Pathological image analysis system based on weak supervised learning

Abstract

The invention provides a pathological image analysis system based on weak supervised learning. The pathological image analysis system comprises a model processing module, an information transmission module and a data feedback module, the model processing module performs image processing and intelligent analysis on the digital pathological section by adopting weak supervised learning to obtain an intelligent analysis result of the digital pathological section; the information transmission module encrypts and transmits the digital pathological section intelligent analysis result to the data feedback module; and the data feedback module generates feedback information according to the intelligent analysis result of the digital pathological section, and presents the feedback information to a pathologist. According to the pathological image analysis system based on weak supervised learning provided by the invention, the dependence on the labeled data is reduced through weak supervised learning, the cost consumption is effectively reduced, the system can adapt to the characteristics of different types of digital pathological sections, the limitation of the pathological image analysis system is eliminated, and the application range of the pathological image analysis system is widened.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a pathological image analysis system based on weakly supervised learning. Background Art

[0002] In recent years, medical imaging technology has developed rapidly, and digital pathology section technology has been increasingly used in clinical diagnosis. Traditional pathology diagnosis relies on pathologists to observe and manually analyze digital pathology sections with the help of observation tools, which is not only time-consuming and labor-intensive, but also easily affected by subjective factors, resulting in errors in the analysis results of digital pathology sections, thus failing to guarantee the accuracy and consistency of the diagnosis results. With the continuous advancement of artificial intelligence and deep learning technologies, artificial intelligence and deep learning technologies are gradually applied to medical imaging, with the help of artificial intelligence for intelligent analysis of data pathology sections, allowing pathologists to make diagnoses based on the intelligent analysis results of data pathology sections, providing data information for pathologists to make diagnoses, reducing the tasks of pathologists and improving the efficiency of data pathology section analysis.

[0003] At present, the existing technical solutions usually adopt a fully supervised learning model to perform intelligent analysis on digital pathological sections, which is costly and can only analyze digital pathological sections of specific parts, making the pathological image analysis system have limitations. Therefore, the present invention proposes a pathological image analysis system based on weak supervised learning, which reduces the dependence on labeled data through weak supervised learning, effectively reduces cost consumption, and can also adapt to the characteristics of different types of digital pathological sections, eliminate the limitations of the pathological image analysis system, and improve the scope of application of the pathological image analysis system. Summary of the invention

[0004] The purpose of the present invention is to provide a pathological image analysis system based on weakly supervised learning to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a pathological image analysis system based on weakly supervised learning, comprising: a model processing module, an information delivery module and a data feedback module;

[0006] The model processing module is used to perform image processing and intelligent analysis on digital pathology sections using weakly supervised learning to obtain intelligent analysis results of digital pathology sections;

[0007] The information transmission module is used to encrypt and transmit the results of the intelligent analysis of digital pathological sections to the data feedback module;

[0008] The data feedback module is used to generate feedback information according to the intelligent analysis results of the digital pathology sections and present the feedback information to the pathologist.

[0009] Furthermore, the information delivery module includes: a monitoring confirmation unit, an encryption processing unit and a transfer delivery unit;

[0010] The monitoring confirmation unit is used to monitor the results of the intelligent analysis of digital pathological sections with respect to the model processing module, and obtain the results of the intelligent analysis of digital pathological sections;

[0011] The encryption processing unit is used to divide information for the intelligent analysis results of digital pathology slices, and perform encryption processing according to the information division results matching the corresponding encryption method to obtain encrypted information of the intelligent analysis results of digital pathology slices;

[0012] The transfer unit is used to transmit encrypted information of the intelligent analysis result of the digital pathology slice.

[0013] Furthermore, the data feedback module generates feedback information according to the intelligent analysis results of the digital pathology sections, including:

[0014] Determine the target object of the digital pathology slice according to the intelligent analysis result of the digital pathology slice;

[0015] Retrieve the feedback template according to the target object to obtain the target feedback template;

[0016] Parse the intelligent analysis results of digital pathology slices, and fill the parsing information of the intelligent analysis results of digital pathology slices into the target feedback template to obtain feedback information in the initial state;

[0017] The feedback information in the initial state is marked with key points and image annotated to obtain the feedback information in the final state.

