WO2023001190A1 - 结直肠息肉图像的识别方法、装置及存储介质 - Google Patents

结直肠息肉图像的识别方法、装置及存储介质 Download PDF

Info

Publication number
WO2023001190A1
WO2023001190A1 PCT/CN2022/106764 CN2022106764W WO2023001190A1 WO 2023001190 A1 WO2023001190 A1 WO 2023001190A1 CN 2022106764 W CN2022106764 W CN 2022106764W WO 2023001190 A1 WO2023001190 A1 WO 2023001190A1
Authority
WO
WIPO (PCT)
Prior art keywords
colorectal
image
colorectal polyp
polyp
classification
Prior art date
Application number
PCT/CN2022/106764
Other languages
English (en)
French (fr)
Inventor
李佳昕
王玉峰
Original Assignee
天津御锦人工智能医疗科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 天津御锦人工智能医疗科技有限公司 filed Critical 天津御锦人工智能医疗科技有限公司
Priority to EP22845355.1A priority Critical patent/EP4350617A1/en
Publication of WO2023001190A1 publication Critical patent/WO2023001190A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/56—Extraction of image or video features relating to colour
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/778—Active pattern-learning, e.g. online learning of image or video features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10068—Endoscopic image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30028—Colon; Small intestine
    • G06T2207/30032—Colon polyp
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/03—Recognition of patterns in medical or anatomical images
    • G06V2201/032—Recognition of patterns in medical or anatomical images of protuberances, polyps nodules, etc.

