NL2028271B1 - Method of performing radiomics analysis on image data. - Google Patents
Method of performing radiomics analysis on image data. Download PDFInfo
- Publication number
- NL2028271B1 NL2028271B1 NL2028271A NL2028271A NL2028271B1 NL 2028271 B1 NL2028271 B1 NL 2028271B1 NL 2028271 A NL2028271 A NL 2028271A NL 2028271 A NL2028271 A NL 2028271A NL 2028271 B1 NL2028271 B1 NL 2028271B1
- Authority
- NL
- Netherlands
- Prior art keywords
- data
- image
- image data
- model
- radiomics
- Prior art date
Links
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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
-
- 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
- G06T2207/10116—X-ray 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
- G06T2207/10—Image acquisition modality
- G06T2207/10132—Ultrasound 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
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- 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
-
- 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
- G06T2207/30108—Industrial image inspection
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Measuring And Recording Apparatus For Diagnosis (AREA)
Abstract
The invention is directed at a method of performing radiomics analysis on imaging data from an imaging system. A processing unit receives image data of a part of a body, including a region of interest. For each of a plurality of image features, an image feature parameter value is derived by the processing unit, the value being a quantitative representation of the image feature. Using a signature model, signature model values are derived from image feature parameter values, for performing radiomics analysis. The imaging system comprises a sensor array With sensors for providing sensor signals, and receiving image data comprises retrieving sensor signal data of sensor signals directly from the sensor. The image data is raW data formed by said sensor signal data for performing the radiomics analysis directly on said raW image data, for enabling to provide at least one classification result associated With the region of interest. [Fig. 1]
Description
Title: Method of performing radiomics analysis on image data. Field of the invention The present invention is directed at a method of performing radiomics analysis on imaging data obtained from at least one imaging system, the method comprising the steps of: receiving, by a processing unit from an imaging system, image data of at least a part of a body, including at least a part of a region of interest; deriving, by the processing unit, for each of a plurality of image features associated with the region of interest, an image feature parameter value from the image data, said image feature parameter value being a quantitative representation of said respective image feature, such as to yield a plurality of image feature parameter values; and deriving, by said processing unit using a signature model, one or more signature model values associated with the region of interest from said plurality of image feature parameter values, wherein said signature model includes a functional relation between or characteristic values of said image feature parameter values for deriving said signature model values therefrom, for performing the radiomics analysis for establishing radiomics data based on the one or more signature model values. The invention is further directed at a method of training a machine learning data processing model, to a trained machine learning data processing model and to a system for performing radiomics analysis on imaging data obtained from an imaging system.
Background Radiomics is the quantitative analysis of standard-of-care medical imaging; the information obtained can be applied within clinical decision support systems to create diagnostic, prognostic, and/or predictive models. Radiomics may also be used to evaluate the quality of inert objects or industrial products. Radiomics analysis can be performed by extracting hand-crafted radiomics features or via deep learning algorithms. Radiomics has evolved tremendously in the last decade, becoming a bridge between imaging and precision medicine. Radiomics exploits sophisticated image analysis tools coupled with statistical elaboration to extract the wealth of information hidden inside computed tomography (CT), magnetic resonance (MR),
echography (EC), scintigrahies (SC), Optical Coherence Tomography (OCT) and/or positron emission tomography (PET) scans, routinely performed in the everyday clinical practice. Many efforts have been devoted in recent years to the standardization and validation of radiomics approaches, to demonstrate their usefulness and robustness.
While radiomics has shown promise in the field of personalized medicine, several hurdles are hindering its incorporation in clinical decision support systems, including but not limited to the sensitivity of the majority of deep or handcrafted radiomic features (RFs) to differences in acquisition and reconstruction. Regardless of whether or not imaging systems are of a same type or even from a same manufacturer, in reality the output of every imaging system is different. Upon installation and in use, the settings of each imaging system are individually managed dependent on many factors. The same is true for the settings applied in reconstruction software, to provide the eventual images. Furthermore, settings may also be adapted dependent on the patient, e.g. dependent on which part of the body needs to be imaged or dependent on body characteristics of a patient. A broad implementation of radiomics is hindered thereby, because it affects the robustness and reproducibility of certain radiomics features.
A radiomics signature is a set of one or more functional relations between or characteristic values of certain image features, which if they are encountered in a combination and proportion that satisfies the signature definition, is of medical significance. For example, if from an image of a tumor or a certain disorder, it can be determined that the image features satisfy a certain radiomics signature, this may provide information about the diagnosis, the biology, the prognosis, the optimal treatment or response evaluation to a certain treatment. As may be appreciated, the most valuable radiomics signatures are those that are sufficiently robust to be independent on the image system or reconstruction settings, because these can be applied to any image regardless of it’s origin. Radiomics can also be used to detect and/or segment an organ, a pathology, or a lesion, typically with deep learning-based methods.
Summary of the invention It is an object of the present invention to overcome the above disadvantages and to provide a method of performing radiomics analysis that can be applied ubiquitously to images of any imaging system.
To this end, there is provided herewith a method of performing radiomics analysis as described above, wherein the imaging system comprises a sensor array including a plurality of sensors configured for providing sensor signals for enabling imaging, and wherein the step of receiving image data comprises retrieving, from the imaging system, sensor signal data representative of the sensor signals directly obtained from the sensor, wherein the image data is raw image data formed by said sensor signal data for performing the radiomics analysis directly on said raw image data.
Instead of applying radiomics analysis on reconstructed and processed images from the imaging systems, the present invention uses the raw data coming from the sensors of the imaging systems in order to perform the radiomics signature analysis. The invention is based on the insight that it is not necessary for radiomics analysis to be based on images that are optimized for visualization to the medical practitioner. The raw data coming from the imaging systems contains all information needed to perform radiomics signature analysis. It is the biologics information that needs to be evaluated in the radiomics analysis to evaluate the prognostic value thereof. All the applied system settings, as well as the reconstruction process thereafter, and the optimization settings that aid to the interpretation of an image by a medical practitioner, are not needed for the radiomies analysis and merely add to the biasing thereof.
Typically, however, raw data from the imaging systems is not used nor stored for medical or industrial applications. Occasionally, in a research setting, such raw data is maintained for research purposes in order to reconstruct images again (e.g. to test new reconstruction algorithms). Therefore, in accordance with some embodiments, the method proposes to maintain raw data from imaging systems in order to enable the radiomics analysis therefore. Those image data are then future- compliant, for example if in the future a new reconstruction methods using new technologies is found those raw image data can be resused. Such data may for example be stored in a data repository, such as a central database or a database located in a medical or industrial facility or federated databases. Furthermore, it is possible to store raw data from imaging systems in a large data repository wherein such data from several medical facilities is stored and evaluated. This, for example, enables to identify radiomics signatures that can be used across medical centers, and are independent of whichever image system is used.
