightness 腔· MeanBrightness 腔.MinBrightness 腔.StdDe vBrightness 腔·MaxCompactness 腔· MeanCompactness 腔.MinCompactness 腔· StdDe vCompactness 腔·MaxDensity 腔.MeanDensity 腔.MinDensity 腔.StdDevDensity ^.MaxDiff.ofenclosing.enclosede 腔.MeanDiff.ofenclosing.enclosed -94 (91) 200538734 特性Epithelial nucleus. MinStdevChannel2 Epithelial nucleus StdDevStdevChanne Epithelial nucleus MaxStdevChanneB Epithelial nucleus. MeanStdevChanneB Epithelial nucleus. MinStdevChanneB Epithelial nucleus. StdDevStdevChanne Epithelial nucleus. MaxWidthPxl Epithelial nucleus. MeanWidthPxl.ObjectPxlPxD dxD Epithelium. · MeanAreaPxl chamber .MinAreaPxl cavity · StdDevAreaPxl cavity · SumAreaPxl cavity · MaxAsymmetry chamber .MeanAsymmetry chamber .MinAsymmetry cavity · StdDev Asymmetry chamber .MaxBorderlengthPxl chamber .MeanBorderlengthPxl cavity · MinBorderlengthPxl cavity · StdDevBorderlengthPxl cavity · SumBorderlengthPxl chamber .MaxBrightness cavity · MeanBrightness cavity .MinBrightness chamber. StdDe vBrightness cavity MaxCompactness cavity MeanCompactness cavity MinCompactness cavity StDDe vCompactness cavity MaxDensity cavity MeanDensity cavity MinDensity cavity StdDevDensity ^ .MaxDiff.ofenclosing.enclosede cavity.MeanDiff.closeenc

腔.MinDiff.ofenclosing.enclosede 腔.S tdDe vDiff. ofenclo sing. enclo s 腔· MaxEllipticHt 腔.MeanEllipticFit 腔.MinEllipticFit 腔.StdDevEUipticFit 腔.MaxLengthPxl 腔·MeanLengthPxl 腔·MinLengthPxl 腔· S tdDe vLengthPxl 腔.SumLengthPxl 腔.MaxMax.Diff· 腔.MeanMax.Diff. 腔.MinMax.Diff· 腔.StdDevMax.Diff. 腔.MaxMeanChannell 腔.MeanMeanChannell 腔.MinMeanChannell 腔· StdDe vMeanChannell 腔.MaxMeanChannel2 腔.MeanMeanChannel2 腔·MinMeanChannel2 腔· StdDevMeanChannel2 腔·MaxMeanChannel3 腔.MeanMeanChanneB 腔.MinMeanChannel3 腔· StdDevMeanChanneB 腔.MaxRadiusoflargestenclosedell 腔.MeanRadiusoflargestenclosedel 腔·MinRadiusoflargestenclosedell 腔.StdDevRadiusoflargestenclosed 腔.MaxRadiusofsmallestenclosinge 腔·MeanRadiusofsmallestenclosing 腔.MinRadiusofsmallestenclosinge 腔.StdDevRadiusofsmallestenclosi 腔.MaxStdevChannell 腔·Μ63π8ΐ(1ενΟιαηηε11 腔.MinStdevChannell 腔.StdDevStdevChannel 1 腔.MaxStdevChannel2 腔.MeanStdevChannel2 -95- (92) 200538734 特性_Cavity. MinDiff.ofenclosing.enclosede Cavity.S tdDe vDiff. Ofenclo sing. .Diff. Cavity.MeanMax.Diff. Cavity.MinMax.Diff. Cavity.StdDevMax.Diff. Cavity.MaxMeanChannell Cavity.MeanMeanChannell Cavity. MinMeanChannell Cavity.StdDe vMeanChannell Cavity.MaxMeanChannel2 Cavity.MeanMeanChannel2 Cavity. .MeanMeanChanneB chamber .MinMeanChannel3 cavity · StdDevMeanChanneB chamber .MaxRadiusoflargestenclosedell chamber .MeanRadiusoflargestenclosedel cavity · MinRadiusoflargestenclosedell chamber .StdDevRadiusoflargestenclosed chamber .MaxRadiusofsmallestenclosinge cavity · MeanRadiusofsmallestenclosing chamber .MinRadiusofsmallestenclosinge chamber .StdDevRadiusofsmallestenclosi chamber .MaxStdevChannell chamber · Μ63π8ΐ (1ενΟιαηηε11 chamber .MinStdevChannell chamber .StdDevStdevChannel 1 cavity .MaxStdevChannel2 Cavity. MeanStdevChannel2 -95- (92) 200538734 Features_

腔.MinStdevChannel2 腔· StdDevStdevChannel2 腔·MaxStdevChannel3 腔·MeanStdevChanneB 腔·MinStdevChanneB 腔· StdDevStdevChanneB 腔.MaxWidthPxl 腔.MeanWidthPxl 腔.MinWidthPxl 腔.StdDevWidthPxl 紅血球.Objects 紅血球· Obj ectsPct 紅血球.MaxAreaPxl 紅血球.MeanAreaPxl 紅血球·MinAreaPxl 紅血球.StdDevAreaPxl 紅血球· SumAreaPxl llJilli^.MaxAsyrnmetry 紅血球.MeanAsymmetry 紅血球.MinAsymmetry 紅血球.StdDevAsymmetry 紅血球·MaxBorderlengthPxl 紅血球.MeanBorderlengthPxl 紅血球.MinBorderlengthPxl 紅血球· StdDevBorderlengthPx 紅血球.SumBorderlengthPxl 紅血球.MaxBrightness 紅血球.MeanBrightness 紅血球.MinBrightness 紅血球.StdDevBrightness 紅血球.MaxCompactness 紅血球.MeanCompactness 紅血球.MinCompactness 紅血球.StdDevCompactness 紅血球·MaxDensity 紅血球.MeanDensity 紅血球.MinDensity 紅血球· StdDevDensity 紅血球.MaxDiff.ofenclosing· 紅血球.MeanDiff.ofenclosing 紅血球·MinDiff.ofenclosing. -96 (93) 200538734 特性_· StdDevStdevChannel2 cavity chamber chambers .MinStdevChannel2 · MaxStdevChannel3 cavity · MeanStdevChanneB cavity · MinStdevChanneB cavity · StdDevStdevChanneB chamber .MaxWidthPxl chamber .MeanWidthPxl chamber .MinWidthPxl chamber .StdDevWidthPxl red blood .Objects · Obj ectsPct red blood red blood red blood .MeanAreaPxl .MaxAreaPxl · MinAreaPxl red blood red blood red blood .StdDevAreaPxl · SumAreaPxl llJilli ^ .MaxAsyrnmetry red blood .MeanAsymmetry red blood red blood .StdDevAsymmetry .MinAsymmetry · MaxBorderlengthPxl red blood red blood red blood .MinBorderlengthPxl .MeanBorderlengthPxl · StdDevBorderlengthPx red blood red blood red blood .SumBorderlengthPxl .MaxBrightness red blood .MeanBrightness red blood red blood .StdDevBrightness .MinBrightness red blood red blood .MaxCompactness .MeanCompactness red blood cells. MinCompactness Red Blood Cell.StdDevCompactness Red Blood Cell · MaxDensity Red Blood Cell.MeanDensity Red Blood Cell.MinDensity Red Blood Cell · StdDevDensity Red Blood Cell.MaxDiff.ofenclosing · Red Blood Cell.MeanDiff.ofenclosing Red Blood Cell · MinDiff.ofenclosing. -96 (93) 200538734 Features_

紅血球.StdDevDiff.ofenclosi 紅血球.MaxEllipticFit 紅血球.MeanBllipticFit 紅血球.MinBllipticFit 紅血球.StdDevBllipticFit 紅血球.MaxLengthPxl 紅血球.MeanLengthPxl 紅血球.MinLengthPxl 紅血球.StdDevLengthPxl 紅血球.SumLengthPxl 紅血球.MaxMax.Diff. 紅血球.MeanMax.Diff. 紅血球.MinMax.Diff. 紅血球.StdDevMax.Diff. 紅血球.MaxMeanChannell 紅血球.MeanMeanChannell 紅血球.MinMeanChannell 紅血球.StdDevMeanChannell 紅血球.MaxMeanChannel〗 紅血球.MeanMeanChannel2 紅血球.MinMeanChannel〕 紅血球· StdDevMeanChannel2 紅血球.MaxMeanChanneB 紅血球.MeanMeanChanneB 紅血球.MinMeanChanneB 紅血球.StdDevMeanChannel3 紅血球·MaxRadiusoflargesten 紅血球.MeanRadiusoflargeste 紅血球.MinRadiusoflargesten 紅血球· StdDevRadiusoflarges 紅血球.MaxRadiusofsmalleste 紅血球.MeanRadiusofsmallest 紅血球.MinRadiusofsmalleste 紅血球 StdDevRadmsofsmalle 紅血球.MaxStdevChannell 紅血球.MeanStdevChannell 紅血球.MinStdevChannell 紅血球.StdDevStdevChannell 紅血球.MaxStdevChannel2 紅血球·MeanStdevChannel2 -97- (94) 200538734 特性_Red blood cells, red blood cells .MaxEllipticFit .StdDevDiff.ofenclosi red blood red blood .MeanBllipticFit .MinBllipticFit red blood red blood .MaxLengthPxl .StdDevBllipticFit red blood red blood .MeanLengthPxl .MinLengthPxl red blood red blood .SumLengthPxl .StdDevLengthPxl red blood .MaxMax.Diff. Red blood .MeanMax.Diff. Erythrocyte .MinMax.Diff. RBC .StdDevMax.Diff. .MaxMeanChannell red blood red blood red blood .MinMeanChannell .MeanMeanChannell red blood red blood .StdDevMeanChannell .MaxMeanChannel〗 red blood red blood .MeanMeanChannel2 .MinMeanChannel] · StdDevMeanChannel2 red blood red blood red blood .MaxMeanChanneB .MeanMeanChanneB red blood red blood .StdDevMeanChannel3 .MinMeanChanneB · MaxRadiusoflargesten red blood red blood .MeanRadiusoflargeste Red Blood Cells. MinRadiusoflargesten Red Blood Cells · StdDevRadiusoflarges Red Blood Cells. MaxRadiusofsmalleste Red Blood Cells. MeanRadiusofsmallest Red Blood Cells. MinRadiusofsmalleste Red Blood Cells. StdDevRadmsofsmalle Red Blood Cells. MaxStdevChannell Red Blood Cells. MeanStdevChannell Red Blood Cells. MinStdev Channell Red Blood Cell. StdDevStdevChannell Red Blood Cell. MaxStdevChannel2 Red Blood Cell MeanStdevChannel2 -97- (94) 200538734 Features_

紅血球.MinStdevChannel2 紅血球· StdDevStdevChannel2 紅血球.MaxStdevChanneB 紅血球·MeanStdevChannel3 紅血球.MinStdevChanneB 紅血球.StdDevStdevChannel3 紅血球.MaxWidthPxl 紅血球.MeanWidthPxl 紅血球.MinWidthPxl 紅血球.StdDevWidthPxl 基質.Objects 基質.ObjectsPct 基質·MaxAreaPxl 基質.MeanAreaPxl 基質.MinAreaPxl 基質.StdDevAreaPxl 基質.Sum AreaPxl 基質·MaxAsymmetry 基質· MeanAsymmetry 基質.MinAsymmetry 基質.StdDev Asymmetry 基質.MaxBorderlengthPxl 基質.MeanBorderlengthPxl 基質.MinBorderlengthPxl 基質· StdDevBorderlengthPxl 基質.SumBorderlengthPxl 基質.MaxBrightness 基質.MeanBrightness 基質.MinBrightness 基質.StdDevBrightness 基質.MaxCompactness 基質.MeanCompactness 基質.MinCompactness 基質· StdDevCompactness 基質.MaxDensity 基質.MeanDensity 基質.MinDensity 基質.StdDevDensity 基質.MaxDiff.ofenclosing.enclosed 基質·MeanDiff.ofenclosing.enclose -98 (95) 200538734 特性_· StdDevStdevChannel2 red blood red blood red blood .MinStdevChannel2 .MaxStdevChanneB · MeanStdevChannel3 red blood red blood red blood .MinStdevChanneB .StdDevStdevChannel3 red blood red blood .MeanWidthPxl .MaxWidthPxl red blood red blood .MinWidthPxl .Objects .StdDevWidthPxl matrix matrix matrix .ObjectsPct · MaxAreaPxl matrix .MeanAreaPxl matrix .MinAreaPxl matrix .StdDevAreaPxl matrix. Sum AreaPxl matrix.MaxAsymmetry matrix.MinAsymmetry matrix.StdDev Asymmetry matrix.MaxBorderlengthPxl matrix.MeanBorderlengthPxl matrix.MinBorderlengthPxl matrix. MinCompactness MatrixStdDevCompactness Matrix MaxDensity Matrix MeanDensity Matrix MinDensity Matrix StdDevDensity Matrix MaxDiff.ofenclosing.enclosed Matrix MeanDiff.ofenclosing.enclose -98 (95) 200538734 Properties_