Definitions

  • the present application relates to the field of medical technology, and in particular to a colorectal polyp image recognition method, device and storage medium.
  • Colorectal polyps refer to protruding lesions protruding from the intestinal lumen, and are one of the common intestinal diseases, including adenomatous polyps and non-adenomatous polyps (hamartomatous polyps, metaplastic polyps and inflammatory polyps) , of which adenomatous polyps are precancerous lesions.
  • adenomatous polyps and non-adenomatous polyps hamartomatous polyps, metaplastic polyps and inflammatory polyps
  • the morbidity and mortality of colorectal cancer in China are increasing year by year. Most patients are in the middle and advanced stages when they are discovered, and the 5-year survival rate is less than 50%. Colonoscopy is the most intuitive and effective way to find lesions, which can improve the detection of early intestinal cancer and reduce mortality.
  • Electronic chromoendoscopy is an important progress in the field of digestive endoscopy in recent years, such as narrow-band imaging (NBI), endoscopic intelligent spectroscopic colorimetry (FICE), high-definition intelligent electronic chromoendoscopy (i-scan), etc. Improved detectability of colorectal adenomas.
  • NBI narrow-band imaging
  • FICE endoscopic intelligent spectroscopic colorimetry
  • i-scan high-definition intelligent electronic chromoendoscopy
  • Improved detectability of colorectal adenomas Studies have shown that electronic staining light technology has similar and high sensitivity, specificity and accuracy in distinguishing colorectal polypoid lesions from tumors and non-tumors.
  • Electronic staining typing has good application value in the diagnosis and treatment of colorectal lesions . However, at present, it still relies on doctors to judge the results of typing based on experience and visual observation, and there are subjective differences.
  • the present application provides a colorectal polyp image recognition method, device and storage medium.
  • the present application provides a method for identifying a colorectal polyp image, and the method for identifying a colorectal polyp image includes:
  • it may further include: evaluating the type of colorectal polyps in the colorectal region according to the recognition result.
  • the identification method of the colorectal polyp image also includes:
  • the training classification model includes:
  • the gradient feature fusion map of the colorectal polyp sample image is trained
  • the Mish function is:
  • x is each pixel value of the output feature map of ResNeSt Block.
  • the gradient feature fusion map of the colorectal polyp sample image is trained, including:
  • r second feature maps After performing two convolution operations of 1*1 and 3*3 on the k*r first feature maps, r second feature maps will be obtained again in each group of group convolutions, and the r second feature maps will be fused , input to the Split Attention layer to strengthen features through the attention mechanism, and get k results;
  • optimizing the classification model through the loss function further includes:
  • the positive samples are weighted to calculate the loss value.
  • the color equalization process includes:
  • the image is an image of a colorectal region or a colorectal polyp sample image.
  • the edge feature map fusion includes:
  • the zoomed image and the edge feature map are fused into a gradient feature fusion map.
  • the present application provides a colorectal polyp type result prediction device, the colorectal polyp type result prediction device includes: a memory, a processor, and stored in the memory and can be on the processor running computer programs;
  • the present application provides a computer-readable storage medium, the computer-readable storage medium stores a colorectal polyp typing result prediction program, and the colorectal polyp NICE typing result prediction program is executed by a processor , implementing the steps including the recognition method for colorectal polyp images as described in any one of the above.
  • the present application provides a method for performing colonoscopy on a subject, the method including using any one of the above-mentioned colorectal polyp image recognition methods, or the above-mentioned colorectal polyp
  • the polyp type result prediction device performs real-time prediction or judgment on the type of colorectal polyps under the microscope of the subject.
  • the colorectal polyp classification refers to the morphological classification of non-neoplastic polyps, neoplastic polyps, or adenocarcinoma.
  • the classification standard may be the classification standard under the traditional endoscope or the classification standard under the magnifying glass combined with narrow-band imaging (ME-NBI) endoscope.
  • the classification standard under the ME-NBI endoscope includes, for example, NICE classification standard, Sano classification standard, Pit classification standard, Hiroshima classification standard, Showa classification standard, Jikei classification standard, JNET classification standard standard.
  • NICE classification criteria Preferably the NICE classification criteria.
  • the colorectal polyp image may be an image acquired in real time under an endoscope, or may be a stored image.
  • Each embodiment of the present application builds a recognition sample by performing color equalization processing and edge feature map fusion processing on the colorectal part image, which can effectively unify and enhance image features, and perform feature extraction through ResNeSt Block, which can more accurately predict colorectal polyps
  • the typing results can assist doctors in inferring the pathological properties of colorectal polyps.
  • Fig. 1 is a flow chart of the prediction method for colorectal polyp NICE typing results provided by various embodiments of the present application;
  • Fig. 2 is a schematic diagram of a feature extraction network based on a ResNeSt Block provided by various embodiments of the present application;
  • FIG. 3 is a graph of the improved Mish function of various embodiments of the present application.
  • colonal polyp classification refers to the morphological classification of non-neoplastic polyps, neoplastic polyps or adenocarcinoma;
  • classification standard can be, for example, NICE classification criteria, Sano classification standard, Pit classification standard, Hiroshima classification standard, Showa classification standard, Jikei classification standard, JNET classification standard.
  • An embodiment of the present invention provides a method for predicting the result of NICE typing of colorectal polyps, as shown in FIG. 1 , the method for predicting the result of NICE typing of colorectal polyps includes:
  • the color equalization processing and edge feature map fusion processing are performed on the colorectal part image to construct the recognition sample, which can effectively unify and strengthen the image features, and perform feature extraction through the ResNeSt Block, so that a deep learning-based
  • the computer-aided NICE typing system can more accurately predict the NICE typing results of colorectal polyps, and assist doctors in inferring the pathological properties of colorectal polyps.
  • images of colorectal polyps are acquired through white light, NBI, FICE, BLI, LCI, I-SCAN and other modes.
  • NICE classification results include type 1, type 2, type 3, and normal intestinal tract.
  • the NICE classification classification model based on the attention splitting module ResNeSt Block of the pre-built call it includes:
  • the classification model is trained; and in the process of training the classification model, the classification model is optimized according to the loss function; the loss function optimizes the classification model.
  • Described classification model; Described loss function formula is as follows:
  • the classification model of the NICE classification based on ResNeSt Block is obtained; in the formula, L represents the loss value, n represents the total number of categories, i represents the true category of the current data, and t represents the characteristics passed through. Extract the 1D tensor of the category probability distribution obtained by the network, t[i] represents the predicted probability value of the classification model under the category i, w(i) represents the weighting coefficient for the predicted loss value of the category i, Represents exp and summing all the values of t.
  • the embodiment of the present invention builds training samples by performing color equalization processing and edge feature map fusion processing on colorectal images, which can effectively unify and strengthen image features, thereby improving the prediction accuracy of colorectal polyp NICE typing results; according to the The above loss function optimizes the classification model, effectively improving the training speed of the model.
  • the training classification model includes:
  • the gradient feature fusion map of the colorectal polyp sample image is trained
  • the output of the ResNeSt Block is added to the gradient feature fusion map of the colorectal polyp sample image, and activated by the improved Mish function.
  • the improved Mish function is:
  • x is the pixel value of the output feature map of ResNeSt Block.
  • the embodiment of the present invention effectively improves the feature fitting ability of the network based on the above-mentioned improved Mish function.
  • the gradient feature fusion map of the colorectal polyp sample image is trained, including:
  • r second feature maps After performing two convolution operations of 1*1 and 3*3 on the k*r first feature maps, r second feature maps will be obtained again in each group of group convolutions, and the r second feature maps will be fused , input to the Split Attention layer to strengthen features through the attention mechanism, and get k results;
  • the processing flow of ResNeSt Block includes:
  • the smoother Mish function is selected as the activation function, which improves the feature fitting ability of the network.
  • the graph is shown in Figure 3.
  • each group of Block is composed of three layers of convolution, so its network connection structure is:
  • ResNet The most effective part in ResNet is the residual module.
  • its original residual structure is modified into ResNeSt Block (attention splitting module).
  • the feature extraction network can be composed of any number of ResNeSt Blocks, but its The number of Blocks is inversely proportional to the training/prediction speed; it is directly proportional to the prediction accuracy; but when the number of Blocks is too large, the degradation of the network fitting ability and the decrease in accuracy may occur.
  • 16 groups of ResNeSt Blocks are used, not only Effectively improve prediction accuracy and ensure training speed.
  • optimizing the classification model through the loss function further includes:
  • the positive samples are weighted to calculate the loss value.
  • the data set on ImageNet is used for model pre-training to ensure that the initial parameters have better feature extraction ability and generalization ability.
  • the initial learning rate of model training is 0.0001, and the learning rate is reduced by a fixed step size attenuation method.
  • the loss function is also based on the ratio of the number of positive and negative samples, and the positive samples are weighted to calculate the loss value.
  • the specific formula is:
  • L represents the loss value
  • n represents the total number of categories
  • i represents the true category of the current data
  • t represents the 1D tensor of the category probability distribution obtained through the feature extraction network
  • t[i] represents the classification model in The predicted probability value under category i
  • w(i) represents the weighting coefficient for the predicted loss value of category i
  • the optimizer uses the AdamW optimizer algorithm, which includes weight decay and L2 regularization functions, which can effectively improve the training speed and model accuracy. This model reaches the optimal solution after about 300 iterations.
  • the color equalization process includes:
  • the image is an image of a colorectal region or a colorectal polyp sample image.
  • the edge feature map fusion includes:
  • the zoomed image and the edge feature map are fused into a gradient feature fusion map.
  • the training samples are constructed by superimposing color equalization processing and gradient feature map fusion, and image features can be effectively unified and enhanced through the above two methods.
  • the values of the R, G, and B channels of the image are weighted and multiplied by Kr, Kg, and Kb respectively, for example, using the addWeighted function of python-opencv.
  • the use of color equalization processing can reduce the difference in intestinal color caused by light and the patient itself, make the color tone of the data set more uniform, and improve the learning and detection efficiency of the feature extraction network.
  • edge extraction operator such as Scharr, Canny, Sobel, Laplace, etc.
  • the scaled image H and the edge feature map J are fused into 4-channel data, which is the gradient feature fusion map.
  • Various embodiments of the present invention can construct a computer-aided NICE typing system based on deep learning, which can more accurately predict the NICE typing results of colorectal polyps, and assist doctors in inferring the pathological properties of colorectal polyps.
  • An embodiment of the present invention provides a colorectal polyp NICE typing result prediction device, the colorectal polyp NICE typing result prediction device includes: a memory, a processor, and a device stored in the memory and operable on the processor computer program;
  • the device for predicting the result of colorectal polyp NICE typing may be an endoscope detection device, and the storage may be a cloud storage.
  • An embodiment of the present invention provides a computer-readable storage medium.
  • the computer-readable storage medium stores a colorectal polyp NICE typing result prediction program.
  • the colorectal polyp NICE typing result prediction program is executed by a processor, The steps of the method for predicting the result of NICE typing of colorectal polyps as described in any one of the first embodiment are realized.
  • An embodiment of the present invention provides a method for performing a colonoscopy on a subject, the method includes using the colorectal polyp NICE classification result prediction method described in Example 1, or the colorectal polyp described in Example 2
  • the NICE typing result prediction device performs real-time prediction or judgment on the type of NICE typing of colorectal polyps under the microscope of the subject.
  • Embodiment 2 For the specific realization of Embodiment 2, Embodiment 3, and Embodiment 4, reference may be made to Embodiment 1, which have corresponding technical effects.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Image Analysis (AREA)