As will be appreciated, differences exist to some extent in the type of imaging system, in particular the modality thereof. For example, the raw data of a computed tomography (CT) system or of a positron emission tomography (PET) system cannot be directly compared to the raw data of a magnetic resonance imaging (MRI) system. Although radiomics analysis applied to one of these modalities may not be applied directly to other modalities, it appears to be possible to use style transfer in order to convert the raw data of an imaging system of a first modality into the raw data style of a second modality. For this, reference is made to e.g. Dutch patent application number NL 2026732 owned by the assignee of the present application.
The term ‘raw image data’, as applied herein, may otherwise be referred to as sensor-level imaging data and is provided by unprocessed imaging data coming from the sensors (e.g. digitized sensor signals forming the sensor signal data). The form and structure of this raw data may typically be dependent on the imaging modality applied. The present method may be applied with a variety of different imaging modalities, including (but not limited to) computed tomography imaging (CT), nuclear magnetic resonance (NMR) or magnetic resonance imaging (MRI), positron emission tomography (PET), echography (e.g. ultrasonic imaging), optical coherence tomography (OCT), and X-ray imaging. These modalities are of different nature and typically acquire and store sensor signal data in a different manner. The idea of using sensor-level data is that this data includes the highest fraction of biological information obtainable, whereas during convention image processing with the objective of generating visible images for a medical practitioner, filter and processing methods are applied that enhance the visibility and interpretability, but which are not directed at optimally maintaining the biological information therein.
In some embodiments, the imaging system is a computed tomography (CT) system, and the sensor signal data is at least one of listmode data or sinogram data. In other or further embodiments, the imaging system is a positron emission tomography (PET) system, and the sensor signal data is listmode data. In yet further embodiments, the imaging system is a magnetic resonance imaging (MRI) system, and the sensor signal data is k-space data. In yet further embodiments, the imaging system is an echography system (ECH), and the sensor signal data is ultrasonic sensor data of one or more ultrasonic bursts having a frequency between 1 and 40 MHz. In yet further embodiments, the imaging system is an optical coherence tomography system (OCT), and the sensor signal data is optical sensor data of near-infrared or 5 infrared optical radiation. In some embodiments, the imaging system is an X-ray system, and the sensor signal data is data indicative of a pulse of X-ray radiation. Any of these types of systems, and their raw imaging data formats, can be used in combination with the method of the invention.
In some embodiments, the imaging system is a computed tomography (CT) system, and wherein for obtaining the image data, X-ray type electromagnetic radiation is applied by the imaging system, wherein the X-ray type electromagnetic radiation is applied at a radiation dose in a range smaller than 2 millisievert, preferably smaller than 1 millisievert. These embodiments are based on the further insight that, in order to capture the biological information desired for performing radiomics signature analysis, it is sufficient in a CT scanner as imaging system to apply a low dose or even ultralow dose.
By ‘low dose’ a dose reduction of 80% or higher as compared to a regular applied dose for performing CT is meant, thus a dose of at most 20% of the normal dose. By ‘ultralow dose’ a dose reduction of 90% or higher as compared to a regular applied dose for performing CT is meant, thus a dose of at most 10% of the normal dose. With some embodiments of the invention, it has been found that a mean effective radiation dose of ultralow dose CT (ULDCT) of 0.12 mSv is achievable, which is a mean reduction of 94% compared to the standard dose. The applied dose qualifying as ‘normal dose’ is dependent on which part of the body is to be imaged. For example, for common types of CT scans, the amount of radiation normally absorbed by the body (i.e. the normal dose) from these systems, are as follows. For image of belly or pelvis, typically 10 mSv (millisievert), which is equal to about 3 years of background radiation. For colonography, typically 6 mSv, which is equal to about 2 years of background radiation. For image of the head, typically 2 mSv, which is equal to about 8 months of background radiation. For image of the spine, typically 6 mSv, equal to about 2 years of background radiation. For image of the chest, typically 7 mSv, equal to about 2 years of background radiation. For lung cancer screening, typically 1.5 mSv, equal to about 6 months of background radiation. For coronary angiography (CTA), typically 12 mSv, equal to about 4 years of background radiation. For cardiac images
(for calcium scoring), typically 3 mSv, equal to about 1 year of background radiation. As may be understood, the low dose and ultralow dose regimes for these types of CT in the present invention can be scaled to the above normal dose regimes. Therefore, the low dose regime typically does not exceed 2.5 mSv — preferably 2 mSv — and the ultralow dose regime does not exceed 1.3 mSv — preferably 1 mSv. Dependent on the part of the body to be imaged, the dose may then be determined: e.g. for the head the normal dose is 2 mSv, and thus the low dose regime is at most 0.4 mSv whereas the ultralow dose regime is at most 0.2 mSv. As another example, for coronary angiography (CTA) the normal dose is 12 mSv, which can be reduced in the low dose regime to at most 2.4 mSv and in the ultralow dose regime to at most 1.2 mSv.
In some embodiments, in addition to performing the radiomics analysis directly on said raw image data, the method further comprises a step of processing the sensor signal data such as to obtain processed image data. Herewith, the step of processing includes applying at least one image processing kernel to the raw image data. Within the scope of the invention, in addition to or alternative to using raw imaging data, processing the image data using an image processing kernel may be applied. In more basic embodiments, such image processing may be targeted at additionally providing an image optimized for visibility and interpretability, i.e. as the practitioner may be used to receive. The radiomics performed on the raw image data, in that case, is performed in addition to the image processing to produce such an image. However, in addition or alternatively, also other kernels may be applied. For example, the method may apply image processing kernels optimized for maintaining an as large as possible part of the biological information, or image processing kernels that enhance certain features of the raw image data that are important in the radiomics analysis e.g. to obtain a more outspoken result thereof. Hence, in these embodiments, the at least one image processing kernel is optimized for enhancing, in the processed image data, one or more of said image feature parameter values used for deriving said signature model values therefrom.
Also, in some embodiments, a plurality of different kernels may be applied, including both kernels applied to provide a regular image as well as kernels applied for other purposes as referred to above. For example, in such embodiments, the step of processing the sensor signal data comprises: applying a plurality of mutually different kernels to the raw image data such as to obtain a plurality of processed images, wherein each processed image is obtained using one of the kernels;
wherein the steps of deriving the plurality of image feature parameter values and deriving one or more signature model values are further performed on each processed image of the plurality of processed images, such as to perform a radiomics analysis on each of the plurality of processed images for obtaining, from each radiomics signature analysis, the at least one classification result associated with the region of interest such as to yield a plurality of classification results. The performance of multiple radiomics analysis on processed image data which is processed in accordance with multiple different kernels, may be applied to the benefit of various purposes. For example, it enables to optimize an image differently by applying different kernels, such as to apply different signature models during the radiomics analysis , which may provide results being indicative of multiple different health states (e.g. evaluating tumor hypoxia, tumor heterogeneity, and the occurrence of metastases in parallel). Thereby, each kernel may optimize the raw image data with respect to the particular signature model to be applied per health state to be evaluated. Additionally or alternatively, this may be applied to gain better insight in e.g. a single clinical question and improve the confidence level of the radiomies result.