基質.MinDiff.ofenclosing.enclosed 基質.StdDevDiff.ofenclosing.enclo 基質·MaxEllipticFit 基質.MeanEllipticFit 基質.MinEllipticFit 基質.StdDevEllipticFit 基質.MaxLengtitiPxl 基質.MeanLengthPxl 基質.MinLengthPxl 基質· StdDe vLengthPxl 基質.SumLengthPxl 基質.MaxMax.Diff. 基質.MeanMax.Diff. 基質.MinMax.Diff· 基質.StdDevMax.Diff. 基質.MaxMeanChannell 基質·MeanMeanChannell 基質.MinMeanChannell 基質.StdDe vMeanChannell 基質.MaxMeanChannel2 基質.MeanMeanChannel2 基質.MinMeanChannel2 基質.StdDevMeanChannel2 基質.MaxMeanChannel3 基質.MeanMeanChannel3 基質.MinMeanChannel3 基質.StdDevMeanChannel3 基質.MaxRadiusoflargestenclosedel 基質.MeanRadiusoflargestenclosede 基質.MinRadiusoflargestenclosedel 基質.StdDevRadiusoflargestenclose 基質.MaxRadiusofsmallestenclosing 基質.MeanRadiusofsmallestenclosin 基質.MinRadiusofsmallestenclosing 基質.StdDevRadiusofsmallestenclos 基質·MaxStdevChannell 基質.MeanStdevChannell 基質.MinStdevChannell 基質.StdDevStdevChannell 基質·MaxStdevChannel2 -99 (96) 200538734 特性_Matrix.MinDiff.ofenclosing.enclosed matrix.StdDevDiff.ofenclosing.enclo matrix.MaxEllipticFit matrix. MeanEllipticFit matrix.MinEllipticFit matrix.StdDevEllipticFit matrix.MeanLxl.MaxPengthLengthLengthLengthPxLengthLengthPin MeanMax.Diff.Matrix.MinMax.Diff.Matrix.StdDevMax.Diff.Matrix.MaxMeanChannell Matrix.MeanMeanChannell Matrix.MinMeanChannell Matrix.StdDe vMeanChannell Matrix.MaxMeanChannel2 Matrix.MeanMeanChannel2 Matrix.MinMeanChannel2 Matrix.St2DMevMean. Substrate.StdDevMeanChannel3 Substrate.MaxRadiusoflargestenclosedel Substrate.MeanRadiusoflargestenclosede Substrate.MinRadiusoflargestenclosedel Substrate.StdDevRadiusoflargestenclose Substrate.MaxRadiusofsmallestenclosing Substrate.MeanRadiusofsmallest.StevestingMaximc. evChannell matrix. StdDevStdevChannell matrixMaxStdevChannel2 -99 (96) 200538734 Features_

基質.MeanStdevChannel2 基質· MinStde vChannel2 基質.StdDe vStde vChannel2 基質.MaxStdevChanneB 基質.MeanStdevChannel3 基質.MinStdevChanneB 基質.StdDevStde vChannel3 基質.MaxWidthPxl 基質.MeanWidthPxl 基質.MinWidthPxl 基質.StdDevWidthPxl 基質核.Objects 基質核.ObjectsPct 基質核.MaxAreaPxl 基質核.MeanAreaPxl 基質核.MinAreaPxl 基質核.StdDevAreaPxl 基質核.SumAreaPxl 基質核.MaxAsymmetry 基質核·MeanAsymmetry 基質核.MinAsymmetry 基質核.StdDevAsymmetry 基質核.MaxBorderlengthPxl 基質核.MeanBorderlengthPxl 基質核.MinBorderlengthPxl 基質核.StdDevBorderlengthPxl 基質核.SumBorderlengthPxl 基質核.MaxBrightness 基質核.MeanBrightness 基質核.MinBrightness 基質核.StdDevBrightness 基質核·MaxCompactness 基質核·MeanCompactness 基質核.MinCompactness 基質核.StdDevCompactness 基質核.MaxDensity 基質核·MeanDensity 基質核.MinDensity 基質核.StdDevDensity 基質核.MaxDiff.ofenclosing.e -100 (97) 200538734 特性_Matrix. MeanStdevChannel2 matrixMinStde vChannel2 matrix. StdDe vStde vChannel2 matrix. MaxStdevChanneB matrix. MeanStdevChannel3 matrix. MinStdevChanneB matrix. StdDevStde vChannel3 matrix. MaxWidthPxl matrix. MeanWidthPxl matrix. MinWidthPxl matrix. CoreWidth.xPxl matrix. Nucleus. MeanAreaPxl matrix core. MinAreaPxl matrix core. StdDevAreaPxl matrix core. SumAreaPxl matrix core. MaxAsymmetry matrix core. MeanAsymmetry matrix core. MinAsymmetry matrix core. StdDevAsymmetry matrix core. MaxBorderlengthPxl matrix core. Plexi.x.D. SumBorderlengthPxl matrix core. MaxBrightness matrix core. MeanBrightness matrix core. MinBrightness matrix core. StdDevBrightness matrix core. MaxCompactness matrix core. MeanCompactness matrix core. MinCompactness matrix core. StdDevCompactness matrix core. MaxDensity matrix core. NuclearDensDityDensDity Nuclear.MaxDiff.ofenclosing.e -100 (97) 200538734 Features_

基質核.MeanDiff ofenclosing. 基質核.MinDiff.ofenclosing.e 基質核.StdDevDiff.ofenclosin 基質核·MaxBllipticFit 基質核.MeanEllipticFit 基質核.MinEllipticFit 基質核· StdDe vEllipticFit 基質核.MaxLengthPxl 基質核.MeanLengthPxl 基質核.MinLengthPxl 基質核.StdDevLengthPxl 基質核.SumLengthPxl 基質核.MaxMax.Diff· 基質核.MeanMax.Diff. 基質核.MinMax.Diff. 基質核.StdDevMax.Diff. 基質核.MaxMeanChannell 基質核.MeanMeanChannell 基質核.MinMeanChannell 基質核.StdDevMeanChannell 基質核.MaxMeanChannel2 基質核.MeanMeanChannel2 基質核.MinMeanChannel2 基質核.StdDevMeanChannel2 基質核.MaxMeanChanneB 基質核.MeanMeanChanneB 基質核.MinMeanChannel3 基質核.StdDevMeanChannel3 基質核.MaxRadiusoflargestenc 基質核·MeanRadiusoflargesten 基質核.MinRadiusoflargestenc 基質核.StdDevRadiusoflargest 基質核.MaxRadiusofsmallesten 基質核.MeanRadiusofsmalleste 基質核.MinRadiusofsmallesten 基質核.StdOevRadiusofsinalles 基質核.MaxStdevChannell 基質核.MeanStdevChannell 基質核.MinStdevChannell 基質核.StdDevStdevChannell 基質核·MaxStdevChannel2 基質核.MeanStdevChannel2 基質核.MinStdevChannel2 基質核.StdDevStdevChannel2 -101 - (98) 200538734 特性_ 基質核.MaxStdevChannel3Matrix nucleus. MeanDiff of enclosing. Matrix nucleus. MinDiff.ofenclosing.e Matrix nucleus. StdDevDiff.ofenclosin Matrix nucleus. MaxBllipticFit Matrix nucleus. MeanEllipticFit Matrix nucleus. MinEllipticFit Matrix nucleus. StdDe vEllipticFit matrix nucleus P.xLength MxLength. Core. StdDevLengthPxl Matrix Core. SumLengthPxl Matrix Core. MaxMax.Diff. Matrix Core. MeanMax.Diff. Matrix Core. MinMax.Diff. Matrix Core. StdDevMax.Diff. Matrix Core. MaxMeanChannell Matrix Core. MeanMeanChannell Matrix Core. MinMeanChannell. StdDevMeanChannell Matrix Core.MaxMeanChannel2 Matrix Core. MeanMeanChannel2 Matrix Core.MinMeanChannel2 Matrix Core.StdDevMeanChannel2 Matrix Core.MaxMeanChanneB Matrix Core.MeanMeanChanneB Matrix Core.MinMeanChannel3 Matrix Core.StdDevMeanChannel3 Matrix Core.Radium Core LargeRadRadius. Nucleus. MaxRadiusofsmallesten matrix nucleus. MeanRadiusofsmalleste matrix nucleus. MinRadiusofsmallesten matrix nucleus. StdOevRadiusofsinalles Matrix core. MaxStdevChannell Matrix core. MeanStdevChannell Matrix core. MinStdevChannell Matrix core. StdDevStdevChannell Matrix core. MaxStdevChannel2 Matrix core. MeanStdevChannel2 Matrix core. MinStdevChannel2 Matrix core. StdDevStdevChannel2 -101-(98) 200538734 Features_ Matrix core.

基質核.MeanStdevChanneBStromal nucleus. MeanStdevChanneB

基質核.MinStdevChanneB 基質核.StdDevStdevChannel3 基質核.MaxWidthPxl 基質核.MeanWidthPxl 基質核.MinWidthPxl 基質核.StdDevWidthPxlMatrix nucleus. MinStdevChanneB Matrix nucleus. StdDevStdevChannel3 Matrix nucleus. MaxWidthPxl Matrix nucleus. MeanWidthPxl Matrix nucleus. MinWidthPxl Matrix nucleus. StdDevWidthPxl

C2ENC2EN

EN2SN L2CoreEN2SN L2Core

C2LC2L

CEN2LCEN2L

表2.形態學特性Table 2. Morphological characteristics

Script ν2·0 (350 特性） 特性 橋作物 Mean Area Pxl 橋作物 StdDev Area Pxl 矯作物 Mean Asymmetry 橋作物 StdDev Asymmetry 橋作物 Mean Border index 橋作物 StdDev Border index 橋作物 Mean Border length Pxl 橋作物 StdDev Border length Pxl 橋作物 Mean Brightness 橋作物 StdDev Brightness 橋作物 Mean Compactness 橋作物 StdDev Compactness 橋作物 Mean Density 橋作物 StdDev Density 橋作物 Mean Diff. of enclosing/enclosed ellipse 橋作物 StdDev Diff. of enclosing/enclosed ellipse 橋作物 Mean Elliptic Fit 橋作物 StdDev Elliptic Fit 橋作物 Mean Length Pxl 橋作物 StdDev Length Pxl 橋作物 Mean Length/width 矯作物 StdDev Length/width 橋作物 Mean Main direction 橋作物 StdDev Main direction 矯作物 Mean Max.Diff. 矯作物 StdDev Max.Diff. 橋作物 Mean Mean Channel 1 矯作物 StdDev Mean Channel 1 矯作物 Mean Mean Channel 2 橋作物 StdDev Mean Channel 2 -102- (99) 200538734 特性_ 橋作物 Mean Mean Channel 3 矯作物 StdDev Mean Channel 3 矯作物 Mean Radius of largest enclosed ellipse 矯作物 StdDev Radius of largest enclosed ellipse 橋作物 Mean Radius of smallest enclosing ellipse 橋作物 StdDev Radius of smallest enclosing ellipse 矯作物 Mean Rectangular Fit 矯作物 StdDev Rectangular Fit 橋作物 Mean Shape index 矯作物 StdDev Shape index 橋作物 Mean Stddev Channell 橋作物 StdDev Stddev Channell 矯作物 Mean Stddev Channel2Script ν2 · 0 (350 characteristics) Characteristics Bridge crop Mean Area Pxl Bridge crop StdDev Area Pxl Correction crop Mean Asymmetry Bridge crop StdDev Asymmetry Bridge crop Mean Border index Bridge crop StdDev Border index Bridge crop Mean Border length Pxl Bridge crop StdDev Border length Pxl Bridge Crop Mean Brightness Bridge Crop StdDev Brightness Bridge Crop Mean Compactness Bridge Crop StdDev Compactness Bridge Crop Mean Density Bridge Crop StdDev Density Bridge Crop Mean Diff. Of enclosing / enclosed ellipse Bridge Crop StdDev Diff. Of enclosing / enclosed ellipse Bridge Fit Elean StdDev Elliptic Fit Bridge Crop Mean Length Pxl Bridge Crop StdDev Length Pxl Bridge Crop Mean Length / width Correct Crop StdDev Length / width Bridge Crop Mean Main direction Bridge Crop StdDev Main direction Correct Crop Mean Max.Diff. Correct Crop StdDev Max.Diff. Bridge Crop Mean Mean Channel 1 Correction Crop StdDev Mean Channel 1 Correct Mean Crop 2 Mean Mean Channel 2 Bridge Crop StdDev Mean Channel 2 -102- (99) 200538734 Characteristics _ Bridge Crop Mean Mean Channel 3 Correction StdDev Mean Channel 3 Corrected crop Mean Radius of largest enclosed ellipse Corrected crop StdDev Radius of largest enclosed ellipse Bridge crop Mean Radius of smallest enclosing ellipse Bridge crop StdDev Radius of smallest enclosing ellipse Corrected crop Mean Rectangular Fit Corrected crop StdDev Rectangular Fit Bridge crop Mean Shape index Orthopedic crop StdDev Shape index Bridge crop Mean Stddev Channell Bridge crop StdDev Stddev Channell Orthotics Crop Stddev Channel2