Abstract

本申请涉及一种结直肠息肉图像的识别方法、装置及存储介质。所述结直肠息肉图像的识别方法包括：对结直肠部位图像进行色彩均衡化处理和边缘特征图融合，得到所述结直肠部位图像的梯度特征融合图；调用预先构建的基于注意力拆分模块ResNeSt Block的结直肠息肉分型的分类模型，对所述结直肠部位图像的梯度特征融合图进行识别。本申请通过对结直肠部位图像进行色彩均衡化处理和边缘特征图融合处理构建识别样本，可有效的统一及加强图像特征，并通过基于ResNeSt Block构建的特征提取网络进行特征提取，能更准确预测结直肠息肉分型结果，辅助医生完成对结直肠息肉的病理性质推断。

Description

结直肠息肉图像的识别方法、装置及存储介质
交叉引用说明
本申请要求于2021年7月23日提交中国专利局、申请号为202110833639.8，发明名称为“结直肠息肉图像的识别方法、装置及存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及医疗技术领域，尤其涉及一种结直肠息肉图像的识别方法、装置及存储介质。
背景技术
结直肠息肉是指突出于肠腔内的***性病变，是肠道常见的疾病之一，包括腺瘤性息肉和非腺瘤性息肉(错构瘤性息肉、化生性息肉和炎症性息肉)，其中腺瘤性息肉属于癌前病变。中国结直肠癌发病率和死亡率逐年上升，多数患者被发现时已属于中晚期，5年生存率不到50％。结肠镜是发现病变最直观有效的方法，可以提高肠道早期癌的发现从而降低死亡率。电子染色内镜是近年来消化内镜检查领域的重要进展，如窄带成像(NBI)、内镜智能分光比色技术(FICE)、高清智能电子染色内镜(i-scan)等已经广泛用于改善结直肠腺瘤的可检测性。有研究表明，在鉴别结直肠息肉样病变肿瘤和非肿瘤性方面，电子染色光技术有着相似且较高的灵敏度、特异性和准确性，电子染色分型在结直肠病变诊治方面具有良好应用价值。但目前尚依赖于医生凭经验及肉眼观察判断分型的结果，存在主观差异性。
发明内容
为了解决上述技术问题或者至少部分地解决上述技术问题，本申请提供了一种结直肠息肉图像的识别方法、装置及存储介质。
第一方面，本申请提供了一种结直肠息肉图像的识别方法，所述结直肠息肉图像的识别方法包括：
对结直肠部位图像进行色彩均衡化处理和边缘特征图融合，得到所述结直肠部位图像的梯度特征融合图；
调用预先构建的基于注意力拆分模块ResNeSt Block的结直肠息肉分型的分类模型，对所述结直肠部位图像的梯度特征融合图进行识别。
进一步地，为了预测或判断结直肠息肉的分型，在上述识别方法后还可以包括：根据识别结果，评估结直肠部位的结直肠息肉分型。
可选地，所述结直肠息肉图像的识别方法还包括：
对预先获取的结直肠息肉样本图像进行色彩均衡化处理和边缘特征图融合，得到结直肠息肉样本图像的梯度特征融合图；
利用所述结直肠息肉样本图像的梯度特征融合图和预先构建的基于ResNeSt Block的特征提取网络，训练所述分类模型；并在训练所述分类模型的过程中，根据损失函数优化所述分类模型；
在所述分类模型训练完成后，得到所述基于ResNeSt Block的结直肠息肉分型的分类模型。
可选地，所述根据所述结直肠息肉样本图像的梯度特征融合图和预先构建的基于ResNeSt Block的特征提取网络，训练分类模型，包括：
在基于构建的所述ResNeSt Block特征提取网络中，对所述结直肠息肉样本图像的梯度特征融合图进行训练；
将ResNeSt Block的输出通过Mish函数激活。
可选地，所述Mish函数为：
f(x)＝x*tanh(ln(1+e x))；
式中，x为ResNeSt Block的输出特征图的每个像素值。
可选地，所述在所述ResNeSt Block特征提取网络中，对所述结直肠息肉样本图像的梯度特征融合图进行训练，包括：
对所述结直肠息肉样本图像的梯度特征融合图拆分，得到k个分组卷积；
将每个分组卷积拆分得到r个第一特征图；
对k*r个第一特征图先进行1*1和3*3两次卷积运算后，每组分组卷积中将重新得到r个第二特征图，将r个第二特征图进行融合，输入至Split Attention层通过注意力机制加强特征，得到k个结果；
将k个结果继续进行融合后输入至一个1*1的卷积层进行卷积计算，得到ResNeSt Block的输出；k和r均为超参数。
可选地，所述在训练所述分类模型过程中，通过所述损失函数优化所述分类模型还包括：
根据所述结直肠息肉样本图像的正负样本的数量配比，对正样本进行加权计算损失值。
可选地，所述色彩均衡化处理包括：
根据图像的色彩均值，确定图像的RGB通道权重值；
根据图像的RGB通道权重值，对所述图像的RGB通道加权；所述图像为结直肠部位图像或者结直肠息肉样本图像。
可选地，所述边缘特征图融合包括：
对RGB通道加权后的图像进行缩放，得到缩放图像；
将所述缩放图像进行边缘提取，得到边缘特征图；
将所述缩放图像与所述边缘特征图融合成梯度特征融合图。
第二方面，本申请提供了一种结直肠息肉分型结果预测装置，所述结直肠息肉分型结果预测装置包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序；
所述计算机程序被所述处理器执行时，实现包括如上任一项所述的结直肠息肉图像的识别方法的步骤。
第三方面，本申请提供了一种计算机可读存储介质，所述计算机 可读存储介质上存储有结直肠息肉分型结果预测程序，所述结直肠息肉NICE分型结果预测程序被处理器执行时，实现包括如上任一项所述的结直肠息肉图像的识别方法的步骤。
第四方面，本申请提供了一种对受试者进行结直肠镜检查的方法，所述方法中包括利用如上任意一项所述的结直肠息肉图像的识别方法，或如上所述的结直肠息肉分型结果预测装置对受试者镜下结直肠息肉的分型进行实时预测或判断。
在本说明书中，所述结直肠息肉分型是指针对非肿瘤性息肉、肿瘤性息肉或腺癌等在形态学上的分型。
所述分型的标准可以为传统内窥镜下的分型标准或放大镜结合窄带成像(ME-NBI)内窥镜下的分型标准。所述ME-NBI内窥镜下的分型标准包括，例如，NICE分型标准、Sano分型标准、Pit分型标准、Hiroshima分型标准、Showa分型标准、Jikei分型标准、JNET分型标准。优选为NICE分型标准。