Regarding the latter of the above examples, in some embodiment, the method further comprises a step of evaluating the plurality of classification results obtained from the plurality radiomics analysis such as to obtain an overall classification result associated with the region of interest. For example, ten different radiomics analysis may be applied to gain insight in a single clinical question, and the results of all radiomics analyses may be combined (e.g. averaged, ensembled, selected, or statistically evaluated) to provide an overall classification result based thereon.
In some embodiments, the overall classification result is obtained by at least one of: determining whether one or more classification results of the plurality of classification results are identical or similar; counting, for each classification result, a number of identical or similar classification results, and determining the overall classification result to be equal to the majority of identical or similar classification results; or associating a weighing value to each of the mutually different kernels, and calculating, for each classification result, a sum of weighing values associated with those kernels that yielded identical or similar classification results, for determining the overall classification result to be equal to the classification result providing the highest sum of weighing values. In the last embodiment, using weighing values, the classification result of each radiomics analysis is first weighed, and thereafter a decision step is applied based thereon to determine the overall classification. The weighing, however, may likewise be applied in an averaging step to determine an overall classification result. This may depend on the situation.
In some embodiments, the method further comprises a step of calculating a confidence interval for the overall classification result, wherein the confidence interval is calculated based on the plurality of classification results obtained from the plurality radiomics analysis . If all analyses point in a same direction, this obviously raises confidence in the outcome and provides an overall result leaving no or very little room for a different clinical outcome. If the results are highly variable, this in itself is indicative of an overall classification result having a lower confidence level.
As mentioned already above, in some embodiments wherein an image processing kernel is applied, the step of processing comprises a step of image reconstruction for providing based on the image data an enhanced image suitable for visual evaluation. This may be an image comparable to a conventional image, but in some preferred embodiments, the image may be augmented with additional information obtained from the radiomics analyses. For example, this processing may further include image reconstruction, and/or in accordance with some embodiments, a step of contour recognition performed on the processed image data, for establishing contour data indicative of a contour of the neoplasm or the lesion. The processed images, which can be used for regular medical purposes to be reviewed by a practitioner, may alternatively or additionally be stored along with raw image data such as to be useable as training data for training a machine learning data processing model. This will be discussed further down below.
As referred to above, the images may be augmented. Therefore in other or further embodiments, the method further comprises a step of adding one or more information items or information labels to the enhanced image, wherein the information items or information labels include at least one of a group comprising: semantic labelling associated with an ontology, a data model or a machine readable vocabulary, and indicative of the at least one classification of the region of interest; or machine-readable non imaging data such as clinical data, biological data, familial data, or therapeutic data. Machine-readable non imaging data may for example include data such as: clinical data, for example age, body mass index (BMI), history, activities, professional exposures; biological data, e.g. genomic, proteomic or blood biomarkers; familial data, for example a familial history of cancer, heart diseases or diabetes; therapeutic data, for example data about past and/or present treatments. In some embodiments, the image data from the imaging system is stored in a data repository, and the step of receiving the image data includes a step of obtaining the image data from the data repository. Here, the image data is not directly acquired from an imaging system, but may first have been stored in the data repository. It is also possible that the imaging data is received via a data communication network, either directly from an imaging system or from a (central) database.
In some embodiments, the image data includes further processed image data which has been processed prior to storing thereof in the data repository, and wherein the step of receiving the image data further comprises a step of reverse processing of the further processed image data such as to obtain pseudo raw image data, the pseudo raw image data providing mimicked sensor signal data indicative of an original sensor signal data associated with the further processed image; wherein the raw image data is provided by the pseudo raw image data. In this special case, image data which has already been processed (e.g. conventionally processed image data) may be reverse processed to get back the “raw” data. It will be understood that the real raw data from these images may not be obtainable, but the method proposes to generate pseudo raw image data, i.e. image data mimicking the original sensor-level data as good as possible, by performing the reverse processing e.g. using a particular kernel therefore. This pseudo raw image data can then be used as if it were the original raw image data. For example, it can be applied as training data for training a machine learning data processing model to identify and define signature models. The term machine learning data processing model is intended to include a model based on artificial intelligence in the broad sense including machine learning and deep learning. In this process, data annotations (delineations, masks) can also be back projected to enrich the pseudo-raw data. This may later be transferred to real raw data.
In some embodiments, the signature model is obtained by obtaining signature model data indicative of the functional relation between or the characteristic values of said image feature parameter values for deriving said signature model values, and in accordance with various different and/or complementary embodiments, the signature model data may be obtained by at least one of: obtaining, from a user, user input including the signature model data; obtaining the signature model data from a server or data repository; or obtaining the signature model data from a machine learning data processing model. The signature model may be obtained by user input or from a data repository/server, e.g. when using existing signature models. However, in accordance with the above classes of embodiments, another possibility is to obtain the signature model data from a first machine learning data processing model. These embodiments thus rely on deep learning methods to identify radiomics signatures from image data collections. In combination with the storing of raw image data as briefly described above, the deep learning methods can be used to identify robust radiomics signatures that are based on raw image data, the advantages of which have been described above.
In some embodiments, the method further comprises a step of storing, in one or more data repositories, training data for enabling training a machine learning data processing model using a central or distributed training method, wherein the training data comprises the raw image data together with at least one of: the one or more signature model values; the radiomics data; the at least one classification result; or where relevant any one or more of: the processed image data; at least one of the plurality of processed images or the classification results; the overall classification result; the contour data; at least a part of the one or more information items or information labels; or the signature model data. Storing this data in a training data repository enables training of further machine learning data processing models, such as the first or second machine learning data processing model described herein.
In yet further embodiments, e.g. wherein step of processing includes applying at least one image processing kernel or a plurality of processing kernels to the raw image data, the method one or more or all of these kernels may be obtained from a second machine learning data processing model. Such a second machine learning data processing model can be trained to define kernels that optimize the images for a certain purpose, e.g. by enhancing certain data features or by maintaining as much as possible biological data. Such training methods will also be described below. For example in some embodiments, the or each kernel obtained from the second machine learning data processing model is configured for retaining or enhancing one or more data features of the sensor signal data in the processed image data, wherein the one or more data features enhanced or retained comprise at least one element of a group comprising: features associated with high frequency parts of a power spectrum of the sensor signal data; features associated with image data not visible by the human eye; features indicative of tissue hypoxia such as tumor hypoxia; features indicative of heterogeneity; features indicative of blood flow or perfusion of tissue or ventilation of a lung; genomic features; epigenetic features; proteomic features; features related to the concentration of therapeutics, such as concentrations or accumulations of: antibodies, single chain antibodies, single-domain antibodies, chemicals, peptides, bacteria, viruses, metal fractions, contrast product, boron for boron neuron therapy, radioisotopes of radio-immunotherapy (iodine-131 (**'), yttrium-90 (*°Y), lutetium-177 (Lu), Bismuth-213 (Bi). astatine-211 (?!'At), actinium-225 (***Ac)...), chimeric antigen receptor T cells; features indicative of a presence of immune cells, such as CD8+ cells, regulatory T cells, macrophage type I or II, neutrophils, or erythropoietic cells; features related to treatment associated effects, such as DNA damage, reoxygenation, heterogeneity decrease, senescent cells, fibrotic content, or content in inflammatory cells; features related to chemical compositions, such as iron content in hemochromatosis, heta-amyloid deposition in Alzheimer disease, deposition of asbestos, physiological aging score.