矯作物 StdDev Stddev Channel2 橋作物 Mean Stddev Channel3 橋作物 StdDev Stddev Channel3 矯作物 Mean Width Pxl 矯作物 StdDev Width Pxl 細胞質 Mean Area Pxl 細胞質 StdDev Area Pxl 細胞質 Mean Asymmetry 細胞質 StdDev Asymmetry 細胞質 Mean Border index 細胞質 StdDev Border index 細胞質 Mean Border length Pxl 細胞質 StdDev Border length Pxl 細胞質 Mean Brightness 細胞質 StdDev Brightness 細胞質 Mean Compactness 細胞質 StdDev Compactness 細胞質 Mean Density 細胞質 StdDev Density 細胞質 Mean Diff· of enclosing/enclosed ellipse 細胞質 StdDev Diff_ of enclosing/enclosed ellipse 細胞質 Mean Elliptic Fit 細胞質 StdDev Elliptic Fit 細胞質 Mean Length Pxl 細胞質 StdDev Length Pxl 細胞質 Mean Length/width 細胞質 StdDev Length/width 細胞質 Mean Main direction 細胞質 StdDev Main direction 細胞質 Mean Max.Diff. 細胞質 StdDev Max.Diff· 細胞質 Mean Mean Channel 1 細胞質 StdDev Mean Channel 1 103- (100) 200538734 特性_ 細胞質 Mean Mean Channel 2 細胞質 StdDev Mean Channel 2 細胞質 Mean Mean Channel 3 細胞質 StdDev Mean Channel 3 細胞質 Mean Radius of largest enclosed ellipse 細胞質 StdDev Radius of largest enclosed ellipse 細胞質 Mean Radius of smallest enclosing ellipse 細胞質 StdDev Radius of smallest enclosing ellipse 細胞質 Mean Rectangular Fit 細胞質 StdDev Rectangular Fit 細胞質 Mean Shape index 細胞質 StdDev Shape index 細胞質 Mean Stddev Channel 1Correction Crop StdDev Stddev Channel2 Bridge Crop Mean Stddev Channel3 Bridge Crop StdDev Stddev Channel3 Correct Crop Mean Width Pxl Correct Crop StdDev Width Pxl Cytoplasm Mean Area Pxl Cytoplasm StdDev Area Pxl Cytoplasm Mean Asymmetry Cytoplasm StdDev Asymmetry Cytoplasm Mean Border Cell Index length Pxl Cytoplasmic StdDev Border length Pxl Cytoplasmic Mean Brightness Cytoplasmic StdDev Brightness Cytoplasmic Mean Compactness Cytoplasmic StdDev Compactness Cytoplasmic Mean Density Cytoplasmic StdDev Density Cytoplasmic Mean Diff of of enclosing / enclosed ellipse Cytoplasmic StdDev Diff_ of enclosingDevlipfit Cell Ellipd Fitted Cytoplasmic Mean Length Pxl cytoplasmic StdDev Length Pxl cytoplasmic Mean Length / width cytoplasmic StdDev Length / width cytoplasmic Mean Main direction cytoplasmic StdDev Main direction cytoplasmic Mean Max.Diff. Cytoplasmic StdDev Max.Diff · cytoplasmic Mean Mean Channel 1 cytoplasmic StdDev Mean Channel 1 103- (100) 20 0538734 Characteristics_ Cytoplasmic Mean Mean Channel 2 Cytoplasmic StdDev Mean Channel 2 Cytoplasmic Mean Mean Channel 3 Cytoplasmic StdDev Mean Channel 3 Cytoplasmic Mean Radius of largest enclosed ellipse Cytoplasmic StdDev Radius of largest enclosed ellipse Cytoplasmic Mean Radius of smallest enclosing ellipse Cytoplasmic StdDev Radius of smallest enclosing ellipse Cytoplasmic Mean Rectangular Fit Cytoplasmic StdDev Rectangular Fit Cytoplasmic Mean Shape index Cytoplasmic StdDev Shape index Cytoplasmic Mean Stddev Channel 1

細胞質 StdDev Stddev Channel 1 細胞質 Mean Stddev Channel 2 細胞質 StdDev Stddev Channel 2 細胞質 Mean Stddev Channel 3 細胞質 StdDev Stddev Channel 3 細胞質 Mean Width Pxl 細胞質 StdDev Width Pxl 上皮核 Mean Area Pxl 上皮核 StdDev Area Pxl 上皮核 Mean Asymmetry 上皮核 StdDev Asymmetry 上皮核 Mean Border index 上皮核 StdDev Border index 上皮核 Mean Border length Pxl 上皮核 StdDev Border length Pxl 上皮核 Mean Brightness 上皮核 StdDev Brightness 上皮核 Mean Compactness 上皮核 StdDev Compactness 上皮核 Mean Density 上皮核 StdDev Density 上皮核 Mean Diff. of enclosing/enclosed ellipse 上皮核 StdDev Diff· of enclosing/enclosed ellipse 上皮核 Mean Elliptic Fit 上皮核 StdDev Elliptic Fit 上皮核 Mean Length Pxl 上皮核 StdDev Length Pxl 上皮核 Mean Lengthlwidth 上皮核 StdDev Length/width 上皮核 Mean Main direction 上皮核 StdDev Main direction 上皮核 Mean Max.Diff. 上皮核 StdDev Max.Diff. -104- (101) 200538734 特性_ 上皮核 Mean Mean Channel 1 上皮核 StdDev Mean Channel 1 上皮核 Mean Mean Channel 2 上皮核 StdDev Mean Channel 2 上皮核 Mean Mean Channel 3 上皮核 StdDev Mean Channel 3 上皮核 Mean Radius of largest enclosed ellipse 上皮核 StdDev Radius of largest enclosed ellipse 上皮核 Mean Radius of smallest enclosing ellipse 上皮核 StdDev Radius of smallest enclosing ellipse 上皮核 Mean Rectangular Fit 上皮核 StdDev Rectangular Fit 上皮核 Mean Shape indexCytoplasmic StdDev Stddev Channel 1 cytoplasmic Mean Stddev Channel 2 cytoplasmic StdDev Stddev Channel 2 cytoplasmic Mean Stddev Channel 3 cytoplasmic StdDev Stddev Channel 3 cytoplasmic Mean Width Pxl cytoplasmic StdDev Width Pxl epithelial nucleus Mean Area Pxl epithelial nucleus StdDev Area Pxlym epithelium Asymmetry Epithelial Core Mean Border index Epithelial Core StdDev Border index Epithelial Core Mean Border length Pxl Epithelial Core StdDev Border length Pxl Epithelial Core Mean Brightness Epithelial Core StdDev Brightness Epithelial Core Mean Compactness Epithelial Core StdDev Compactness Epithelial Core Mean Density Epithelial Core StdDev Diff. Of enclosing / enclosed ellipse Epithelial Nucleus StdDev Diff · of enclosing / enclosed ellipse Epithelial Nucleus Mean Elliptic Fit Epithelial Nucleus StdDev Elliptic Fit Epithelial Nucleus Mean Length Pxl Epithelial Nucleus StdDev Length Pxl Epithelial Nucleus Mean LengthlWidth Epithelial Nucleus StdDev Main direction Epithelial nucleus StdDev Main direction Epithelial nucleus Mean Max.Diff. Epithelial nucleus StdDev Max.Diff. -104- ( 101) 200538734 Characteristics_ Epithelial Mean Mean Channel 1 Epithelial Nucleus StdDev Mean Channel 1 Epithelial Nucleus Mean Mean Channel 2 Epithelial Nucleus StdDev Mean Channel 2 Epithelial Nucleus Mean Mean Channel 3 Epithelial Nucleus StdDev Mean Channel 3 Epithelial Nucleus Mean Radius of Largest Enclosed Ellipse Epithelium StdDev Radius of largest enclosed ellipse Epithelial nucleus Mean Radius of smallest enclosing ellipse Epithelial nucleus StdDev Radius of smallest enclosing ellipse Epithelial nucleus Mean Rectangular Fit Epithelial nucleus StdDev Rectangular Fit Epithelial Mean Shape index

上皮核 StdDev Shape index 上皮核 Mean Stddev Channel 1 上皮核 StdDev Stddev Channel 1 上皮核 Mean Stddev Channel 2 上皮核 StdDev Stddev Channel 2 上皮核 Mean Stddev Channel 3 上皮核 StdDev Stddev Channel 3 上皮核 Mean Width Pxl 上皮核 StdDev Width Pxl 腔 Mean Area Pxl 腔 StdDev Area Pxl 腔 Mean Asymmetry 腔 StdDev Asymmetry 腔 Mean Border index 腔 StdDev Border index 腔 Mean Border length Pxl 腔 StdDev Border length Pxl 腔 Mean Brightness 腔 StdDev Brightness 腔 Mean Compactness 腔 StdDev Compactness 腔 Mean Density 腔 StdDev Density 腔 Mean Diff. of enclosing/enclosed ellipse 腔 StdDev Diff. of enclosing/enclosed ellipse 腔 Mean Elliptic Fit 腔 StdDev Elliptic Fit 腔 Mean Length Pxl 腔 StdDev Length Pxl 腔 Mean Length/width 腔 StdDev Length/width 腔 Mean Main direction 腔 StdDev Main direction -105- (102) 200538734 特性_ 腔 Mean Max.Diff. 腔 StdDev Max.Diff. 腔 Mean Mean Channel 1 腔 StdDev Mean Channel 1 腔 Mean Mean Channel 2 腔 StdDev Mean Channel 2 腔 Mean Mean Channel 3 腔 StdDev Mean Channel 3 腔 Mean Radius of largest enclosed ellipse 腔 StdDev Radius of largest enclosed ellipse 腔 Mean Radius of smallest enclosing ellipse 腔 StdDev Radius of smallest enclosing ellipse 腔 Mean Rectangular FitEpithelial core StdDev Shape index Epithelial core Mean Stddev Channel 1 Epithelial core StdDev Stddev Channel 1 Epithelial core Mean Stddev Channel 2 Epithelial core StdDev Stddev Channel 2 Epithelial core Mean Stddev Channel 3 Epithelial core StdDev Stddev Channel 3 Epithelial core Mean Width PxDev Epithelial Pxl Cavity Mean Area Pxl Cavity StdDev Area Pxl Cavity Mean Asymmetry Cavity StdDev Asymmetry Cavity Mean Border index Cavity StdDev Border index Cavity Mean Border length Pxl Cavity StdDev Border Length Pxl Cavity Mean Brightness Cavity StdDev Brightness Cavity Mean Compactness Cavity StDDev DevDev Density cavity Mean Diff. Of enclosing / enclosed ellipse cavity StdDev Diff. Of enclosing / enclosed ellipse cavity Mean Elliptic Fit cavity StdDev Elliptic Fit cavity Mean Length Pxl cavity StdDev Length Pxl cavity Mean Length / width cavity StdDev Length / direction Width cavity StdDev Main direction -105- (102) 200538734 Features_ Cavity Mean Max.Diff. Cavity StdDev Max.Diff. Cavity Mean Mean Channel 1 Cavity StdDev Mean Channel 1 Cavity Mean Mean Channel 2 Cavity StdDev Mean Channel 2 cavity Mean Mean Channel 3 cavity StdDev Mean Channel 3 cavity Mean Radius of largest enclosed ellipse cavity StdDev Radius of largest enclosed ellipse cavity Mean Radius of smallest enclosing ellipse cavity StdDev Radius of smallest enclosing ellipse cavity Mean Rectangular Fit