所述结直肠息肉图像可以为内窥镜下实时获取的图像，也可以为存储的图像。
本申请实施例提供的上述技术方案与现有技术相比具有如下优点：
本申请各实施例通过对结直肠部位图像进行色彩均衡化处理和边缘特征图融合处理构建识别样本，可有效的统一及加强图像特征，并通过ResNeSt Block进行特征提取，能更准确预测结直肠息肉分型结果，辅助医生完成对结直肠息肉的病理性质推断。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，示出了符合本发明的实施例，并与说明书一起用于解释本发明的原理。
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，对于本领域普通技术人员而言，在不付出创造性劳动性的前 提下，还可以根据这些附图获得其他的附图。
图1为本申请各个实施例提供的结直肠息肉NICE分型结果预测方法的流程图；
图2为本申请各个实施例提供的基于一个ResNeSt Block的特征提取网络示意图；
图3为本申请各个实施例的改进的Mish函数的曲线图。
具体实施方式
应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于限定本发明。
在后续的描述中，使用用于表示元件的诸如“模块”、“部件”或“单元”的后缀仅为了有利于本发明的说明，其本身没有特定的意义。因此，“模块”、“部件”或“单元”可以混合地使用。
在本说明书中，术语“结直肠息肉分型”是指针对非肿瘤性息肉、肿瘤性息肉或腺癌等在形态学上的分型；分型的标准可以为，例如，NICE分型标准、Sano分型标准、Pit分型标准、Hiroshima分型标准、Showa分型标准、Jikei分型标准、JNET分型标准。
实施例一
本发明实施例提供一种结直肠息肉NICE分型结果预测方法，如图1所示，所述结直肠息肉NICE分型结果预测方法包括：
S101，对结直肠部位图像进行色彩均衡化处理和边缘特征图融合，得到所述结直肠部位图像的梯度特征融合图；
S102，调用预先构建的基于注意力拆分模块ResNeSt Block的NICE分型分类模型，对所述结直肠部位图像的梯度特征融合图进行识别；
S103，根据识别结果，评估结直肠部位的结直肠息肉NICE分型类型。
本发明实施例通过对结直肠部位图像进行色彩均衡化处理和边缘 特征图融合处理构建识别样本，可有效的统一及加强图像特征，并通过ResNeSt Block进行特征提取，从而可以构建出基于深度学习的计算机辅助NICE分型***，能更准确预测结直肠息肉NICE分型结果，辅助医生完成对结直肠息肉的病理性质推断。
在具体实现过程中，通过白光、NBI、FICE、BLI、LCI、I-SCAN等模式下获取结直肠息肉图像。NICE分型结果包括一型、二型、三型、正常肠道。
在一些实施例中，所述调用预先构建的基于注意力拆分模块ResNeSt Block的NICE分型分类模型之前，包括：
对预先获取的结直肠息肉样本图像进行色彩均衡化处理和边缘特征图融合，得到结直肠息肉样本图像的梯度特征融合图；
根据所述结直肠息肉样本图像的梯度特征融合图和预先构建的基于ResNeSt Block的特征提取网络，训练分类模型；并在训练分类模型过程中，根据损失函数优化所述分类模型；损失函数优化所述分类模型；所述损失函数公式如下：
Figure PCTCN2022106764-appb-000001
在所述分类模型训练完成后，得到所述基于ResNeSt Block的NICE分型的分类模型；式中L代表损失值，n代表总类别数量，i代表当前数据的真实类别，t代表经过所述特征提取网络得到的类别概率分布的1D张量，t[i]代表所述分类模型在i类别下的预测概率值，w(i)代表对i类别预测损失值的加权系数，
Figure PCTCN2022106764-appb-000002
代表对t的所有数值进行exp再加和。
本发明实施例通过对结直肠部位图像进行色彩均衡化处理和边缘 特征图融合处理构建训练样本，可有效的统一及加强图像特征，进而提高结直肠息肉NICE分型结果的预测准确性；根据所述损失函数优化所述分类模型，有效提高模型的训练速度。
在一些实施例中，所述根据所述结直肠息肉样本图像的梯度特征融合图和预先构建的基于ResNeSt Block的特征提取网络，训练分类模型，包括：
在所述ResNeSt Block特征提取网络中，对所述结直肠息肉样本图像的梯度特征融合图进行训练；
将ResNeSt Block的输出与所述结直肠息肉样本图像的梯度特征融合图相加，并通过改进的Mish函数激活。
可选地，所述改进的Mish函数为：
f(x)＝x*tanh(ln(1+e x))；
式中，x为ResNeSt Block的输出特征图的像素值。
本发明实施例基于上述改进的Mish函数有效提高了网络的特征拟合能力。
可选地，所述在所述ResNeSt Block特征提取网络中，对所述结直肠息肉样本图像的梯度特征融合图进行训练，包括：
对所述结直肠息肉样本图像的梯度特征融合图拆分，得到k个分组卷积；
将每个分组卷积拆分得到r个第一特征图；
对k*r个第一特征图先进行1*1和3*3两次卷积运算后，每组分组卷积中将重新得到r个第二特征图，将r个第二特征图进行融合，输入至Split Attention层通过注意力机制加强特征，得到k个结果；
将k个结果继续进行融合后输入至一个1*1的卷积层进行卷积计算，得到ResNeSt Block的输出；k和r均为超参数。
详细地，如图2所示，ResNeSt Block的处理流程包括：
a)将input(输入梯度特征融合图)进行拆分，得到Cardinal 1-Cardinal k，共计k个分组卷积；
b)每个分组卷积中继续拆分图像，得到Split 1–Split r，共计r个第一特征图；
c)之后对k*r个第一特征图先进行1*1的卷积计算，再进行3*3的卷积计算，两次卷积运算后，每组分组卷积中将重新得到r个特征图，将其进行融合，输入至Split Attention层通过注意力机制加强特征，最终得到k个结果；
d)将k个结果继续进行融合，结果输入至一个1*1的卷积层进行卷积计算，得到输出output
e)ResNeSt Block的输出与input相加，进行激活函数激活，得到最终特征图。