In other or further embodiments, the signature model data is obtained from the first machine learning data processing model, and wherein the first machine learning data processing model is trained to provide the signature model data based on processed image data, wherein the processed image data is obtained using an image processing kernel obtained from the second data processing model, such that the signature model is optimized for deriving signature model values based on the processed image data. In this embodiments, the two machine learning data processing models cooperate to process the data optimally in order to apply a radiomies analysis based on an deep learning acquired signature model.
In some embodiments, the method further comprising a step of: identifying, in said image data, at least one data section containing data relating to a region of interest. By doing so, this enables the method to be concentrated on the region of interest only, and thereby reduce the required arithmetic capacity for carrying out the analysis. In some of these embodiments, where storing of data is performed as earlier described, the step of storing may comprise the storing of the data of the at least one data section relating to the region of interest, while discarding image data that is not included in the at least one data section. This enables very efficient storing of raw image data, which is a great advantage in view of the typical size of sensor level image data for a complete image.
In some embodiments, the radiomics analysis is performed for a medical use, such as for one or more of: screening, diagnosis, biology, prognosis, treatment selection, or response evaluation to a treatment. However, in other or further embodiments, the radiomics analysis is performed for an industrial use, such as for one or more of: quality testing of an object, age evaluation, determination of a maturity of an organic produet such as a plant or a fruit; or determination of a disease status of an organic product such as a plant or a fruit.
In accordance with a second aspect, there is provided a method of training a first machine learning data processing model using a method according to any one or more of the preceding claims, wherein the method comprises: providing, to the machine learning data processing model from an imaging system, image data for a plurality of images, each of the images being of at least a part of a human or animal body comprising at least a part of a neoplasm or a lesion, and wherein the image data is raw image data formed by sensor signal data obtained from a sensor array of the imaging system, the sensor array including a plurality of sensors configured for providing sensor signals for enabling imaging; providing, to the machine learning data processing model by a processing unit, for each image at least one of: one or more derived signature model values or derived radiomics data, wherein the derived signature model values or the derived radiomics data are obtained using a method according to any one or more of the preceding claims; and training the machine learning data processing model based on the image data and the at least one of: the one or more derived signature model values or the derived radiomics data, for enabling the machine learning data processing model to provide signature model data indicative of a signature model for classifying the neoplasms or lesions imaged in the plurality of images, based on the image data as input. Using the above training method enables to train machine learning data processing models to perform automatic determination of e.g. new signature models. The systems is thereby fed with raw image data and with radiomics data or e.g. medical data, and may analyze which image feature characteristies are typically encountered with certain medical states, neoplasm phenotypes, or disorder states or characteristics, to give some examples.
In some embodiments, one or more of: the image data; the one or more signature model values; the radiation data; or further training data is obtained from training data stored in a data repository in accordance with an embodiment earlier described.
In some embodiments, the training method is a reinforced learning method, further comprising the steps of: generating, using the machine learning data processing model and based on a raw image data as input, trial signature model data indicative of a trial signature model; deriving, by a processing unit, for each of a plurality of image features associated with the region of interest, an image feature parameter value from the image data, said image feature parameter value being a quantitative representation of said respective image feature, such as to yield a plurality of image feature parameter values; and deriving, by said processing unit using the trial signature model, one or more trial signature model values associated with the region of interest from said plurality of image feature parameter values, for performing a radiomics analysis for establishing radiomics data based on the one or more trial signature model values; comparing, by the processing unit, the one or more trial signature model values with reference data, wherein the reference data comprises at least one of: the one or more derived signature model values; and generating, based on the step of comparing, a positive feedback value or a negative feedback value dependent on whether or not the one or more trial signature model values matches the one or more derived signature model values. A reinforced learning model has been found to be well suited for identifying signature models useable in a method of the present invention.
In accordance with a third aspect, there is provided a method of training a second machine learning data processing model, wherein the method comprises: providing, to the second machine learning data processing model from an imaging system, image data for a plurality of images, each of the images being of at least a part of a human or animal body comprising at least a part of a neoplasm or a lesion, and wherein the image data is raw image data formed by sensor signal data obtained from a sensor array of the imaging system, the sensor array including a plurality of sensors configured for providing sensor signals for enabling imaging; providing, to the second machine learning data processing model by a processing unit, a trial image processing kernel; providing, to the second machine learning data processing method, one or more data feature criteria; performing an initial processing step wherein the image data is processed using the image processing kernel such as to yield a processed image data and determining a quality parameter indicative of how well the processed image matches the data feature criteria; wherein the method further comprises iterating of the steps of: a) modifying the trial image processing kernel; b) processing the processed image data using the trial image processing kernel; ¢) comparing the processed image data against the data feature criteria such as to determine the quality parameter; and d) determine whether the quality parameter is indicative of an improvement, and based thereon, determine whether to perform a further iteration or whether to cease further iteration; wherein the method, after ceasing the iteration, comprises providing the trial image processing kernel as image processing kernel at the output of the second machine learning data processing model. In accordance with a fourth aspect, there is provided a machine learning data processing model for performing radiomics signature analysis, wherein the machine learning data processing model is trained using a method according to the second aspect such as to yield a first machine learning data processing model; or wherein the machine learning data processing model is trained using a method according to the third aspect such as to yield a second machine learning data processing model.
Yet in accordance with a fifth aspect, there is provided a system for performing radiomics analysis on imaging data obtained from an imaging system, wherein the system is cooperatively connected to the imaging system, and wherein the system is cooperatively connected to, and wherein the system comprises a memory and a processing unit, wherein the processing unit is configured for: receiving, from the imaging system, image data of at least a part of a human or animal body, wherein the part of the human or animal body comprises at least a part of a neoplasm or a lesion; deriving for each of a plurality of image features associated with the region of interest, an image feature parameter value from the image data, said image feature parameter value being a quantitative representation of said respective image feature, such as to yield a plurality of image feature parameter values; and deriving, using a signature model stored in the memory, one or more signature model values associated with the region of interest from said plurality of image feature parameter values, wherein said signature model includes a functional relation between or characteristic values of said image feature parameter values for deriving said signature model values therefrom, for performing the radiomics analysis for establishing radiomics data based on the one or more signature model values; wherein the imaging system comprises a sensor array including a plurality of sensors configured for providing sensor signals for enabling imaging, and wherein for receiving said image data the processing unit is configured for retrieving, from the imaging system, sensor signal data representative of the sensor signals directly obtained from the sensor, wherein the image data is raw image data formed by said sensor signal data for performing the radiomics analysis directly on said raw image data. Brief description of the drawings The invention will further be elucidated by description of some specific embodiments thereof, making reference to the attached drawings. The detailed description provides examples of possible implementations of the invention, but is not to be regarded as describing the only embodiments falling under the scope. The scope of the invention is defined in the claims, and the description is to be regarded as illustrative without being restrictive on the invention. In the drawings: Figure 1 schematically illustrates a method in accordance with the present invention; Figure 2 schematically illustrates a method in accordance with an embodiment of the present invention; Figure 3 schematically illustrates a method in accordance with an embodiment of the present invention; Figure 4 schematically illustrates an embodiment of a method in accordance with the present invention; Figure 5 schematically illustrates a method in accordance with an embodiment of the present invention; Figure 6 schematically illustrates an embodiment of the method in accordance with the present invention; Figure 7 is a continuation of figure 4 illustrating an embodiment of the method of the present invention; Figure 8 schematically illustrates an example analysis carried out using the present invention; Figures 9A and 9B illustrate a reconstructed and raw MRI image; Figures 10A and 10B illustrate a reconstructed and raw CT image.