腔 StdDev Rectangular Fit 腔 Mean Shape index 腔 StdDev Shape index 腔 Mean Stddev Channel 1 腔 StdDev Stddev Channel 1 腔 Mean Stddev Channel 2 腔 StdDev Stddev Channel 2 腔 Mean Stddev Channel 3 腔 StdDev Stddev Channel 3 腔 Mean Width Pxl 腔 StdDev Width Pxl 基質 Mean Area Pxl 基質 StdDev Area Pxl 基質 Mean Asymmetry 基質 StdDev Asymmetry 基質 Mean Border index 基質 StdDev Border index 基質 Mean Border length Pxl 基質 StdDev Border length Pxl 基質 Mean Brightness 基質 StdDev Brightness 基質 Mean Compactness 基質 StdDev Compactness 基質 Mean Density 基質 StdDev Density 基質 Mean Diff· of enclosing/enclosed ellipse 基質 StdDev Diff. of enclosing/enclosed ellipse 基質 Mean Elliptic Fit 基質 StdDev Elliptic Fit 基質 Mean Length Pxl 基質 StdDev Length Pxl 基質 Mean Length/width 基質 StdDev Length/width (103) 200538734 特性_ 基質 Mean Main direction 基質 StdDev Main direction 基質 Mean Max.Diff. 基質 StdDev Max.Diff. 基質 Mean Mean Channel 1 基質 StdDev Mean Channel 1 基質 Mean Mean Channel2 基質 StdDev Mean Channel2 基質 Mean Mean Channel3Cavity StdDev Rectangular Fit Cavity Mean Shape index Cavity StdDev Shape index Cavity Mean Stddev Channel 1 Cavity StdDev Stddev Channel 1 Cavity Mean Stddev Channel 2 Cavity StdDev Stddev Channel 2 Cavity Mean Stddev Channel 3 Cavity StdDev Stddev Channel 3 Cavity Mean Width Pxl Cavity StdDev Matrix Mean Area Pxl Matrix StdDev Area Pxl Matrix Mean Asymmetry Matrix StdDev Asymmetry Matrix Mean Border index Matrix StdDev Border index Matrix Mean Border length Pxl Matrix StdDev Border length Pxl Matrix Mean Brightness Matrix StdDev Brightness Matrix Mean Compactness Matrix StdDevevity Matrix Mean Diff of of enclosing / enclosed ellipse Matrix StdDev Diff. Of enclosing / enclosed ellipse Matrix Mean Elliptic Fit Matrix StdDev Elliptic Fit Matrix Mean Length Pxl Matrix StdDev Length Pxl Matrix Mean Length / width Matrix StdDev Length / width (_103) 2005387 Matrix Mean Main direction Matrix StdDev Main direction Matrix Mean Max.Diff. Matrix StdDev Max.Diff. Matrix Mean Mean Channel 1 Substrate StdDev Mean Channel 1 Substrate Mean Mean Channel2 Substrate StdDev Mean Channel2 Substrate Mean Mean Channel3

基質 StdDev Mean ChanneB 基質 Mean Radius of largest enclosed ellipse 基質 StdDev Radius of largest enclosed ellipse 基質 Mean Radius of smallest enclosing ellipseMatrix StdDev Mean ChanneB matrix Mean Radius of largest enclosed ellipse matrix StdDev Radius of largest enclosed ellipse matrix Mean Radius of smallest enclosing ellipse

基質 StdDev Radius of smallest enclosing ellipse 基質 Mean Rectangular Fit 基質 StdDev Rectangular Fit 基質 Mean Shape index 基質 StdDev Shape index 基質 Mean Stddev Channel 1 基質 StdDev Stddev Channell 基質 Mean Stddev Channel2 基質 StdDev Stddev Channel2 基質 Mean Stddev Channel3 基質 StdDev Stddev Channel3 基質 Mean Width Pxl 基質 StdDev Width Pxl 基質核 Mean Area Pxl 基質核 StdDev Area Pxl 基質核 Mean Asymmetry 基質核 StdDev Asymmetry 基質核 Mean Border index 基質核 StdDev Border index 基質核 Mean Border length Pxl 基質核 StdDev Border length Pxl 基質核 Mean Brightness 基質核 StdDev Brightness 基質核 Mean Compactness 基質核 StdDev Compactness 基質核 Mean Density 基質核 StdDev Density 基質核 Mean Diff· of enclosing/enclosed ellipse 基質核 StdDev Diff. of enclosing/enclosed ellipse 基質核 Mean Elliptic Fit 基質核 StdDev Elliptic Fit 基質核 Mean Length Pxl 基質核 StdDev Length Pxl -107- (104) 200538734 特性_ 基質核 Mean Length/width 基質核 StdDev Length/width 基質核 Mean Main direction 基質核 StdDev Main direction 基質核 Mean Max.Diff. 基質核 StdDev Max.Diff. 基質核 Mean Mean Channel 1 基質核 StdDev Mean Channel 1 基質核 Mean Mean Channel 2 基質核 StdDev Mean Channel2 基質核 Mean Mean Channel 3 基質核 StdDev Mean Channel 3Matrix StdDev Radius of smallest enclosing ellipse Matrix Mean Rectangular Fit Matrix StdDev Rectangular Fit Matrix Mean Shape index Matrix StdDev Shape index Matrix Mean Stddev Channel 1 Matrix StdDev Stddev Channell Matrix Mean Stddev Channel2 Matrix StdDev Stddev Channel2 Matrix Mean Stddev Channel3 Width Pxl Matrix StdDev Width Pxl Matrix Core Mean Area Pxl Matrix Core StdDev Area Pxl Matrix Core Mean Asymmetry Matrix Core StdDev Asymmetry Matrix Core Mean Border index Matrix Core StdDev Border index Matrix Core Mean Border Length Pxl Matrix Core StdDev Border Length Pxl Matrix Core Bright Matrix StdDev Brightness Matrix Core Mean Compactness Matrix Core StdDev Compactness Matrix Core Mean Density Matrix Core StdDev Density Matrix Core Mean Diff · of enclosing / enclosed ellipse Matrix Core StdDev Diff. Of enclosing / enclosed ellipse Matrix Core Elliptic Fit Elliptic Fit Matrix Stromal core Mean Length Pxl StdDev Length Pxl -107- (104) 20053 8734 Features_ Matrix Core Length / width Matrix Core StdDev Length / width Matrix Core Mean Main direction Matrix Core StdDev Main direction Matrix Core Mean Max.Diff. Matrix Core StdDev Max.Diff. Matrix Core Mean Mean Channel 1 Matrix Core StdDev Mean Channel 1 Matrix nucleus Mean Mean Channel 2 Matrix nucleus StdDev Mean Channel 2 Matrix nucleus Mean Mean Channel 3 Matrix nucleus StdDev Mean Channel 3

基質核 Mean Radius of largest enclosed ellipse 基質核 StdDev Radius of largest enclosed ellipse 基質核 Mean Radius of smallest enclosing ellipse 基質核 StdDev Radius of smallest enclosing ellipse 基質核 Mean Rectangular Fit 基質核 StdDev Rectangular Fit 基質核 Mean Shape index 基質核 StdDev Shape index 基質核 Mean Stddev Channel 1 基質核 StdDev Stddev Channel 1 基質核 Mean Stddev Channel 2 基質核 StdDev Stddev Channel 2 基質核 Mean Stddev Channel 3 基質核 StdDev Stddev Channel 3 基質核 Mean Width Pxl 基質核 StdDev Width Pxl 矯作物面積Pxl 細胞質面積Pxl 上皮核面積Pxl 腔面積Pxl 紅血球面積Pxl 基質面積Pxl 基質核面積Pxl 矯作物物件數 細胞質物件數 上皮核物件數 腔物件數 紅血球物件數 基質物件數 基質核物件數 紅血球 Mean Area Pxl 紅血球 StdDev Area Pxl 紅血球 Mean Asymmetry 紅血球 StdDev Asymmetry -108- (105) 200538734 特性_ 紅血球 Mean Border index 紅血球 StdDev Border index 紅血球 Mean Border length Pxl 紅血球 StdDev Border length Pxl 紅血球 Mean Brightness 紅血球 StdDev Brightness 紅血球 Mean Compactness 紅血球 StdDev Compactness 糸工血球Mean Density 紅血球 StdDev Density 紅血球 Mean Diff. of enclosing/enclosed ellipseMatrix core Mean Radius of largest enclosed ellipse Matrix core StdDev Radius of largest enclosed ellipse Matrix core Mean Radius of smallest enclosing ellipse Matrix core StdDev Radius of smallest enclosing ellipse Matrix core Mean Rectangular Fit Matrix core StdDev Rectangular Fit Matrix core Mean Shape index StdDev Shape index Matrix Core Stddev Channel 1 Matrix Core StdDev Stddev Channel 1 Matrix Core Mean Stddev Channel 2 Matrix Core StdDev Stddev Channel 2 Matrix Core Mean Stddev Channel 3 Matrix Core StdDev Stddev Channel 3 Matrix Core Mean Width Pxl Matrix Core StdDev Width Pxl Corrective Crop Area Pxl Cytoplasmic area Pxl Epithelial nucleus area Pxl Cavity area Pxl Red blood cell area Pxl Matrix area Pxl Matrix nucleus area Pxl Orthopedic crop object Number of cytoplasmic objects Number of epithelial core objects Number of cavity objects Number of matrix cells Number of matrix objects Number of matrix nuclear objects Number of red blood cells Mean Area Pxl Red Blood Cell StdDev Area Pxl Red Blood Cell Mean Asymmetry Red Blood Cell StdDev Asymmetry -108- (105) 200538734 Characteristics_ Red Blood Cell Mean Border index Red Blood Cell StdD ev Border index Red Blood Cell Mean Border length Pxl Red Blood Cell StdDev Border length Pxl Red Blood Cell Mean Brightness Red Blood Cell StdDev Brightness Red Blood Cell Mean Compactness Red Blood Cell StdDev Compactness Maternity Blood Cell Mean Density Red Blood Cell StdDev Density Red Blood Cell Mean Diff. of enclosing / enclosed ellipse