其中，激活函数选择更加平滑的Mish函数，提升了网络的特征拟合能力，曲线图如图3所示。
ResNeSt Block特征提取网络中共有16组ResNeSt Block，每组Block中又由三层卷积构成，所以其网络连接结构为：
Input→Conv*1→ResNeSt Block(Conv*2+Attention+Conv*1)*16→全连接层。
ResNet中最有效果的部分为残差模块，本发明实施例中其原本的残差结构修改为了ResNeSt Block(注意力拆分模块)，所述特征提取网络可由任意数量的ResNeSt Block组成，但其Block数量与训练/预测速度成反比；与预测准确率成正比；但当Block数量过多时可能会出现网络拟合能力退化、准确率下降的情况，本发明实施例中采用16组ResNeSt Block，不仅有效提高预测准确性，并保证了训练速度。
可选地，所述在训练分类模型过程中，通过所述损失函数优化所述分类模型还包括：
根据所述结直肠息肉样本图像的正负样本的数量配比，对正样本进行加权计算损失值。
进一步地，训练分类模型过程中，使用了ImageNet上的数据集上进行模型预训练，保证起始参数具有较优的特征提取能力及泛化能力。 模型训练的初始学习率均为0.0001，采用固定步长衰减的方式减小学习率，基于此，损失函数同时基于正负样本的数量配比，对正样本进行加权计算损失值，具体公式为：
Figure PCTCN2022106764-appb-000003
(式中L代表损失值，n代表总类别数量，i代表当前数据的真实类别，t代表经过所述特征提取网络得到的类别概率分布的1D张量，t[i]代表所述分类模型在i类别下的预测概率值，w(i)代表对i类别预测损失值的加权系数，
Figure PCTCN2022106764-appb-000004
代表对x的所有数值进行exp再加和。)。
优化器均采用AdamW优化器算法，该优化器包含了权重衰减及L2正则化功能，可以有效提高训练速度及模型准确度，本模型在迭代约300轮之后达到最优解。
在一些实施例中，所述色彩均衡化处理包括：
根据图像的色彩均值，确定图像的RGB通道权重值；
根据图像的RGB通道权重值，对所述图像的RGB通道加权；所述图像为结直肠部位图像或者结直肠息肉样本图像。
其中，所述边缘特征图融合包括：
对RGB通道加权后的图像进行缩放，得到缩放图像；
将所述缩放图像进行边缘提取，得到边缘特征图；
将所述缩放图像与所述边缘特征图融合成梯度特征融合图。
本发明实施例在原有的图片样本数据基础上，又叠加采用了色彩均衡化处理及梯度特征图融合的方式构建训练样本，通过上述两方式可有效的统一及加强图像特征。
其中，色彩均衡化：
1.首先将图像所有像素的R、G、B三个通道的值，分别进行加总并除以像素点个数，得到RGB各通道均值，即v(r)、v(g)、v(b)；
2.求出总均值K，K＝(v(r)+v(g)+v(b))/3；
3.计算出各通道权重值，即Kb＝K/v(b)；Kg＝K/v(g)；Kr＝K/v(r)；
4.对图像的R、G、B三通道的值分别与Kr、Kg、Kb完成加权相乘，例如使用python-opencv的addWeighted函数。
使用色彩均衡化处理可以减少光线及病人本身原因造成的肠道颜色差异，让数据集色调更统一，提升特征提取网络的学习及检测效率。
梯度特征图融合：
1.将数据集统一缩放成m像素*n像素，得到缩放图像H；例如，430像素*368像素；
2.将缩放图像转为灰度图后使用边缘提取算子(例如Scharr、Canny、Sobel、Laplace等)进行边缘提取，得到边缘特征图J；
将缩放图像H及边缘特征图J融合成4通道的数据，即为梯度特征融合图。
本发明各个实施例可以构建出基于深度学习的计算机辅助NICE分型***，能更准确预测结直肠息肉NICE分型结果，辅助医生完成对结直肠息肉的病理性质推断。
实施例二
本发明实施例提供一种结直肠息肉NICE分型结果预测装置，所述结直肠息肉NICE分型结果预测装置包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序；
所述计算机程序被所述处理器执行时，实现如实施例一中任一项所述的结直肠息肉NICE分型结果预测方法的步骤。
其中，结直肠息肉NICE分型结果预测装置可以是内窥镜检测设备，存储器可以是云存储器。
实施例三
本发明实施例提供一种计算机可读存储介质，所述计算机可读存 储介质上存储有结直肠息肉NICE分型结果预测程序，所述结直肠息肉NICE分型结果预测程序被处理器执行时，实现如实施例一中任一项所述的结直肠息肉NICE分型结果预测方法的步骤。
实施例四
本发明实施例提供一种对受试者进行结直肠镜的方法，所述方法中包括利用实施例一所述的结直肠息肉NICE分型结果预测方法，或实施例二所述的结直肠息肉NICE分型结果预测装置对受试者镜下结直肠息肉的NICE分型的类型进行实时预测或判断。
实施例二、实施例三和实施例四的具体实现可以参阅实施例一，具有相应的技术效果。
需要说明的是，在本文中，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者装置不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者装置所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括该要素的过程、方法、物品或者装置中还存在另外的相同要素。
上述本发明实施例序号仅仅为了描述，不代表实施例的优劣。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件，但很多情况下前者是更佳的实施方式。基于这样的理解，本发明的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端(可以是手机，计算机，服务器，空调器，或者网络设备等)执行本发明各个实施例所述的方法。
上面结合附图对本发明的实施例进行了描述，但是本发明并不局限于上述的具体实施方式，上述的具体实施方式仅仅是示意性的，而 不是限制性的，本领域的普通技术人员在本发明的启示下，在不脱离本发明宗旨和权利要求所保护的范围情况下，还可做出很多形式，这些均属于本发明的保护之内。