Detailed description In figure 1, the method 1 enables to perform a radiomics analysis on imaging data obtained from an imaging system 10. The imaging system 10 may for example be a computed tomography (CT) system comprising a table 6 onto which a patient 5 may be situated, which is examined by the scanner 10. The scanner 10 comprises a plurality of sensors in a sensor system, with which the imaging data is obtained. A computed tomography system 10 for example irradiates x ray radiation from an x ray source, which is received by the plurality of sensors within the imaging system 10. Note that the sensors in imaging system 10 are not visible in figure 1. The sensor signals are digitized as sensor signal data by the system 10, and provide the raw data 8 upon which an image 30 may later be based. The method 1 in figure 1 is explained on the basis of a computed tomography system 10 as imaging system 10. It is to be understood, though, that the imaging system 10 may alternatively be a different imaging system, for example a magnetic resonance imaging (MRI) system, a positron emission tomography (PET) system, an ultrasound based system, such as an echography system (ECH), an optical coherence tomography system (OCT), or any other imaging system. The schematic imaging system 10 illustrated in figure 1 typically may be recognized as a CT system, but could be embodied different in case a different imaging system is applied. Typically from a CT system or a PET system, the raw data 8 is obtained as list mode data from the system 10. In case the imaging system 10 would be an MRI system, the raw data at the input of the method 1 would preferably be k-space data. The exact manner wherein the data is stored is not essential to the present invention. It is to be understood that at the input of the method 1 in accordance with the present invention, sensor level data from the imaging system 10 is to be obtained, such as sensor level data 8 in figure 1.
In step 12, the method continues by identifying, from the raw imaging data 8 obtained from the imaging system 10, a plurality of imaging features that are relevant for performing radiomics analysis later. The output of analysis step 12 will be provided as image feature data 13. The image feature data 13 will be the input to step 17 for performing radiomics signature analysis. The radiomics analysis 17 is performed on the basis of a signature model provided by signature model] data 16. The signature model data 16 will for example be obtained from a database 15, but may likewise be obtained from a different source. Alternatively, the signature model data 16 may be manually provided by an operator of the system. In yet another class of embodiments, the signature model data 16 is obtained using a machine learning data processing mall, as will be explained later in this document.
The radiomics analysis step applies signature model 16 to the image feature data 13 in order to determine signature model values 20 therefrom. For example, a signature model may provide algorithms or relations that enable to calculate signature model values 20 based on the image features 13 determined in the analysis step 12. These algorithms may include all types of algorithms such as algorithms based on image features 13 or combinations of image features 13 or statistical data relations. Another possibility is that these algorithms or relations are provided by charaeteristic values of the image feature data 13, or for example threshold values (e.g. whether or not the heterogeneity of the image data or the concentration of a certain drug or pathogen is above a threshold level). The specific signature model 16 that is applied for performing the radiomics analysis 17 may typically be dependent on the clinical question to be answered. For example, the clinical question could be to classify a certain tumor that is found in the patient 5. Another clinical question could be to classify an abnormality or a disorder pattern found in the patient 5. For example, concentrations of certain chemical composition which can be imaged with the imaging system 10 may have a certain clinical relevancy and may therefore be part of a signature model that enables to perform an objective overall clinical assessment of the patient’s 5 health state.
In step 22, the signature model values 20 are evaluated in order to classify a region of interest in the image obtained from the imaging system 10. The step 22 will provide a classification 25 as output of the method 1. The classification 25 may for example be indicative of the presence a certain type of disorder or a lesion, a bone fracture, cicatrices, or a certain type of tissue. For example, the classification 25 may be indicative of a type of tumor or how well it responds to treatment, or may indicate the presence of an infarct or constriction of an artery or vain or e.g. thrombosis.
Optionally, the method 1 may additionally provide a regular type of image reconstruction step 27 based on a image processing terminal 28. The regular image reconstruction step 27 allows to provide an image 30 of the patient 5. Here, in the present example, an image of the patient's lungs 23 is illustrated showing an abnormality 24 in the patient's left lung (the right sight of the image 30). In the example provided, the region of interest could be the part of the image 30 showing the abnormality 24. The abnormality 24 for example could be a tumor or lesion in the lung
23. Method 1 in that case particularly is directed that analyzing the region of interest, being the abnormality 24. Thus the steps 12 and 17 are in particularly focused on obtaining the image features 13 from the region of interest including the abnormality
24 and calculating the signature model values 20 from these image features 13. Alternatively, for example in a situation wherein a patient 5 is scanned for identifying metastases of cancer, the whole image 30 may be region of interest, in a sense that any abnormality found in the image 30 is to be evaluated using a signature model 16 in order to determine whether or not this may be a metastasis. In a different example, an X-ray image may be made of a broken bone, in order to accordingly determine the seriousness of the injury. In principle, the present method may be generally applied in the clinical field. Even outside this field, the present method may be applied for evaluation of imaging data. For example, the imaging data may be an image of a plant or other subject, or an internal material scan of a physical object in order to determine its integrity. For example, it is also possible to apply the present invention in combination with internal imaging of semiconductor structures to identify defects or internal damage. In another field, internal imaging of materials may be evaluated in this manner to detect certain patterns in e.g. an interface between two layers. A method of the present invention may be applied broadly, but provides particular advantages in the clinical field.
Figure 2 schematically illustrates a method in accordance with the further embodiment of the present invention. In the figure 2 the imaging system 10 again includes the table 6 on to which a patient may be situated. While data from the imaging system 10 is first provided, like in figure 1, to an analysis step 12-1 in order to determine image features therefrom. Thereafter, in step 17-1, the radiomics analysis is performed on the image features determined in step 12-1. From the signature model values provided at the output of step 17-1, a classification {el1) 25-1 is obtained from this radiomics analysis step 17-1.