[0018] Furthermore, when the model processing module uses weakly supervised learning to perform image processing and intelligent analysis on digital pathology sections, image label enhancement processing is performed on the digital pathology sections to generate instance-level labels for the digital pathology sections, and then a supervised method is used to perform intelligent analysis based on the instance-level labels of the digital pathology sections to obtain intelligent analysis results for the digital pathology sections.

[0019] Furthermore, image label enhancement processing is performed on the digital pathology slides, including:

[0020] Perform instance segmentation on digital pathology sections to obtain multiple inherited instances;

[0021] Train the MIL model based on the inherited instance and obtain the trained MIL model;

[0022] Use the trained MIL model to output the results for the inheritance instance, and build a fully supervised instance dataset based on the output results;

[0023] The cMIL model is combined with a fully supervised instance dataset to perform image segmentation and annotation on digital pathology sections, and instance-level labels of digital pathology sections are obtained.

[0024] Furthermore, instance segmentation is performed on the digital pathology slices, including:

[0025] Perform preliminary image instance recognition on digital pathology sections to obtain the overall size of the instances in the digital pathology sections;

[0026] Get the image size of digital pathology slides;

[0027] Determine a segmentation scale factor according to an image size of the digital pathology slice and an overall size of instances in the digital pathology slice;

[0028] The digital pathology slices are segmented into multiple inheritance instances according to the segmentation scale factor.

[0029] Furthermore, the MIL model includes: forward propagation and backward propagation. In the forward propagation, the MIL model is trained according to the instance selection criteria to obtain the training results of the inherited instance; in the backward propagation, the image-level label of the inherited instance is obtained, the cross-entropy loss is calculated according to the image-level label of the inherited instance and the training results of the inherited instance, and the parameters of the MIL model are updated according to the cross-entropy loss.

[0030] Furthermore, when training the MIL model according to the instance selection standard, two instance selection standards are used to train the MIL model respectively, one is the Max-Max instance selection standard, and the other is the Max-Min instance selection standard. By training the MIL model according to the instance selection standard, two trained MIL models are obtained, and the initial configuration of the MIL model when trained with different instance selection standards is the same.

[0031] Furthermore, when the trained MIL model is used to output the results for the inheritance instance, two trained MIL models are used to screen and predict the inheritance instance respectively to obtain the label of the inheritance instance, and then the label of the inheritance instance is matched with the inheritance instance, and the matching results are combined, and the combined inheritance instances are combined together to obtain a fully supervised instance data set.

[0032] Furthermore, the cMIL model is a MIL model that combines two instance selection criteria. The cMIL model is used in combination with a fully supervised instance dataset to perform image segmentation and annotation on digital pathology sections, including:

[0033] Training the cMIL model to obtain a cMIL training model;

[0034] Segment the digital pathology slices into instances to obtain digital pathology slice segmentation instances;

[0035] Use the cMIL training model to classify digital pathology slide segmentation examples;

[0036] According to the classification results and the fully supervised instance dataset, the digital pathology slice segmentation instances are annotated, the instance labels of the digital pathology slices are enhanced, and the instance-level labels of the digital pathology slices are obtained.

[0037] The present invention realizes intelligent analysis of digital pathological sections by using weakly supervised learning through the pathological image analysis system, reduces the dependence on labeled data, reduces the need for manual labeling of pathological images of digital pathological sections, thereby reducing the labeling cost consumption, effectively reducing the cost consumption of the pathological image analysis system, and can also adapt to the characteristics of different types of pathological images, so that the pathological image analysis system can perform intelligent analysis on different types of digital pathological sections, eliminates the limitations of the pathological image analysis system, expands the scope of application of the pathological image analysis system, and improves the scalability of the pathological image analysis system.