Claims (13)

  1. 一种结直肠息肉图像的识别方法，其特征在于，所述结直肠息肉图像的识别方法包括：
    对结直肠部位图像进行色彩均衡化处理和边缘特征图融合，得到所述结直肠部位图像的梯度特征融合图；
    调用预先构建的基于注意力拆分模块ResNeSt Block的结直肠息肉分型的分类模型，对所述结直肠部位图像的梯度特征融合图进行识别。
  2. 根据权利要求1所述的结直肠息肉图像的识别方法，其特征在于，所述结直肠息肉图像的识别方法还包括：
    对预先获取的结直肠息肉样本图像进行色彩均衡化处理和边缘特征图融合，得到结直肠息肉样本图像的梯度特征融合图；
    利用所述结直肠息肉样本图像的梯度特征融合图和预先构建的基于ResNeSt Block的特征提取网络，训练所述分类模型；并在训练所述分类模型的过程中，通过损失函数优化所述分类模型；所述损失函数公式如下：
    Figure PCTCN2022106764-appb-100001
    在所述分类模型训练完成后，得到所述基于ResNeSt Block的结直肠息肉分型的分类模型；式中L代表损失值，n代表总类别数量，i代表当前数据的真实类别，t代表经过所述特征提取网络得到的类别概率分布的1D张量，t[i]代表所述分类模型在i类别下的预测概率值，w(i)代表对i类别预测损失值的加权系数，
    Figure PCTCN2022106764-appb-100002
    代表对t的所有数值进行exp再加和。
  3. 根据权利要求2所述的结直肠息肉图像的识别方法，其特征在 于，所述利用所述结直肠息肉样本图像的梯度特征融合图和预先构建的基于ResNeSt Block的特征提取网络，训练所述分类模型，包括：
    在所述预先构建的基于ResNeSt Block特征提取网络中，对所述结直肠息肉样本图像的梯度特征融合图进行训练；
    将ResNeSt Block的输出通过Mish函数激活。
  4. 根据权利要求3所述的结直肠息肉图像的识别方法，其特征在于，所述Mish函数为：
    f(x)＝x*tanh(ln(1+e x))；
    式中，x为ResNeSt Block的输出特征图的像素值。
  5. 根据权利要求3所述的结直肠息肉图像的识别方法，其特征在于，所述在所述预先构建的基于ResNeSt Block特征提取网络中，对所述结直肠息肉样本图像的梯度特征融合图进行训练，包括：
    对所述结直肠息肉样本图像的梯度特征融合图拆分，得到k个分组卷积；
    将每个分组卷积拆分得到r个第一特征图；
    对k*r个第一特征图先进行1*1和3*3两次卷积运算后，每组分组卷积中将重新得到r个第二特征图，将r个第二特征图进行融合，输入至Split Attention层通过注意力机制加强特征，得到k个结果；
    将k个结果继续进行融合后输入至一个1*1的卷积层进行卷积计算，得到ResNeSt Block的输出；k和r均为超参数。
  6. 根据权利要求2所述的结直肠息肉图像的识别方法，其特征在于，所述在训练所述分类模型过程中，通过损失函数优化所述分类模型还包括：
    根据所述结直肠息肉样本图像的正负样本的数量配比，对正样本进行加权计算损失值。
  7. 根据权利要求1所述的结直肠息肉图像的识别方法，其特征在于，所述色彩均衡化处理包括：
    根据图像的色彩均值，确定图像的RGB通道权重值；
    根据图像的RGB通道权重值，对所述图像的RGB通道加权；所述图像为结直肠部位图像或者结直肠息肉样本图像。
  8. 根据权利要求7所述的结直肠息肉图像的识别方法，其特征在于，所述边缘特征图融合包括：
    对RGB通道加权后的图像进行缩放，得到缩放图像；
    将所述缩放图像进行边缘提取，得到边缘特征图；
    将所述缩放图像与所述边缘特征图融合成梯度特征融合图。
  9. 根据权利要求1所述的结直肠息肉图像的识别方法，其特征在于，所述结直肠息肉分型的标准为NICE分型标准、Sano分型标准、Pit分型标准、Hiroshima分型标准、Showa分型标准、Jikei分型标准或JNET分型标准。
  10. 根据权利要求9所述的结直肠息肉图像的识别方法，其特征在于，所述结直肠息肉分型的标准为NICE分型标准。
  11. 一种结直肠息肉分型结果预测装置，其特征在于，所述结直肠息肉分型结果预测装置包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序；
    所述计算机程序被所述处理器执行时，实现包括如权利要求1所述的结直肠息肉图像的识别方法的步骤。
  12. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质上存储有结直肠息肉分型结果预测程序，所述结直肠息肉分型结果预测程序被处理器执行时，实现包括如权利要求1所述的结直肠息肉图像的识别方法的步骤。
  13. 一种对受试者进行结直肠镜检查的方法，其特征在于，所述方法中包括利用如权利要求1所述的结直肠息肉图像的识别方法，或如权利要求11所述的结直肠息肉分型结果预测装置对受试者镜下结直肠息肉的分型进行实时预测或判断。
PCT/CN2022/106764 2021-07-23 2022-07-20 结直肠息肉图像的识别方法、装置及存储介质 WO2023001190A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
EP22845355.1A EP4350617A1 (en) 2021-07-23 2022-07-20 Colorectal polyp image recognition method, apparatus, and storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110833639.8A CN113284146B (zh) 2021-07-23 2021-07-23 结直肠息肉图像的识别方法、装置及存储介质
CN202110833639.8 2021-07-23