In addition to this first radiomics analysis branch described above, the method in figure 2 further includes a plurality of image processing steps. The image processing steps each apply a kernel 28-2, 28-3 and 28-4 to the imaging data considered. In accordance with the present invention, the kernels 28-2, 28-3 and 28-4 may be applied directly on the raw imaging data 8 obtained from the imaging system
10. The kernels 28-2 through 28-4 are mutually different from each other, and may for example optimize the imaging data for certain radiomics analysis steps to be performed. For example, each of the kernels k1, k2 and k3 may individually enhance one of more image features in the raw imaging data obtained from imaging system 10. The kernels k1 through k3 (28-2 through 28-4) thereby each prepare the raw imaging data for the radiomics analysis that follows, by for example enhancing or augmenting certain imaging features that are important for evaluating the image data in accordance with the respective signature model to be applied. These enhancements or optimizations may focus on any of a large number of different image features that can be directly or statistically or in accordance with another algorithm be determined from the raw data.
The image data processed in accordance with each of the respective image processing kernels 28-2, 28-3 and 28-4 is respectively analyzed in order to determine the imaging features therefrom in steps 12-2, 12-3 and 12-4. Thereafter, in the radiomics analysis steps 17-2 through 17-4, the radiomics analysis is performed in accordance with the signature model 16 associated with each of the kernels k1 through k3, for example. Alternatively of additionally, the radiomics analysis steps 17-2, 17-3 and 17-4 may all perform the same signature model 16, but the difference is merely in the preprocessing of the data in accordance with the kernels k1 through k3. At output of the radiomics signature steps, a plurality of different signature model values 20 will be obtained for each of the radiomics analysis steps 17-2 through 17-4. This will provide a classification 25 for each of these radiomics analysis steps, in figure 2 being classifications cl2 25-2, ¢l3 25-3 and cl4 25-4.
All of the classifications cll, ¢l2, ¢l3 and cl4 (25-1 through 25-4) are provided to an evaluation step 35 in order to evaluate the outcome. The evaluation step 35 will provide an overall classification that is based on the various classifications 25-1 through 25-4 obtained from the performed radiomics signature analysis. For example, the evaluation step may involve a comparison step wherein it is compared whether or not the classifications 25-1 through 25-4 confirm each other, or alternatively are very different from each other. If for example three of the four classifications are similar, or are indicative of a same clinical pattern, the overall classification 38 may be determined to match this clinical pattern. Furthermore, rather than simply considering the majority of classifications, each of the classification results 25-1 through 25-4 may for example include a weighing value to allow some of the classifications 25-1 through 25-4 to be more important in the end result than other classifications. Eventually, an overall classification 38 will be determined in evaluation step 35. In addition, based on this analysis of classification, and optionally based on the weighing values, the confidence level may be determined. As may be appreciated, in case all of the classifications 25-1 through 25-4 point in a same direction, the confidence level in the outcome is very high. However, if one or more of the classifications point in a different direction, this may lower the overall confidence level in the overall classification 38. For example, if only two of the classifications point in the same direction while two other classifications both point in a same different direction, the overall classification 38 may be ambiguous but at least point in these two directions. If two of the classifications point in the same direction while two other classifications each point in a completely different direction, some confidence in the classification designated by the first two classifications that point in the same direction may be gained, although this being two unlimited extend. In the present example, the evaluation step 35 is based on four different classifications 25-1 through 25-4. The skilled person will appreciate that any number of classifications may be applied here, and the invention is not limited to a certain number of classifications. For example in case the skilled person desires to apply one hundred different radiomics signature steps 17, providing one hundred different classifications, the method of the present invention may be applied to provide a very accurate overall classification result 38 which also indicates the confidence level 39. However, the method may also be applied with a different number of radiomics signature steps 17, e.g. only one or two, or maybe ten, or five hundred.
A further embodiment of the present invention is schematically illustrated in figure 3. The left side of figure 3, in particular steps 12-1, 17-1, 25-1, 35, 38, 39 and the steps 28-2 through 28-5, 17-2 through 17-5 and the classifications 25-2 through 25-4, are similar to the method in accordance with the embodiment illustrated in figure 2. Additional to this embodiment is the image reconstruction step 40 which is applied on the basis of kernel 48. This image reconstructions step emphasizes to provide an interpretable image that can be evaluated by a medical practitioner of the patient itself. For example, an optional step of contour recognition 45 may by applied in order to emphasize the contours of an abnormality shown in the image of the patient 5. In the embodiment illustrated in figure 3, from the radiomics analysis steps 17-2 through 17-5, clinical analysis data may be obtained in step 50. For example, if one of the radiomics analysis particularly focus on tumor heterogeneity, then the clinical analysis data may by indicative of those parts of the image step are indicative of a high level of heterogeneity. Alternatively, if some of the radiomics analysis steps 17-2 through 17-5 focusses on the presence of certain biomarkers in the image, the locations where these biomarkers are found may be obtained as clinical analysis data in step 50. The clinical analysis data obtained in step 50 can be obtained directly from the signature model values and/or the image feature parameter values determined in the analysis steps 12 and radiomics analysis steps 17-2 through 17-5. The clinical analysis data may be further enhanced by obtaining other clinical data which is not necessarily related to the image, but which may be related to the patient itself. For example this may be information about his age or his gender, or information about his medical history or genetics. In step 52 the image obtained from steps 40 and 45 is augmented with the clinical analysis data. For example, parts of the images may be colored or textured and labels may be edited to be indicative of certain image features and clinically relevant data. All this information may be combined in step 53 in a single comprehensive image or in a plurality of comprehensive images that can be interpreted by a medical practitioner, the patient or another person. For example, if a plurality of comprehensive images is provided in step 53, each of these comprehensive images may focus on a certain aspect that is relevant for the overall clause 38 determined by the method. To give an example, if a specific type of tumor is classified by the method illustrated in figure 3 such that the overall clause 38 is indicative of this type of tumor, the comprehensive images provided in step 53 may provide information that explains to the medical practitioner or the patient how the system of the present invention came to the conclusion that this type of tumor is to be considered.
Figure 3 further illustrates the database 60. The database 60 indicates that the raw data 8 from the imaging system 10 may be directly provided to the analysis system performing the method of the present invention, or may be obtained from a database 60. For example, the raw data may be stored in the database 60 as indicated by the dotted arrow from the imaging system 10 to the database 60, and may be retrieved from that database 60 by a remote analysis system, as indicated by the dotted arrow between database 60 and the start of the method of the present invention. For example, the database 60 may be a central database wherein raw imaging data of a plurality of imaging systems, for example imaging systems that are remotely located in the plurality of departments of a hospital, or in a plurality of hospitals and medical facilities, and may centrally store this data in database 60 remotely therefrom.
Figure 4 schematically illustrates the further embodiment of the present invention. In principle, figure 4 only shows a part of the method of the present invention wherein the raw data 8 is processed in step 27 (image processing step) using a kernel 28, and thereafter radiomics analysis in step 17 is performed to provide a certain class 25. Thus, the left part of figure 4 may be indicative of any of the branches performing radiomics in each of the figures 1 through 3. What is special about figure 4, ig that the kernel 28 used for image processing in step 27, is obtained using a machine learning data processing model 27 (hereinafter referred to as ‘second machine learning data processing model’). For example, the machine learning data processing model at it’s input may obtain data from a database 66 which is indicative of a certain clinical question to be answered, or a certain signature model to be applied in radiomics signature analysis.