紅血球 StdDev Diff. of enclosing/enclosed ellipse 紅血球 Mean Elliptic Fit 糸工血球 StdDev Elliptic Fit 紅血球 Mean Length Pxl 紅血球 StdDev Length Pxl 紅血球 Mean Length/width 紅血球 StdDev Length/width 紅血球 Mean Main direction 紅血球 StdDev Main direction 紅血球 Mean Max.Diff. 紅血球 StdDev Max.Diff. 糸工血球 Mean Mean Channel 1 紅血球 StdDev Mean Channel 1 紅血球 Mean Mean Channel 2 紅血球 StdDev Mean Channel 2 紅血球 Mean Mean Channel 3 紅血球 StdDev Mean Channel 3 紅血球 Mean Radius of largest enclosed ellipse 紅血球 StdDev Radius of largest enclosed ellipse 紅血球 Mean Radius of smallest enclosing ellipse 紅血球 StdDev Radius of smallest enclosing ellipse 紅血球 Mean Rectangular Fit 紅血球 StdDev Rectangular Fit 紅血球 Mean Shape index 紅血球 StdDev Shape index 紅血球 Mean Stddev Channel 1 紅血球 StdDev Stddev Channel 1 紅血球 Mean Stddev Channel 2 紅血球 StdDev Stddev Channel 2 紅血球 Mean Stddev Channel 3 紅血球 StdDev Stddev Channel 3 紅血球 Mean Width Pxl 紅血球 StdDev Width Pxl -109- 200538734 (106) 【圖式簡單說明】 爲了對本發明具體實例有較佳了解，要參考下面的說 明及配合所附之圖式，其中相同的指示數字於全文中係指 稱相同的部份，且於其中： 圖1 A和1 B爲使用一預測模型來治療、診斷或預測一 醫療狀況的發生之系統方塊圖解； 圖1 C爲產生一預測模型所用系統的方塊圖解； # 圖2顯示出可由預測模型輸出的患者範例結果； 圖3爲在處理組織影響中所包括的範例階段之流程圖 圖4爲篩選一醫療狀況的抑制劑化合物中包含的範例 階段之流程圖； 圖5 a和5 b分別顯示出在影像分割與分類後，健康與 異常***組織樣品的灰標度數位影像； 圖6顯示出一預測***癌復發所用模型中使用的各 β 種臨床、分子和電腦產生之形態學特性； 圖7a和7b顯示出染色組織樣品，展示出兩種分子特 性之存在，特別是雄激素受體（AR)和CD 34 ; 圖8爲一 Kaplan-Meier曲線圖，展示出將患者分類爲 低風險、中等風險、或高風險、係根據圖6的特性之模型 所預測的發生***癌復發率所示者； 圖9顯示出一模型預測***癌復發所用的各種臨床 、分子、和電腦產生的形態學特性； 圖10爲一 Kaplan-Meier曲線，展示出由根據圖9的 -110- 200538734 (107) 特性之模型預測出會經歷***癌復發之低風險、中等風 險、或高風險之患者分類； 圖11顯示出一模型預測***癌整體存活率所用之 各種臨床、分子和電腦產生的形態學特性； 圖12爲一 Kaplan-Meier曲線，展示出由根據圖11的 特性之模型預測出因任何肇因所致死亡之低風險、中等風 險、或高風險的患者分類；及 φ 圖1 3顯示出一預測在患者接受***切除術之後發 生侵入性疾病所用的各種臨床和電腦產生之形態學特性。 【主要元件之符號說明】 102 :預測模型 104 :診斷設施 106 :遠程存取裝置 108 :網際網路服務提供者 • 110/112 :通信網路 1 2 2 :檢驗套組 124 :設施 1 3 2 :分析工具 134 :數據庫 1 3 6 :影像處理工具 -111 -Red Blood Cell StdDev Diff. Of enclosing / enclosed ellipse Red Blood Cell Mean Elliptic Fit StdDev Elliptic Fit Red Blood Cell Mean Length Pxl Red Blood Cell StdDev Length Pxl Red Blood Cell Mean Length / width Red Blood Cell StdDev Length / width Red Blood Cell Main Main direction Red Blood Cell StdDDDev Main Direction Red Blood Cells StdDev Max.Diff. Maternity Blood Cells Mean Mean Channel 1 Red Blood Cells StdDev Mean Channel 1 Red Blood Cells Mean Mean Channel 2 Red Blood Cells StdDev Mean Channel 2 Red Blood Cells Mean Mean Channel 3 Red Blood Cells StdDev Mean Channel 3 Red Blood Cells Mean Radius of largest enclosed ellipse Red Blood Cells StdDev Radius of Largest enclosed ellipse Red blood cells Mean Radius of smallest enclosing ellipse Red blood cells StdDev Radius of smallest enclosing ellipse Red blood cells Mean Rectangular Fit Red blood cells StdDev Rectangular Fit Red blood cells Mean Shape index Red blood cells StdDev Shape index Red blood cells Mean Stddev Channel 1 Red blood cells StdDev Stddev Stddev Channel 1 Stddev Channel 2 Red Blood Cell M ean Stddev Channel 3 Red blood cell StdDev Stddev Channel 3 Red blood cell Mean Width Pxl Red blood cell StdDev Width Pxl -109- 200538734 (106) [Simplified illustration of the drawing] For a better understanding of the specific examples of the present invention, please refer to the following description and the accompanying Figures, where the same indicator numbers refer to the same parts throughout, and in which: Figures 1 A and 1 B are block diagrams of a system using a predictive model to treat, diagnose, or predict the occurrence of a medical condition; 1 C is a block diagram of the system used to generate a predictive model; # FIG. 2 shows example patient outcomes that can be output by the predictive model; FIG. 3 is a flowchart of example stages involved in processing organizational impact. Figure 5a and 5b show grey scale digital images of healthy and abnormal prostate tissue samples after image segmentation and classification, respectively; Figure 6 shows a predicted prostate cancer Clinical, molecular, and computer-generated morphological characteristics of each beta used in the model used for relapse; Figures 7a and 7b show Stained tissue samples showing the presence of two molecular characteristics, particularly androgen receptor (AR) and CD 34; Figure 8 is a Kaplan-Meier plot showing the classification of patients as low-risk, medium-risk, or high The risk is shown by the model of prostate cancer recurrence rate predicted based on the characteristics of Figure 6; Figure 9 shows various clinical, molecular, and computer-generated morphological characteristics used by a model to predict prostate cancer recurrence; Figure 10 is A Kaplan-Meier curve showing the classification of low-, medium-, or high-risk patients predicted to experience prostate cancer recurrence by a model based on the characteristics of -110-200538734 (107) in Figure 9; Figure 11 shows a model Various clinical, molecular, and computer-generated morphological characteristics used to predict overall survival of prostate cancer; Figure 12 is a Kaplan-Meier curve showing the low level of death predicted by any cause by a model based on the characteristics of Figure 11 Classification of patients at high, medium, or high risk; and φ Figure 13 shows a graph used to predict invasive disease after a patient undergoes prostatectomy. Kinds of clinical and morphological characteristics of computer-generated. [Symbol description of main components] 102: Prediction model 104: Diagnostic facility 106: Remote access device 108: Internet service provider 110/112: Communication network 1 2 2: Inspection set 124: Facility 1 3 2 : Analysis Tool 134: Database 1 3 6: Image Processing Tool -111-

Claims (1)