Publications (1)

Publication Number Publication Date
WO2023001190A1 true WO2023001190A1 (zh) 2023-01-26

Family

ID=77287044

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/106764 WO2023001190A1 (zh) 2021-07-23 2022-07-20 结直肠息肉图像的识别方法、装置及存储介质

Country Status (3)

Country Link
EP (1) EP4350617A1 (zh)
CN (1) CN113284146B (zh)
WO (1) WO2023001190A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116580027A (zh) * 2023-07-12 2023-08-11 中国科学技术大学 一种结直肠内窥镜视频的实时息肉检测***及方法
CN117036714A (zh) * 2023-10-09 2023-11-10 安徽大学 融合混合注意力机制的肠息肉分割方法、***及介质
CN117058467A (zh) * 2023-10-10 2023-11-14 湖北大学 一种胃肠道病变类型识别方法及***
CN117830226A (zh) * 2023-12-05 2024-04-05 广州恒沙云科技有限公司 一种基于边界约束的息肉分割方法及***

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113284146B (zh) * 2021-07-23 2021-10-22 天津御锦人工智能医疗科技有限公司 结直肠息肉图像的识别方法、装置及存储介质
CN116030303B (zh) * 2023-01-09 2024-01-30 深圳市大数据研究院 一种基于半监督孪生网络的视频结直肠病变分型方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109447973A (zh) * 2018-10-31 2019-03-08 腾讯科技（深圳）有限公司 一种结肠息肉图像的处理方法和装置及***
WO2020222985A1 (en) * 2019-04-30 2020-11-05 The Trustees Of Dartmouth College System and method for attention-based classification of high-resolution microscopy images
CN112200773A (zh) * 2020-09-17 2021-01-08 苏州慧维智能医疗科技有限公司 一种基于空洞卷积的编码器和解码器的大肠息肉检测方法
CN113284146A (zh) * 2021-07-23 2021-08-20 天津御锦人工智能医疗科技有限公司 结直肠息肉图像的识别方法、装置及存储介质