Based thereon, the second machine learning data processing model 72 has been trained in order to provide a kernel 28 that optimizes the raw data 8 for specifically performing radiomies analysis in step 17 using this signature model 16, or a signature model that specifically addresses the clinical question obtained as input from the database 66. It is also possible that kernels 28 are stored in the database 61 from which they are directly retrieved in step 27. The radiomics analysis step 17 uses a signature model 16 which is also obtained from a machine learning processing model 71 (here and after refer to first machine learning data processing model). The first machine learning data processing model 71 main be trained to provide a signature model 16 that is based on a clinical question obtained from database 66. For example, if a image is to be analyzed in order to address a certain clinical question, then bases of the information obtained from database 66, the correct signature model can be identified or provided using the first machine learning data processing model 71 trained to perform this task.
In the method illustrated in figure 4, the kernel 28 and the signature model 16 are optimized for each other, such as to perform the radiomics analysis step 17 and provide classification 25. As may be appreciated in different embodiments either the first machine learning data processing model 71 or the second machine learning data processing model 72 may be absent, such that only the signature model 16 is provided by a first machine learning processing model or that only the kernel 28 is provided by a second machine learning data processing model 72. Also, in addition or alternative to the clinical data or clinical question data, other information may be provided to the machine learning data processing models 71 and 72, such as the image modality of the imaging system used (1.2. a CT-scanner of a MRI-scanner for example).
Figure 5 schematically illustrates an embodiment of a method in accordance with the present invention which integrates a plurality of analysis methods in order to provide a training database 66 on the basis of which the first machine learning data processing model 71 of the second machine learning data processing model 72 can be trained.
In the embodiment illustrated in figure 5, a plurality of imaging systems 10-1, 10-2 and 10-3 provide imaging data which is analyzed in methods 1,1" and 17 in a manner that has been described herein before.
The output of this analysis, for example including imaging data (raw imaging data or processed imaging data), classification results, signature models applied, kernels applied, specifics of the case (clinical data for example) or other data obtained in the methods 1,1" and 1”. All this data may be stored in the database 66 such as to enable use thereof during a training method 70. In addition to the above, image data from a database 60 which has been processed using a conventional image processing scene to provide medical images that are to be evaluated by a medical practitioner, may be reversely processed in step 65 to provide pseudo raw image data 68. The pseudo raw image data 68 mimics the original raw imaging data from the imaging system with which it is obtained.
As may be appreciated, the described disadvantage above was that conventionally processed imaging data is sub-optimal because a part of the biological information is lost.
This biological information cannot be regained, however by reversely processing the data in step 65 to obtain pseudo raw data 68, it is possible to use this pseudo raw imaging data 68 in an analysis method 17 in order to use it as training data for training for example a first machine learning data processing model 71 to identify a radiomies signature.
As may be appreciated, the clinical images obtained from the database 60 and reversely processed in step 65 may already have been evaluated by a medical practitioner and a type of tumor of the clinical specifics of the case are known.
Therefore, this information may well be used in order to train the machine learning processing model.
The training method illustrated in figure 7 may for example be a reinforced learning training method for obtaining a machine learning data processing model.
For example, a plurality of random images from database 60 may be loaded into for example first machine learning data processing model 71, and evaluated according to the model 71 to provide radiomies signature model 16. If the first machine learning data processing model 71 provides a radiomics signature model 16 that is sub-optimal for assessing the image data at the input 73, a negative feedback value is generated and stored.
If a correct signature is predicted by the machine learning data processing model 71, a positive feedback value is determined and stored. The first machine learning data processing model 71 thereby learns to identify the correct radiomics signature 16 with the image data it receives in 73. Similarly, for the processing kernel 28 this training method may be applied for optimizing the imaging data to perform a certain analysis. Alternatively, another possibility is that the training method 70 start with determining a trial kernel 28 and modifies the kernel parameters while evaluating the processing result to determine whether the correct image features are emphasized or enhanced by the kernel 28. In this way, the second machine learning data processing model 72 can be trained to provide the correct kernel 28 on the basis of the image data 74 obtained from database
60.
Figure 6 illustrates a method in accordance with the present invention, carrying out radiomics analysis on raw image data. In accordance with the present invention, raw image data is received in step 120 from an imaging system. The raw imaging data, which may be sinogram or k-space data for example, comprises data obtained by scanning a part of a body of a patient, for example a patient's thorax in the example of figure 6. The data enables to visualize the inside of the thorax, such as to aid diagnosis and treatment of a medical condition.
Thereafter, in the method in accordance with the present invention, without the particular need for pre-processing of the raw data, a segmentation step 124 is performed on the raw imaging data. The segmentation step 124 slices the raw data. Segmentation step 124 may slice the data in different ways, for example in three different directions. Although presently the method works on the basis of raw imaging data (which is not reconstructed as suggested in figure 6), figure 6 for the interpretation of the figure includes a reconstructed lung to illustrate how segmentation is performed. In the example of figure 6, the segmentation step 124 includes a step 140 which divides the initial volume of the raw imaging data obtained from step 120 into three different arrays, wherein each of the arrays provides a three- dimensional arrangement of pixel values indicative of the image data. Each of the three-dimensional arrays comprise consecutive slices of two-dimensional data. For example, the segmentation step 124 may comprise slicing of the data in three directions, in the axial, coronal, and sagittal direction. This is indicated by substeps 142, 144 and 146. From each of these slicing steps, the slices are combined to form a single three-dimensional array for each slicing direction. This yields three different three-dimensional arrays of voxel data sliced by the segmentation step 124. The imaging data 148 in figure 6 is, for example, sliced in the axial direction in substep 142 which yields slices such as slice 150. Slicing in the coronal direction in substep 144 yields slices, such as slice 151; and slicing in the sagittal slicing direction in substep 146 yields slices such as slice 152. In case of forming projection arrays, the method continues with the reconstructions step 126, which is described herein below with reference to figure 7.
In figure 7, the reconstruction step 126 has been schematically illustrated.
During the reconstruction step 126, the data from the different three-dimensional arrays obtained during segmentation in step 124 is reconstructed to provide a single three-dimensional data set. This is performed in substep 160. As visualized in figure 7, data from each of the three-dimensional arrays 161, 162 and 163 may be combined by unification in various ways. For example, a weighted averaging, mean averaging or different algorithm may be applied to calculate each of the values in the three- dimensional dataset to be formed. A convolutional neural network (CNN) can be utilized to find the best way to merge the projection arrays, for example for some tumors, which are better visible from sagittal projection preference, would be given to the values in the sagittal projection array. The output of reconstruction step 126 with substep 160, yields the three-dimensional dataset that may be used for further analysis. After completion of step 126, the output is provided in step 128 of the method. In principle, thereafter it is possible to again perform a contour recognition step in order to, for example, identify the contours of a tumour or other neoplasm that is present in the imaging data. For example, in the example of figures 6 and 7, the lungs of a patient have been examined in order to identify whether or not a tumour is present therein. The contour recognition step may again apply a trained machine learning data processing model in order to identify the contours of such a neoplasm. This is visualized in figure 6 which provides real imaging data indicating the presence of a tumour in the upper part of the left lung of a patient. Optionally, a subset of imaging data containing only the data for the identified neoplasm or non-malignant lesion may be extracted from the total imaging data for further analysis. This may for example yield the subset illustrated in plot 165 in figure 7.