200538734 (1) 十、申請專利範圍 1 · 一種評定一患者醫療狀況發生風險之方法，該方 法包括： 接受該患者的患者資料組；及 使用可預測醫療狀況的模型評定該患者資料組，其中 該模型係以一或更多種臨床特性，一或更多種分子特性， 及一或更多種從一或更多幅組織影像產生的電腦產生之形 態學特性爲基礎，藉此評定該患者醫療狀況的發生風險。 2. 如申請專利範圍第1項之方法，其進一步包括使 用電腦從該患者的組織影像產生一或更多項形態學特性以 包含在該患者資料組內，其中該形態學特性的產生包括： 將組織影像分割成一或更多物件； 將該一或更多物件分類成一或更多物件類別；及 經由對該一或更多物件類別採取一或更多測量而測定 形態學特性。 3. 如申請專利範圍第2項之方法，其中將一或更多 物件分類成一或更多物件類別之步驟包括將一或更多物件 分別分類成來自基質、細胞質、上皮核、基質核、腔、紅 血球、組織矯件物、和組織背景所構成的類別群組中之一 類別。 4. 如申請專利範圍第2項之方法，其中該一或更多 針對一或更多物件類別的測量包括進行一或更多的針對一 或更多物件類別的一或更多光譜性質及/或一或更多形狀 性質之測量。 -112- 200538734 (2) 5 .如申請專利範圍第1項之方法，其中該模型係預 測***癌復發率且一或更多形態學特性係來自紅血球形 態學特性、上皮核形態學特性、基質形態學特性、腔形態 學特性、細胞質形態學特性、和組織背景形態學特性所構 成的形態學特性群組之中。 6 ·如申請專利範圍第5項之方法，其中一或多種臨 床特性來自圖6中所列臨床特性群組之中。 p 7·如申請專利範圍第6項之方法，其中一或多種分 子特性來自圖6中所列臨床特性群組之中。 8 ·如申請專利範圍第7項之方法，其中該預測模型 包括一至少約0.8 8的和諧指數。 9 ·如申請專利範圍第7項之方法，其中該預測模型 包括對於對數等級檢驗低於約0.000 1之p値。 1 0·如申請專利範圍第1項之方法，其中該模型係預 測***癌復發率且一或更多形態學特性係來自紅血球形 i 態學特性、上皮核形態學特性、基質形態學特性、腔形態 學特性、和細胞質形態學特性所構成的形態學特性群組中 〇 11. 如申請專利範圍第1 0項之方法，其中一或更多 臨床特性來自圖9中所列的臨床特性群組。 12. 如申請專利範圍第1 1項之方法，其中一或更多 分子特性來自圖9中所列的分子特性群組。 1 3 ·如申請專利範圍第1 2項之方法，其中該預測模 型包括至少約0.87之和諧指數。 -113- 200538734 (3) 14.如申請專利範圍第1 2項之方法，其中該預測模 型包括對於對數等級檢驗低於約0.0 0 0 1之p値。 1 5 ·如申請專利範圍第1項之方法，其中該方法係預 測***癌存活率且一或更多形態學特性係來自紅血球形 態學特性、上皮核形態學特性，和基質形態學特性所構成 的形態學特性群組之中。 16·如申請專利範圍第15項之方法，其中一或更多 臨床特性來自圖1 1中所列的臨床特性群組。 1 7 ·如申請專利範圍第1 6項之方法，其中一或更多 分子特性來自圖1 1中所列的分子特性群組。 1 8 .如申請專利範圍第1 7項之方法，其中該預測模 型包括至少約〇 . 8 0的和諧指數。 1 9.如申請專利範圍第1 7項之方法，其中該預測模 型包括對於對數等級檢驗低於約0.0001之P値。 2 0.如申請專利範圍第1項之方法，其中用該預測模 型評定該患者資料組之步驟包括定出該患者的診斷計分。 2 1.如申請專利範圍第1項之方法，其中用該預測模 型評定該患者資料之步驟包括患者醫療狀況發生之可能時 22.如申請專利範圍第1項之方法’其中該評定一醫 療狀況發生風險的步驟包括測定該患者對一治療法的可能 回應性或未回應性。 23 ·如申請專利範圍第1項之方法’其進一步包括輸 出表示評定結果之資料。 -114 - 200538734 (4) 24.如申請專利範圍第23項之方法，其中該結果包 括來自診斷計分，表明一或多項以預測模型分析所得患者 資料組的特性之資訊，表明預測模型的準確性之資訊、或 彼等的組合所構成的結果群組之中的結果。 2 5.如申請專利範圍第2 3項之方法，其中該接收該 患者的患者資料組之步驟包括從遙遠位置接收患者的資料 且其中該輸出表明評定結果的資料之步驟包括將結果傳遞 到遙遠位置。 26.如申請專利範圍第25項之方法，其中該接收和 該傳遞包括在一或更多通訊網路接收和傳遞。 2 7 . —種評定一患者***癌復發風險之方法’該方 法包括： 接收該患者的患者資料組；及 用預測***癌復發的模型評定該患者資料組’其中 該模型係以雄激素（AR )的染色指數爲基準’藉此評定該 患者的***癌復發風險。 2 8 . —種產生可預測醫療狀況的模型之方法’該方法 包括= 對於其針對該醫療狀況的結果爲至少部份已知的二或 更多受試者中每一受訊者接收一訊練受試者資料組’該資 料組包括一或更多項臨床特性、一或更多分子特性、和一 或更多從組織影像產生的電腦產生之形態學特性；及 對該訓練受試者資料組實施多元分析’藉此產生可預 測該醫療狀況的模型。 -115- 200538734 (5) 2 9 ·如申請專利範圍第2 8項之方法，其進一步包括 使用一電腦從組織影像產生一或更多形態學特性，其中該 產生該形態學特性之步驟包括： 將一組織影像分割成一或更多物件； 將該一或更多物件分類成一或更多物件類別；及 經由對該一或更多物件類別進行一或更多測量以定出 該形態學特性。 • 3 0·如申請專利範圍第2 9項之方法，其中該將該一 或更多物件分類成一或更多物件類別之步驟包括將該一或 更多物件中的每一者分類成由基質、細胞質、上皮核、基 質核、腔、紅血球、組織矯作物、和組織背景所構成的類 別群組中之類別。 3 1 ·如申請專利範圍第2 9項之方法，其中該進行有 關該一或更多物件類別的一或更多測量之步驟包括進行該 一或更多物件類別的一或更多光譜性質及/或一或更多形 • 狀性質之一或更多測量。 3 2.如申請專利範圍第2 8項之方法，其中該一或更 多臨床特性係來自表4中所列的臨床特性群組中。 33·如申請專利範圍第28項之方法，其中一或更多 分子特性係來自表6中所列的分子特性群組中。 34.如申請專利範圍第28項之方法，其中該模型係 以所選的該訓練目標資料組的一或更多臨床特性、一或更 多分子特性、和一或更多電腦產生的形態學特性之亞組爲 基礎，該亞組係經選擇成可預測該醫療狀況者。 -116- 200538734 (6) 35.如申請專利範圍第3 4項之方法，其中該模型係 預測***癌復發率且該選定的特定亞組包括來自紅血球 形態學特性、上皮核形態學特性、基質形態學特性、腔形 態學特性、細胞質形態學特性、和組織背景形態學特性的 形態學特性群組中之形態學特性。 3 6.如申請專利範圍第3 5項之方法，其中該選定的 特性亞組包括表6所列的一或更多臨床特性。 37.如申請專利範圍第36項之方法，其中該選定的 特性亞組包括表6所列的一或更多分子特性。 3 8 .如申請專利範圍第3 7項之方法，其中該預測模 型包括至少約〇 . 8 8的和諧指數。 39. 如申請專利範圍第37項之方法，其中該預測模 型包括對於對數等級檢驗低於約0.000 1之P値。 40. 如申請專利範圍第34項之方法，其中該模型係 預測***癌復發率且該選定的特定亞組包括來自紅血球 形態學特性、上皮核形態學特性、基質形態學特性、腔形 態學特性、和細胞質形態學特性所構成的形態學特性群組 中之一或更多形態學特性。 4 1 ·如申請專利範圍第4 0項之方法，其中該選定的 特性亞組包括來自圖9所列臨床特性群組中的一或更多臨 床特性。 42.如申請專利範圍第41項之方法，其中該選定的 特性亞組包括來自圖9所列分子特性群組中的一或更多臨 床特性。 -117- 200538734 (7) 43. 如申請專利範圍第42項之方法，其中該預測模 型包括至少約0.87的和諧指數。 44. 如申請專利範圍第42項之方法，其中該預測模 型包括對於對數等級檢驗低於約0.000 1之P値。 45. 如申請專利範圍第34項之方法，其中該模型係 預測***癌復發率且該選定的特定亞組包括來自紅血球 形態學特性、上皮核形態學特性、和基質形態學特性所構 成的形態學特性群組中的一或更多形態學特性。 46. 如申請專利範圍第45項之方法，其中該選定的 特性亞組包括來自圖1 1所列臨床特性群組中之一或更多 臨床特性。 47. 如申請專利範圍第46項之方法，其中該選定的 特性亞組包括來自圖1 1所列分子特性群組中之一或更多 臨床特性。 48·如申請專利範圍第47項之方法，其中該預測模 型包括至少約0.80的和諧指數。 49. 如申請專利範圍第4 7項之方法，其中該預測模 型包括對於對數等級檢驗低於約0.000 1之P値。 50. 如申請專利範圍第2 8項之方法，其中該實施多 元分析之步驟包括在該資料組實施支援向量回歸。 5 1 ·如申請專利範圍第2 8項之方法，其中該實施多 元分析之步驟包括使用實質地根據和諧指數（CI )近似的 目標函數來訓練一神經網路。 52· —種篩選一抑制劑化合物之方法，該方法包括·· -118- 200538734 (8) 接收一患者的第一資料組； 使用可預測醫療狀況的模型評估該第一資料組，其中 該模型係一或更多臨床特性、一或更多分子特性、和一或 更多從一或更多組織影像產生的電腦產生之形態學特性爲 基礎； 投予該患者一試驗化合物； 在投予該試驗化合物之後接收該患者的第二資料組； 使用該模型評估該第二資料組；及 將第一資料組的評估結果與該第二資料組的評估結果 相比較，其中在該第二資料組的結果中相對於第一資料組 所得結果之變化指示出該試驗化合物爲一種抑制劑化合物 〇 5 3 . —種評估一患者的醫療狀況發生風險所用裝置， 該裝置包括： 一可預測該醫療狀況之模型，其中該模型係以一或更 多臨床特性，一或更多分子特性、和一或更多從一或更多 組織影像產生的電腦產生之形態學特性爲基礎，其中該模 型係經構組成： 接收該患者的患者資料組；及 根該模型評估該患者資料組，藉此評定該患者的 醫療狀況發生風險。 54.如申請專利範圍第53項之裝置，其進一步包括 一影像處理工具，該工具係經構組成從該患者的一組織影 像產生一或更多形態學特性以包括在該患者資料組之內， -119- 200538734 (9) 其中該產生形態學特性的步驟包括： 將該組織影像使用該影像處理工具分割成一或更多物 件； 將該一或更多目標以該影像處理工具分類成一或更多 物件類別；及 用該影像處理工具對該一或更多物件類別進行一或更 多測量以定出該形態學特性。 5 5.如申請專利範圍第 5 4項之裝置，其中該以該影 像處理工具將一或更多物件分類成一或更多物件類別的步 驟包括以該影像處理工具將一或更多物件分別分類成選自 基質、細胞質、上皮核、基質核、腔、紅血球、組織橋作 物、和組織背景所構成的類別群組中之類別。 56. 如申請專利範圍第54項之裝置，其中該使用該 影像處理工具對該一或更多類別進行一或更多測量之步驟 包括使用該影像處理工具對一或更多物件類別的一或更多 光譜性質及/或一或更多形狀性質之一或更多測量。 57. 如申請專利範圍第5 3項之裝置，其中該模型係 預測***癌復發率且一或更多形態學特性係選自紅血球 形態學特性、上皮核形態學特性、基質形態學特性、腔形 態學特性、細胞質形態學特性、和組織背景形態學特性所 構成的形態學特性群組之中。 5 8.如申請專利範圍第5 7項之裝置，其中該一或更 多臨床特性係來自圖6中所列的臨床特性群組。 59·如申請專利範圍第58項之裝置，其中該一或更 -120- 200538734 (10) 多分子特性係來自圖6中所列分子特性群組。 60·如申請專利範圍第59項之裝置，其中該預測模 型包括至少約〇·88的和諧指數。 6 1.如申請專利範圍第5 9項之裝置，其中該預測模 型包括對於對數等級檢驗一低於約0 · 0 0 0 1的p値。 62. 如申請專利範圍第5 3項之裝置，其中該模型係 預測***癌復發率且一或更多形態學特性係來自紅血球 φ 形態學特性、上皮核形態學特性、基質形態學特性、腔形 態學特性、和細胞質形態學特性所構成的形態學特性群組 〇 63. 如申請專利範圍第62項之裝置，其中一或更多 臨床特性來自圖9中所列的臨床特性群組。 64. 如申請專利範圍第63項之裝置，其中一或更多 分子特性來自圖9中所列分子特性群組。 65 ·如申請專利範圍第64項之裝置，其中該預測模 Φ 型包括至少約〇 · 8 7的和諧指數。 6 6·如申請專利範圍第6 4項之裝置，其中該預測模 型包括對於對數等級檢驗一低於約〇 · 〇 〇 〇 1之p値。 6 7·如申請專利範圍第5 3項之裝置，其中該模型係 預測***癌存活率且一或更多形態學特性係選自紅血球 形態學特性、上皮核形態學特性、和基質形態學特性所構 成的形態學特性群組。 68·如申請專利範圍第67項之裝置，其中一或更多 臨床特性係選自圖1 1中所列臨床特性群組。 -121 200538734 (11) 69·如申請專利範圍第68項之裝置，其中一或更多 分子特性係選自圖1 1中所列分子特性群組。 7〇·如申請專利範圍第69項之裝置，其中該預測模 型包括至少約0 · 8 0之和諧指數。 7 1 ·如申請專利範圍第6 9項之裝置，其中該預測模 型包括對於對數等級檢驗一低於約〇 · 〇 〇 〇 1之ρ値。 7 2.如申請專利範圍第5 3項之裝置，其中該預測模 # 型係經構組成測定該患者的診斷計分。 7 3 ·如申請專利範圍第5 3項之裝置，其中該預測模 型係經構組成測定該患者的醫療狀況發生之可能時間。 74.如申請專利範圍第53項之裝置，其中該預測模 型係經構組成測定該患者對一治療法的可能回應性或回應 性。 7 5 ·如申請專利範圍第5 3項之裝置，其中該預測模 型經進一步構組成輸出指示出評定結果之資料。 • 76·如申請專利範圍第75項之裝置，其中該結果包 括選自診斷計分，指示出一或更多由該預測模型分析過的 患資料組的特性之資訊，指示出該預測模型的準確性之資 訊、或彼等的組合所構成的結果群組中之結果。 77·如申請專利範圍第75項之裝置，其中該預測模 型係經構組成從一遙遠位置接收該患者的患者資料組且輸 出結果以傳遞到遙遠位置。 78·如申請專利範圍第77項之裝置，其中該患者資 料組係透過一或更多通信網路接收且該結果係透過一或更 -122- 200538734 (12) 多通信網路傳遞。 79. 如申請專利範圍第53項之裝置，其中該預測模 型包括一神經網路。 80. 如申請專利範圍第53項之裝置，其中該預測模 型包括一支援向量機。 8 1 . —種評估一患者的***癌復發風險之裝置，該 裝置包括： φ 一預測***癌復發率的模型，其中該模型係以雄激 素受體（AR )的染色指數爲基底，且其中該模型係經構組 以： 接收該患者的患者資料組；及 預測該患者資料組，藉此評定患者的***癌復發風 險。 82. —種產生可預測醫療狀況的模型之裝置，該裝置 包括： • 一分析工具，其經構組成： 對於二或更多受試者的每一者，其醫療狀況的結 果係至少部份已知者，接收包括一或更多臨床特性、一或 更多分子特性、與一或更多從組織影像產生的電腦產生之 形態學特性的訓練受試者資料組；及 對該訓練受試者資料組實施多元分析，藉此產生 可預測醫療狀況的模型。 83·如申請專利範圍第82項之裝置，其進一步包括 一影像處理工具，其經構組以用電腦從該組織影像產生一 -123- 200538734 (13) 或更多形態學特性’其中該產生形態學特性的步驟包括： 用該影像處理工具將該組織影像分割成一或更多物件 將該一或更多目標用該影像處理工具分類成一或更多 物件類別；及 用該影像處理工具對該一或更多物件類別進行一或更 多測量。 84. 如申請專利範圍第8 3項之裝置，其中該使用該 影像處理工具將該一或更多物件分類成一或更多物件類別 之步驟包括以該影像處理工具將一或更多物件的每一者分 類成選自基質、細胞質、上皮核、基質核、腔、紅血球、 組織矯作物、和組織背景所構成的類別群組中之類別。 85. 如申請專利範圍第8 3項之裝置，其中該對一或 更多物件類別以該影像處理工具進行一或更多測量之步驟 包括用該影像處理工具針對該一或更多目標類別的一或更 多光譜性質及/或一或更多形狀性質進行一或更多測量。 86. 如申請專利範圍第82項之裝置，其中該一或更 多臨床特性係選自表4中所列的臨床特性群組。 87. 如申請專利範圍第82項之裝置，其中該一或更 多分子特性係選自表6中所列的分子特性群組。 88. 如申請專利範圍第82項之裝置，其中該模型係 以該訓練目標資料組中的該一或更多臨床特性、一或更多 分子特性、及一或更多電腦產生的形態學特性所構成之選 定亞組爲基礎，該亞組係經選定爲可預測該醫療狀況。 -124- 200538734 (14) 89. 如申請專利範圍第88項之裝置，其中該模型係 預測***癌復發率且該選定的特性亞組包括選自紅血球 形態學特性、上皮核形態學特性、基質形態學特性、腔形 態學特性、細胞質形態學特性、和組織背景形態學特性所 構成的形態學特性群組中之一或更多形態學特性。 90. 如申請專利範圍第89項之裝置，其中該選定的 特性亞組包括一或更多在圖6中所列的臨床特性。 91. 如申請專利範圍第90項之裝置，其中該選定的 特性亞組包括一或更多在圖6中所列的分子特性。 92 .如申請專利範圍第9 1項之裝置，其中該預測模 型包括至少約0 · 8 8的和諧指數。 93 ·如申請專利範圍第9 1項之裝置，其中該預測模 型包括對於對數等級檢驗小於約0.000 1之p値。 94·如申請專利範圍第88項之裝置，其中該模型係 預測***癌復發率且該選定的特性亞組包括選自紅血球 形態學特性、上皮核形態學特性、基質形態學特性、腔形 態學特性、和細胞質形態學特性所構成的形態學特性群組 中之一或更多形態學特性。 95.如申請專利範圍第94項之裝置，其中該所選特 性亞組包括選自圖9中所列臨床特性群組中之一或更多臨 床特性。 96·如申請專利範圍第95項之裝置，其中該選定特 性亞組包括選自圖9中所列分子特性群組中之一或更多分 子特性。 -125- 200538734 (15) 97.如申請專利範圍第96項之裝置，其中該預測模 型包括至少約〇·87之和諧指數。 9 8.如申請專利範圍第9 6項之裝置，其中該預測模 型包括對於對數等級檢驗一低於約0.000 1之p値。 9 9·如申請專利範圍第8 8項之裝置，其中該模型係 預測***癌存活率且該所選特性亞組包括選自紅血球形 態學特性、上皮核形態學特性、和基質形態學特性所構成 的形態學特性群組中之一或更多形態學特性。 100.如申請專利範圍第99項之裝置，其中該選定的 特性亞組包括選自圖1 1中所列臨床特性群組中的一或更 多臨床特性。 101·如申請專利範圍第100項之裝置，其中該選定 的特性亞組包括選自圖11中所列分子特性群組中之一或 更多分子特性。 1 02 ·如申請專利範圍第1 0 1項之裝置，其中該預測 模型包括至少約0.80的和諧指數。 1 〇3 .如申請專利範圍第1 0 1項之裝置，其中該預測 模型包括於使用對數等級檢驗時一低於約0.000 1之p値 〇 104·如申請專利範圍第82項之裝置，其中該分析工 具係經構組以對該資料組實施支援向量回歸。 105.如申請專利範圔第82項之裝置，其中該分析工 具包括一神經網路且該分析工具係經構組成實質地根據和 諧指數（CI )的近似使用一目標函數來訓練該神經網路。 -126- 200538734 (16) 1 0 6. —種篩選抑制劑化合物所用裝置，該裝置包括 一預測模型，其係以一或更多臨床特性，一或更多分 子特性，和一或更多從一或更多組織影像產生的電腦產生 之形態學特性爲基底，其中該預測模型係經構組成： 接收一患者的第一資料組； 根據該模型評估該第一資料組； 在投予試驗化合物給該患者之後接收該患者的第 二資料組；及 根據該模型評估該第二資料組； 其中從該第一資料組評估所得結果與從該第二資 料組的評定結果之比較，於該第二資料組的結果相對於該 第一資料組的結果有變化之時，即表明該試驗化合物爲一 抑制劑化合物。 107. —種評定患者的醫療狀況發生風險之裝置，該 裝置包括： 一可預測該醫療狀況之模型，其中該模型係以從一或 更多組織影像產生的一或更多電腦產生的形態學特性爲基 底且其中該模型係經構組成： 從患者接收患者資料組；及 根據該模型評估該患者資料組，藉此評定該患者的醫 療狀況之發生風險。 10 8.如申請專利範圍第107項之裝置，其中該患者 資料組包括一以該患者的肝組織影像爲基底的患者資料組 -127- 200538734 (17) 且其中該預測模型係經構組成測定該肝組織爲正常 常。 1 0 9 ·如申請專利範圍第1 0 7項之裝置，其中 係以一或更多電腦產生的形態學特性和一或更多臨 的基底。 110·如申請專利範圍第109項之裝置，其中 資料組包括一以該患者的***組織影像爲基底之 φ 料組且其中該預測模型係經構組或針對該患者的前 復發率作出預測。 1 11 ·如申請專利範圍第1 09項之裝置，其中 資料組包括以該患者的***組織影像爲基底之患 組且其中該預測模型係經構組成針對該患者的臨床 出預測。 1 1 2 . —種電腦可讀媒體，其上記錄著電腦可 令以實施包括下列步驟之方法： # 接收一患者的患者資料組；及 使用可預測醫療狀況的模型評定該患者資料組 該模型係以一或更多臨床特性、一或更多分子特性 一或更多組織影像產生的一或更多電腦產生的形態 爲基底，藉此評定該患者的醫療狀況發生風險。 該模型 床特性 該患者 患者資 列腺癌 該患者 者資料 特性作 執行指 ，其中 、和從 學特性 -128-200538734 (1) 10. Scope of patent application1. A method for assessing the risk of a patient's medical condition, the method includes: receiving a patient data set of the patient; and evaluating the patient data set using a model that predicts the medical condition, wherein the The model is based on one or more clinical characteristics, one or more molecular characteristics, and one or more computer-generated morphological characteristics generated from one or more tissue images. The risk of the situation. 2. The method of claim 1, further comprising using a computer to generate one or more morphological characteristics from the patient's tissue image for inclusion in the patient data set, wherein the generation of the morphological characteristics includes: Segment the tissue image into one or more objects; classify the one or more objects into one or more object categories; and determine the morphological characteristics by taking one or more measurements on the one or more object categories. 3. The method according to item 2 of the patent application, wherein the step of classifying one or more objects into one or more object categories includes classifying one or more objects separately from the matrix, cytoplasm, epithelial nucleus, stromal nucleus, cavity , Red blood cells, tissue orthotics, and tissue background. 4. The method according to item 2 of the patent application, wherein the one or more measurements for one or more object categories include performing one or more one or more spectral properties for one or more object categories and / Or measurement of one or more shape properties. -112- 200538734 (2) 5. The method according to item 1 of the scope of patent application, wherein the model predicts the recurrence rate of prostate cancer and one or more morphological characteristics are derived from morphological characteristics of red blood cells, morphological characteristics of epithelial nucleus, matrix Morphological characteristics, morphological characteristics of cavity, morphological characteristics of cytoplasm, and morphological characteristics of tissue background. 6. The method of claim 5 in which one or more clinical characteristics are from the clinical characteristics group listed in FIG. p 7. The method according to item 6 of the patent application, wherein one or more molecular characteristics are from the clinical characteristic group listed in FIG. 8. The method of claim 7 in the scope of patent application, wherein the prediction model includes a harmony index of at least about 0.88. 9-The method of claim 7 in the scope of patent application, wherein the prediction model includes a p 値 below about 0.001 for a log-level test. 1 0. The method according to item 1 of the scope of patent application, wherein the model predicts the recurrence rate of prostate cancer and one or more morphological characteristics are derived from morphological characteristics of red blood cells, morphological characteristics of epithelial nuclei, morphological characteristics of stroma, The morphological characteristics group consisting of luminal morphological characteristics and cytoplasmic morphological characteristics is 〇11. For the method in the scope of patent application No. 10, one or more clinical characteristics come from the clinical characteristic groups listed in FIG. 9 group. 12. The method according to item 11 of the patent application scope, wherein one or more molecular properties are from the molecular property group listed in FIG. 9. 1 3. The method of claim 12 in the patent application range, wherein the prediction model includes a harmony index of at least about 0.87. -113- 200538734 (3) 14. The method according to item 12 of the patent application scope, wherein the prediction model includes a p 値 which is lower than about 0.0 0 0 1 for a logarithmic rank test. 15 · The method according to item 1 of the scope of patent application, wherein the method predicts the survival rate of prostate cancer and one or more morphological characteristics are derived from morphological characteristics of red blood cells, morphological characteristics of epithelial nuclei, and morphological characteristics of stroma. Morphological characteristics group. 