Family Cites Families (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104952089B (zh) * 2014-03-26 2019-02-15 腾讯科技（深圳）有限公司 一种图像处理方法及***
CN105354852A (zh) * 2015-11-27 2016-02-24 浙江工业大学 基于视觉注意机制的内窥图像中病灶定位方法
CN106097279B (zh) * 2016-06-24 2019-01-01 电子科技大学 一种高动态图像增强方法
CN109635866B (zh) * 2018-12-10 2021-07-23 杭州帝视科技有限公司 处理肠图像的方法
CN111839428A (zh) * 2019-04-25 2020-10-30 天津御锦人工智能医疗科技有限公司 一种基于深度学习提高结肠镜腺瘤性息肉检出率的方法
CN111383214B (zh) * 2020-03-10 2021-02-19 长沙慧维智能医疗科技有限公司 实时内窥镜肠镜息肉检测***
CN111784628B (zh) * 2020-05-11 2024-03-29 北京工业大学 基于有效学习的端到端的结直肠息肉图像分割方法
CN111986204B (zh) * 2020-07-23 2023-06-16 中山大学 一种息肉分割方法、装置及存储介质
CN111882548A (zh) * 2020-07-31 2020-11-03 北京小白世纪网络科技有限公司 基于深度学习的病理图像中细胞计数方法及装置
CN112131959B (zh) * 2020-08-28 2024-03-22 浙江工业大学 一种基于多尺度特征强化的2d人体姿态估计方法
CN112418253B (zh) * 2020-12-18 2021-08-24 哈尔滨市科佳通用机电股份有限公司 基于深度学习的撒砂管松脱故障图像识别方法及***
CN112836589A (zh) * 2021-01-13 2021-05-25 苏州元启创人工智能科技有限公司 基于特征融合的视频中人脸表情识别方法
CN113114599B (zh) * 2021-03-12 2022-01-11 电子科技大学 基于轻量级神经网络的调制识别方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109447973A (zh) * 2018-10-31 2019-03-08 腾讯科技（深圳）有限公司 一种结肠息肉图像的处理方法和装置及***
WO2020222985A1 (en) * 2019-04-30 2020-11-05 The Trustees Of Dartmouth College System and method for attention-based classification of high-resolution microscopy images
CN112200773A (zh) * 2020-09-17 2021-01-08 苏州慧维智能医疗科技有限公司 一种基于空洞卷积的编码器和解码器的大肠息肉检测方法
CN113284146A (zh) * 2021-07-23 2021-08-20 天津御锦人工智能医疗科技有限公司 结直肠息肉图像的识别方法、装置及存储介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
HANG ZHANG; CHONGRUO WU; ZHONGYUE ZHANG; YI ZHU; ZHI ZHANG; HAIBIN LIN; YUE SUN; TONG HE; JONAS MUELLER; R. MANMATHA; MU LI; ALEXA: "ResNeSt: Split-Attention Networks", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 19 April 2020 (2020-04-19), 201 Olin Library Cornell University Ithaca, NY 14853 , XP081648111 *

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116580027A (zh) * 2023-07-12 2023-08-11 中国科学技术大学 一种结直肠内窥镜视频的实时息肉检测***及方法
CN116580027B (zh) * 2023-07-12 2023-11-28 中国科学技术大学 一种结直肠内窥镜视频的实时息肉检测***及方法
CN117036714A (zh) * 2023-10-09 2023-11-10 安徽大学 融合混合注意力机制的肠息肉分割方法、***及介质
CN117036714B (zh) * 2023-10-09 2024-02-09 安徽大学 融合混合注意力机制的肠息肉分割方法、***及介质
CN117058467A (zh) * 2023-10-10 2023-11-14 湖北大学 一种胃肠道病变类型识别方法及***
CN117058467B (zh) * 2023-10-10 2023-12-22 湖北大学 一种胃肠道病变类型识别方法及***
CN117830226A (zh) * 2023-12-05 2024-04-05 广州恒沙云科技有限公司 一种基于边界约束的息肉分割方法及***

Also Published As

Publication number Publication date
EP4350617A1 (en) 2024-04-10
CN113284146A (zh) 2021-08-20
CN113284146B (zh) 2021-10-22

Similar Documents

Publication Publication Date Title
WO2023001190A1 (zh) 结直肠息肉图像的识别方法、装置及存储介质
US10957043B2 (en) AI systems for detecting and sizing lesions
Roth et al. A new 2.5 D representation for lymph node detection using random sets of deep convolutional neural network observations
US20210042920A1 (en) Method, apparatus, system, and storage medium for recognizing medical image
CN110600122A (zh) 一种消化道影像的处理方法、装置、以及医疗***
WO2021159811A1 (zh) 青光眼辅助诊断装置、方法及存储介质
CN113989301A (zh) 一种融合多种注意力机制神经网络的结直肠息肉分割方法
CN113781489B (zh) 一种息肉影像语义分割方法及装置
CN110974179A (zh) 一种基于深度学习的电子染色内镜下胃早癌的辅助诊断***
CN115063592B (zh) 一种基于多尺度的全扫描病理特征融合提取方法及***
Yoo et al. Deep learning diagnostics for bladder tumor identification and grade prediction using RGB method
Guleria et al. Deep learning systems detect dysplasia with human-like accuracy using histopathology and probe-based confocal laser endomicroscopy
Shah et al. Colorectal cancer segmentation using atrous convolution and residual enhanced unet
Du et al. Improving the classification performance of esophageal disease on small dataset by semi-supervised efficient contrastive learning
CN116152650A (zh) 基于CNN和Transformer双向协同指导网络的海洋生物检测方法
Tsaku et al. Texture-based deep learning for effective histopathological cancer image classification
Handa et al. Automatic detection of colorectal polyps with mixed convolutions and its occlusion testing
Henriksen et al. Performance of data enhancements and training optimization for neural network: A polyp detection case study
CN116206105A (zh) 一种深度学习相融合的协同学习增强的结肠息肉分割方法
Xue et al. Cleaning highly unbalanced multisource image dataset for quality control in cervical precancer screening
CN111798426A (zh) 用于移动端的胃肠道间质瘤中核***象深度学习检测***
Chong et al. A deep learning network based on multi-scale and attention for the diagnosis of chronic atrophic gastritis
Fan et al. Warp‐based edge feature reinforcement network for medical image segmentation
Li et al. Feature library: A benchmark for cervical lesion segmentation
CN117058467B (zh) 一种胃肠道病变类型识别方法及***

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22845355

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2023580724

Country of ref document: JP

Kind code of ref document: A

WWE Wipo information: entry into national phase

Ref document number: 2022845355

Country of ref document: EP

ENP Entry into the national phase

Ref document number: 2022845355

Country of ref document: EP

Effective date: 20240105

NENP Non-entry into the national phase

Ref country code: DE