Figure 8 schematically illustrates an example analysis performed using one of the embodiments of the present invention. The figure includes the raw data picture 170 of a brain including a suspicious neoplasm. From this, a reconstructed image 176 may be obtained in a conventional manner.
In accordance with the present invention, radiomics (step 178) will also be performed directly on the raw data of image 170. However, in parallel thereto, in step 172 based on the type of analysis to be preferred, a suitable kernel is selected in order to obtain an Al compliant version 174 of the raw image data.
The Al compliant image is analysed, and based on the combined results of the radiomics analysis 178 on the raw data, and the radiomies analysis performed on the AI compliant image, a probability is determined as to whether the suspicious tumor is a glioblastoma (GBM). From this, an augmented reconstrueted image 180 is obtained, from which information may be obtained about the determined probability (98%) and the data patterns supporting the decision.
Figures 9A and 9B illustrate a reconstructed and raw MRI image of a banana.
Radiomics may be performed in order to perform a quality assessment or to determine whether the banana suffers from a disease.
In figures 10A and 10B a reconstructed and raw CT image of a phone are illustrated that are useable in a diagnostic method to detect a cause of malfunction.
Also, the radiomics method may be applied to perform an inspection of final products.
The present invention has been described in terms of some specific embodiments thereof.
It will be appreciated that the embodiments shown in the drawings and described herein are intended for illustrated purposes only and are not by any manner or means intended to be restrictive on the invention.
The context of the invention discussed here is merely restricted by the scope of the appended claims.
Claims (34)
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
NL2028271A NL2028271B1 (en) | 2021-05-21 | 2021-05-21 | Method of performing radiomics analysis on image data. |
PCT/NL2022/050275 WO2022245214A1 (en) | 2021-05-21 | 2022-05-20 | Method of performing radiomics analysis on image data |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
NL2028271A NL2028271B1 (en) | 2021-05-21 | 2021-05-21 | Method of performing radiomics analysis on image data. |
Publications (1)
Publication Number | Publication Date |
---|---|
NL2028271B1 true NL2028271B1 (en) | 2022-12-06 |
Family
ID=77999270
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
NL2028271A NL2028271B1 (en) | 2021-05-21 | 2021-05-21 | Method of performing radiomics analysis on image data. |
Country Status (2)
Country | Link |
---|---|
NL (1) | NL2028271B1 (en) |
WO (1) | WO2022245214A1 (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2024150791A1 (en) * | 2023-01-11 | 2024-07-18 | 国立大学法人九州大学 | Image conversion device, image conversion method, and image conversion program |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018232388A1 (en) * | 2017-06-16 | 2018-12-20 | Rensselaer Polytechnic Institute | Systems and methods for integrating tomographic image reconstruction and radiomics using neural networks |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
NL2026732B1 (en) | 2020-10-22 | 2022-06-16 | Health Innovation Ventures B V | Method of processing medical images by an analysis system for enabling radiomics signature analysis. |
-
2021
- 2021-05-21 NL NL2028271A patent/NL2028271B1/en active
-
2022
- 2022-05-20 WO PCT/NL2022/050275 patent/WO2022245214A1/en active Application Filing
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018232388A1 (en) * | 2017-06-16 | 2018-12-20 | Rensselaer Polytechnic Institute | Systems and methods for integrating tomographic image reconstruction and radiomics using neural networks |
Non-Patent Citations (1)
Title |
---|
GAO YONGFENG ET AL: "Improved computer-aided detection of pulmonary nodules via deep learning in the sinogram domain", VISUAL COMPUTING FOR INDUSTRY, BIOMEDICINE, AND ART, vol. 2, no. 1, 1 December 2019 (2019-12-01), XP055881841, Retrieved from the Internet <URL:https://vciba.springeropen.com/track/pdf/10.1186/s42492-019-0029-2.pdf> [retrieved on 20220120], DOI: 10.1186/s42492-019-0029-2 * |
Also Published As
Publication number | Publication date |
---|---|
WO2022245214A1 (en) | 2022-11-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Sarica et al. | Random forest algorithm for the classification of neuroimaging data in Alzheimer's disease: a systematic review | |
Bi et al. | Synthesis of positron emission tomography (PET) images via multi-channel generative adversarial networks (GANs) | |
Saba et al. | Lung nodule detection based on ensemble of hand crafted and deep features | |
Zhou et al. | Enhancing magnetic resonance imaging-driven Alzheimer’s disease classification performance using generative adversarial learning | |
JP2022525198A (en) | Deep convolutional neural network for tumor segmentation using positron emission tomography | |
JP6807820B2 (en) | Image search device, method and program | |
CN109472263B (en) | Global and local information combined brain magnetic resonance image segmentation method | |
CN101360453A (en) | Method and system of computer-aided quantitative and qualitative analysis of medical images | |
Sreenivasu et al. | [Retracted] Dense Convolutional Neural Network for Detection of Cancer from CT Images | |
Gates et al. | Glioma segmentation and a simple accurate model for overall survival prediction | |
Fashandi et al. | An investigation of the effect of fat suppression and dimensionality on the accuracy of breast MRI segmentation using U‐nets | |
Apostolopoulos et al. | Deep learning-enhanced nuclear medicine SPECT imaging applied to cardiac studies | |
US10896503B2 (en) | Identification of areas of interest in imaging applications | |
Madhumathy et al. | Deep learning based photo acoustic imaging for non-invasive imaging | |
Mohammed et al. | An efficient approach to diagnose brain tumors through deep CNN | |
NL2028271B1 (en) | Method of performing radiomics analysis on image data. | |
Rai | Cancer detection and segmentation using machine learning and deep learning techniques: A review | |
Özbay et al. | Brain tumor detection with mRMR-based multimodal fusion of deep learning from MR images using Grad-CAM | |
Al-Asadi et al. | Medical image analysis using machine learning techniques: A systematic review | |
US20220375077A1 (en) | Method for generating models to automatically classify medical or veterinary images derived from original images into at least one class of interest | |
Khasawneh et al. | [Retracted] Early Detection of Medical Image Analysis by Using Machine Learning Method | |
Carmo et al. | Automated computed tomography and magnetic resonance imaging segmentation using deep learning: a beginner's guide | |
Lamrani et al. | U-Net-based Artificial Intelligence for Accurate and Robust Brain Tumor Diagnosis using Magnetic Resonance Imaging | |
Ramakrishnan et al. | Automated lung cancer nodule detection | |
Praschl et al. | U-Net based vessel segmentation for murine brains with small micro-magnetic resonance imaging reference datasets |