16. The method of claim 15 in which one or more of the clinical characteristics are from the clinical characteristics group listed in FIG. 11. 17 · The method according to item 16 of the patent application, wherein one or more of the molecular properties are from the molecular property group listed in FIG. 11. 18. The method of claim 17 in the scope of patent application, wherein the prediction model includes a harmony index of at least about 0.8. 19. The method of claim 17 in the scope of patent application, wherein the prediction model includes a P 値 below a logarithmic level test of about 0.0001. 20. The method of claim 1 in the scope of patent application, wherein the step of assessing the patient data set using the predictive model includes determining a diagnostic score for the patient. 2 1. The method according to item 1 of the scope of patent application, wherein the step of using the predictive model to assess the patient information includes the possibility of the patient's medical condition 22. If the method according to item 1 of the scope of patent application 'wherein the evaluation of a medical condition The risky step involves determining the patient's likely or unresponsiveness to a treatment. 23 • The method of applying for the scope of patent application No. 1 'further includes outputting information indicating the evaluation result. -114-200538734 (4) 24. The method according to item 23 of the patent application, wherein the result includes information from a diagnostic score indicating one or more characteristics of the patient data set obtained by analyzing the prediction model, indicating that the prediction model is accurate Sexual information, or results in a group of results. 25. The method of claim 23, wherein the step of receiving the patient data set of the patient includes receiving the patient's data from a remote location and wherein the output of the data indicating the evaluation result includes transmitting the result to a remote position. 26. The method of claim 25, wherein the receiving and the transmitting include receiving and transmitting on one or more communication networks. 27. A method of assessing the risk of prostate cancer recurrence in a patient 'The method includes: receiving a patient profile of the patient; and assessing the patient profile using a model predicting prostate cancer recurrence' wherein the model is based on androgen (AR ) 'S staining index as a benchmark to assess the patient's risk of prostate cancer recurrence. 2 8. — A method of generating a model that predicts a medical condition 'The method includes = receiving one message from each of two or more subjects whose results for the medical condition are at least partially known Training subject data set 'the data set includes one or more clinical characteristics, one or more molecular characteristics, and one or more computer-generated morphological characteristics generated from tissue images; and The data set performed a multivariate analysis' to generate a model that predicts the medical condition. -115- 200538734 (5) 2 9 · The method of claim 28, further comprising using a computer to generate one or more morphological characteristics from the tissue image, wherein the step of generating the morphological characteristics includes: Segment a tissue image into one or more objects; classify the one or more objects into one or more object categories; and determine the morphological characteristics by performing one or more measurements on the one or more object categories. • 30. The method of claim 29, wherein the step of classifying the one or more objects into one or more object categories includes classifying each of the one or more objects into a matrix , Cytoplasm, epithelial nucleus, stromal nucleus, cavity, red blood cells, tissue crops, and tissue background. 31. The method of claim 29, wherein the step of performing one or more measurements on the one or more object classes includes performing one or more spectral properties of the one or more object classes and / Or one or more measurements of one or more shape properties. 3 2. The method of claim 28, wherein the one or more clinical characteristics are from the clinical characteristics group listed in Table 4. 33. The method of claim 28, wherein one or more molecular properties are from the molecular property group listed in Table 6. 34. The method of claim 28, wherein the model is based on one or more clinical characteristics, one or more molecular characteristics, and one or more computer-generated morphology of the selected training target data set. Based on a subgroup of characteristics, the subgroup is selected to predict the medical condition. -116- 200538734 (6) 35. The method according to item 34 of the scope of patent application, wherein the model predicts the recurrence rate of prostate cancer and the selected specific subgroup includes morphological characteristics from red blood cells, morphological characteristics of epithelial nuclei, matrix Morphological characteristics, morphological characteristics of cavities, morphological characteristics of cytoplasm, and morphological characteristics of tissue background morphological characteristics. 36. The method of claim 35, wherein the selected characteristic subgroup includes one or more clinical characteristics listed in Table 6. 37. The method of claim 36, wherein the selected property subgroup includes one or more molecular properties listed in Table 6. 38. The method of claim 37, wherein the prediction model includes a harmony index of at least about 0.88. 39. The method of claim 37, wherein the prediction model includes a P 値 below about 0.0001 for a log-level test. 40. The method of claim 34, wherein the model predicts the recurrence rate of prostate cancer and the selected specific subgroup includes morphological characteristics from red blood cells, morphological characteristics of epithelial nuclei, morphological characteristics of stroma, and morphological characteristics of cavity And one or more morphological characteristics of the morphological characteristics group formed by cytoplasmic and morphological characteristics. 41. The method of claim 40, wherein the selected characteristic subgroup includes one or more clinical characteristics from the clinical characteristic group listed in FIG. 42. The method of claim 41, wherein the selected subset of properties includes one or more clinical properties from the molecular property group listed in FIG. -117- 200538734 (7) 43. The method according to item 42 of the patent application scope, wherein the prediction model includes a harmony index of at least about 0.87. 44. The method as claimed in item 42 of the patent application, wherein the prediction model includes a P 値 below about 0.0001 for a log-level test. 45. The method of claim 34, wherein the model predicts the recurrence rate of prostate cancer and the selected specific subgroup includes morphology derived from morphological characteristics of red blood cells, morphological characteristics of epithelial nuclei, and morphological characteristics of stroma. One or more morphological characteristics in the morphological group. 46. The method of claim 45, wherein the selected characteristic subgroup includes one or more clinical characteristics from the clinical characteristic group listed in FIG. 11. 47. The method of claim 46, wherein the selected characteristic subgroup includes one or more clinical characteristics from one of the molecular characteristic groups listed in Figure 11. 48. The method of claim 47, wherein the prediction model includes a harmony index of at least about 0.80. 49. The method according to item 47 of the patent application scope, wherein the prediction model includes a P 値 below about 0.0001 for a logarithmic level test. 50. The method of claim 28, wherein the step of performing multivariate analysis includes implementing support vector regression in the data set. 51. The method of claim 28, wherein the step of performing the multivariate analysis includes training a neural network using an objective function that is substantially approximated by a harmony index (CI). 52 · —A method for screening an inhibitor compound, the method comprising: -118- 200538734 (8) receiving a first data set of a patient; evaluating the first data set using a model predicting a medical condition, wherein the model Based on one or more clinical characteristics, one or more molecular characteristics, and one or more computer-generated morphological characteristics generated from one or more tissue images; administration of the test compound to the patient; administration of the test compound Receiving a second data set for the patient after the test compound; evaluating the second data set using the model; and comparing the evaluation results of the first data set with the evaluation results of the second data set, where the second data set A change in the results relative to the results obtained in the first data set indicates that the test compound is an inhibitor compound 053. A device used to assess the risk of a patient's medical condition, the device includes: a predictable medical condition Model, wherein the model is based on one or more clinical characteristics, one or more molecular characteristics, and one or more tissue images from one or more tissue images. Based on the computer-generated morphological characteristics of the student, the model is composed of: a patient data set that receives the patient; and the model evaluates the patient data set to assess the risk of the patient's medical condition. 54. The device of claim 53, further comprising an image processing tool configured to generate one or more morphological characteristics from a tissue image of the patient for inclusion in the patient data set , -119- 200538734 (9) The step of generating morphological characteristics includes: segmenting the tissue image into one or more objects using the image processing tool; classifying the one or more targets into one or more objects using the image processing tool Multiple object categories; and using the image processing tool to perform one or more measurements on the one or more object categories to determine the morphological characteristics. 5 5. The device as claimed in claim 54, wherein the step of classifying one or more objects into one or more object categories using the image processing tool includes separately classifying one or more objects with the image processing tool Form a category selected from the group consisting of stromal, cytoplasmic, epithelial nucleus, stromal nucleus, cavity, red blood cells, tissue bridge crops, and tissue background. 56. The device of claim 54, wherein the step of using the image processing tool to perform one or more measurements on the one or more categories includes using the image processing tool to one or more of the one or more object categories. One or more measurements of more spectral properties and / or one or more shape properties. 57. The device according to item 53 of the scope of patent application, wherein the model predicts the recurrence rate of prostate cancer and one or more morphological characteristics are selected from morphological characteristics of red blood cells, morphological characteristics of epithelial nucleus, morphological characteristics of stroma, and cavity Morphological characteristics, cytoplasmic morphological characteristics, and morphological characteristics of tissue background. 5 8. The device according to item 57 of the patent application scope, wherein the one or more clinical characteristics are from the clinical characteristic group listed in FIG. 6. 59. The device according to item 58 of the application, wherein the one or more -120-200538734 (10) multimolecular properties are from the molecular property group listed in FIG. 60. The device of claim 59, wherein the prediction model includes a harmony index of at least about 0.88. 6 1. The device according to item 59 of the patent application scope, wherein the prediction model includes a p 値 which is lower than about 0 · 0 0 0 1 for a logarithmic level test. 62. The device according to item 53 of the scope of patent application, wherein the model predicts the recurrence rate of prostate cancer and one or more morphological characteristics are derived from red blood cell φ morphological characteristics, epithelial nucleus morphological characteristics, matrix morphological characteristics, cavity Morphological characteristics group consisting of morphological characteristics and cytoplasmic morphological characteristics. 63. For the device with the scope of application for patent No. 62, one or more clinical characteristics are from the clinical characteristic group listed in FIG. 9. 64. As for the device in the scope of application for item 63, one or more of the molecular properties are from the molecular property group listed in FIG. 9. 65. The device according to item 64 of the patent application scope, wherein the prediction mode Φ type includes a harmony index of at least about 0.87. 6 6. The device according to item 64 of the patent application scope, wherein the prediction model includes a p 値 which is less than about 0.001 for a logarithmic level test. 67. The device according to item 53 of the scope of patent application, wherein the model predicts the survival rate of prostate cancer and one or more morphological characteristics are selected from morphological characteristics of red blood cells, morphological characteristics of epithelial nuclei, and morphological characteristics of stroma. The morphological characteristics group formed. 68. The device according to item 67 of the patent application, wherein one or more clinical characteristics are selected from the clinical characteristics group listed in FIG. 11. -121 200538734 (11) 69. The device according to item 68 of the patent application, wherein one or more of the molecular properties are selected from the molecular property group listed in FIG. 11. 70. The device of claim 69, wherein the prediction model includes a harmony index of at least about 0.8. 7 1 · The device according to item 69 of the patent application range, wherein the prediction model includes a check for logarithmic ranks of ρ 値 which is lower than about 0. 〇 〇 〇 1. 7 2. The device according to item 53 of the scope of patent application, wherein the predictive model # is a constitutional composition to determine the diagnostic score of the patient. 7 3 · The device according to item 53 of the scope of patent application, wherein the prediction model is a constitutional composition to determine the likely time of occurrence of the medical condition of the patient. 74. The device of claim 53, wherein the predictive model is a constitutional composition to determine a patient's possible responsiveness or responsiveness to a treatment. 7 5 · The device according to item 53 of the scope of patent application, wherein the prediction model is further configured to output data indicating the evaluation result. • 76. The device according to item 75 of the patent application scope, wherein the result includes information selected from a diagnostic score indicating one or more characteristics of the patient data set analyzed by the prediction model, indicating the prediction model's Accurate information, or results in a combination of results. 77. The device of claim 75, wherein the prediction model is composed of a patient data set that receives the patient from a remote location and outputs the results for transmission to the remote location. 78. The device according to item 77 of the patent application scope, wherein the patient data set is received through one or more communication networks and the result is transmitted through one or more -122- 200538734 (12) multiple communication networks. 79. The device of claim 53, wherein the prediction model includes a neural network. 80. The device according to item 53 of the patent application, wherein the prediction model includes a support vector machine. 81. A device for assessing the risk of prostate cancer recurrence in a patient, the device includes: φ a model for predicting the recurrence rate of prostate cancer, wherein the model is based on the androgen receptor (AR) staining index, and wherein The model was constructed to: receive the patient profile of the patient; and predict the patient profile to assess the patient's risk of prostate cancer recurrence. 82. —A device for generating a model for predicting a medical condition, the device comprising: • an analysis tool, which is composed of: for each of two or more subjects, the results of the medical condition are at least partly The known person receives a training subject data set including one or more clinical characteristics, one or more molecular characteristics, and one or more computer-generated morphological characteristics generated from tissue images; and the training subject The patient data group performs multivariate analysis to generate models that predict medical conditions. 83. The device of claim 82, further comprising an image processing tool configured to generate a -123- 200538734 (13) or more morphological characteristics from the tissue image using a computer. The step of morphological characteristics includes: segmenting the tissue image into one or more objects using the image processing tool, classifying the one or more objects into one or more object categories using the image processing tool, and using the image processing tool to One or more object categories take one or more measurements. 84. The device according to claim 83, wherein the step of classifying the one or more objects into one or more object categories using the image processing tool includes classifying each of the one or more objects with the image processing tool. One is classified into a category selected from the group consisting of stromal, cytoplasm, epithelial nucleus, stromal nucleus, cavity, red blood cells, tissue crops, and tissue background. 85. The device of claim 83, wherein the step of performing one or more measurements on one or more object categories with the image processing tool includes using the image processing tool to target one or more target categories. One or more measurements of one or more spectral properties and / or one or more shape properties. 86. The device of claim 82, wherein the one or more clinical characteristics are selected from the clinical characteristics group listed in Table 4. 87. The device of claim 82, wherein the one or more molecular properties are selected from the group of molecular properties listed in Table 6. 88. The device of claim 82, wherein the model is based on the one or more clinical characteristics, one or more molecular characteristics, and one or more computer-generated morphological characteristics in the training target data set. Based on the selected subgroup formed, the subgroup is selected to predict the medical condition. -124- 200538734 (14) 89. The device of claim 88 in the scope of patent application, wherein the model predicts the recurrence rate of prostate cancer and the selected characteristic subgroup includes selected from the morphological characteristics of red blood cells, morphological characteristics of epithelial nucleus, matrix One or more of the morphological characteristics formed by the morphological characteristics, cavity morphological characteristics, cytoplasmic morphological characteristics, and tissue background morphological characteristics. 90. The device of claim 89, wherein the selected characteristic subgroup includes one or more of the clinical characteristics listed in FIG. 91. The device of claim 90, wherein the selected characteristic subgroup includes one or more of the molecular characteristics listed in FIG. 92. The device of claim 91, wherein the prediction model includes a harmony index of at least about 0.88. 93. The device as claimed in claim 91, wherein the prediction model includes p 値 which is less than about 0.0001 for a log-level test. 94. The device according to item 88 of the scope of patent application, wherein the model predicts the recurrence rate of prostate cancer and the selected characteristic subgroup includes selected from the morphological characteristics of red blood cells, morphological characteristics of epithelial nuclei, morphological characteristics of stroma, and morphology of cavity One or more morphological characteristics in the morphological characteristic group consisting of characteristics and cytoplasmic morphological characteristics. 95. The device of claim 94, wherein the selected characteristic subgroup includes one or more clinical characteristics selected from the clinical characteristic group listed in FIG. 96. The device of claim 95, wherein the selected characteristic subgroup includes one or more molecular characteristics selected from the molecular characteristic group listed in FIG. -125- 200538734 (15) 97. The device of claim 96, wherein the prediction model includes a harmony index of at least about 0.87. 9 8. The device of claim 96 in the scope of patent application, wherein the prediction model includes a p 値 below about 0.0001 for a logarithmic level test. 9 9. The device according to item 88 of the scope of patent application, wherein the model predicts the survival rate of prostate cancer and the selected characteristic subgroup includes a member selected from the group consisting of morphological characteristics of red blood cells, morphological characteristics of epithelial nuclei, and morphological characteristics of stroma. One or more morphological characteristics of the formed morphological characteristic group. 100. The device of claim 99, wherein the selected subset of characteristics includes one or more clinical characteristics selected from the group of clinical characteristics listed in FIG. 11. 101. The device of claim 100, wherein the selected characteristic subgroup includes one or more molecular characteristics selected from the molecular characteristic group listed in FIG. 1 02 · The device according to item 101 of the patent application range, wherein the prediction model includes a harmony index of at least about 0.80. 1 〇3. If the device of the scope of patent application 101, the prediction model includes a p 値 〇104 which is lower than about 0.0001 when using a logarithmic level test. · The device of the scope of patent application 82, where The analysis tool is structured to perform support vector regression on the data set. 105. The device of claim 82, wherein the analysis tool includes a neural network and the analysis tool is constructed to substantially train the neural network using an objective function according to the approximation of the harmony index (CI) . -126- 200538734 (16) 1 0 6. —A device for screening inhibitor compounds, the device includes a predictive model based on one or more clinical characteristics, one or more molecular characteristics, and one or more from Computer-generated morphological characteristics from one or more tissue images are used as a basis, wherein the prediction model is composed of: receiving a first data set of a patient; evaluating the first data set based on the model; Receiving the patient's second data set after giving it to the patient; and evaluating the second data set according to the model; wherein the comparison between the evaluation result from the first data set and the evaluation result from the second data set is in the first When the result of the second data set is changed compared to the result of the first data set, it indicates that the test compound is an inhibitor compound. 107. A device for assessing the risk of a medical condition of a patient, the device comprising: a model predicting the medical condition, wherein the model is based on one or more computer-generated morphology generated from one or more tissue images The characteristics are base and the model is composed of warp: receiving a patient data set from a patient; and evaluating the patient data set according to the model to assess the risk of the patient's medical condition. 10 8. The device as claimed in claim 107, wherein the patient data set includes a patient data set based on the patient's liver tissue image -127- 200538734 (17), and wherein the prediction model is a constitutional composition determination The liver tissue is normal. 1 0 9 The device according to item 107 of the scope of patent application, wherein it is based on one or more computer-generated morphological characteristics and one or more adjacent substrates. 110. The device according to item 109 of the patent application, wherein the data set includes a φ data set based on the patient's prostate tissue image and wherein the prediction model is a warp group or a prediction is made for the patient's previous recurrence rate. 1 11 · The device according to the scope of patent application No. 1 09, wherein the data set includes the patient group based on the patient's prostate tissue image, and wherein the prediction model is a prediction of the clinical composition of the patient. 1 1 2. — A computer-readable medium having recorded thereon a method that the computer can perform to include the following steps: # Receive a patient data set for a patient; and evaluate the patient data set using a model that predicts a medical condition It is based on one or more clinical characteristics, one or more molecular characteristics, or one or more computer-generated morphologies generated from tissue images, thereby assessing the risk of the patient's medical condition. The model bed characteristics of the patient patient with adenocarcinoma of the patient's personal information are used as executive instructions, of which, and learning characteristics -128-