[0038] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0039] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0041] Figure 1 is a schematic diagram of the pathological image analysis system of the present invention;

[0042] Figure 2 A schematic diagram of an information transmission module in the pathological image analysis system of the present invention;

[0043] Figure 3 A schematic diagram of the steps of the data feedback module in the pathological image analysis system of the present invention;

[0044] Figure 4 A schematic diagram of image label enhancement processing steps of a model processing module in the pathological image analysis system of the present invention;

[0045] Figure 5 A schematic diagram of step A1 of image label enhancement processing of the model processing module in the pathological image analysis system of the present invention;

[0046] Figure 6 This is a schematic diagram of step A4 of the image label enhancement processing of the model processing module in the pathological image analysis system of the present invention. DETAILED DESCRIPTION

[0047] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0048] like Figure 1 As shown, an embodiment of the present invention provides a pathological image analysis system based on weakly supervised learning, including: a model processing module, an information delivery module and a data feedback module;

[0049] The model processing module is used to perform image processing and intelligent analysis on digital pathology sections using weakly supervised learning to obtain intelligent analysis results of digital pathology sections;

[0050] The information transmission module is used to encrypt and transmit the results of the intelligent analysis of digital pathological sections to the data feedback module;

[0051] The data feedback module is used to generate feedback information according to the intelligent analysis results of the digital pathology sections and present the feedback information to the pathologist.

[0052] In the above technical solution, the model processing module is connected to the data feedback module through the information transmission module.

[0053] In the above technical solution, the information transmission module uses wireless transmission to transmit information.

[0054] In the above technical solution, the feedback information is the graphic and text combined data information about the digital pathology section.

[0055] The above technical solution uses weakly supervised learning to realize intelligent analysis of digital pathological sections through the pathological image analysis system, reduces the dependence on labeled data, reduces the need for manual labeling of pathological images of digital pathological sections, thereby reducing the labeling cost consumption, effectively reducing the cost consumption of the pathological image analysis system, and can also adapt to the characteristics of different types of pathological images, so that the pathological image analysis system can perform intelligent analysis on different types of digital pathological sections, eliminate the limitations of the pathological image analysis system, expand the scope of application of the pathological image analysis system, and improve the scalability of the pathological image analysis system. Through the information transmission module, the intelligent analysis results of digital pathological sections are encrypted and transmitted, which can not only realize the remote transmission of the intelligent analysis results of digital pathological sections, but also ensure the security of the intelligent analysis results of digital pathological sections, avoid malicious tampering with the intelligent analysis results of digital pathological sections in the pathological image analysis system, and affect the accuracy of the pathological image analysis system. In addition, the data feedback module can present standardized feedback information to the pathologist, provide information reference for the pathologist, enable the pathologist to make a diagnosis directly based on the feedback information, provide data information for the pathologist to make a diagnosis, and reduce the workload of the pathologist.

[0056] like Figure 2 As shown, in one embodiment provided by the present invention, the information delivery module includes: a monitoring confirmation unit, an encryption processing unit and a transfer delivery unit;

[0057] The monitoring confirmation unit is used to monitor the results of the intelligent analysis of digital pathological sections with respect to the model processing module, and obtain the results of the intelligent analysis of digital pathological sections;

[0058] The encryption processing unit is used to divide information for the intelligent analysis results of digital pathology slices, and perform encryption processing according to the information division results matching the corresponding encryption method to obtain encrypted information of the intelligent analysis results of digital pathology slices;

[0059] The transfer unit is used to transmit encrypted information of the intelligent analysis result of the digital pathology slice.

[0060] In the above technical solution, the monitoring confirmation unit performs real-time monitoring on the model processing module, and when monitoring the digital pathology slice intelligent analysis results on the model processing module, it is determined whether the model processing module has digital pathology slice intelligent analysis results based on the real-time monitoring signal. When the model processing module has digital pathology slice intelligent analysis results, the integrity analysis of the digital pathology slice intelligent analysis results is performed based on the real-time monitoring signal. If the digital pathology slice intelligent analysis results that appear in the real-time monitoring signal within a preset time interval disappear, then a complete digital pathology slice intelligent analysis result appears at this time, and the digital pathology slice intelligent analysis results obtained at this time are moved to the encryption processing unit for encryption processing.

[0061] In the above technical solution, when the encryption processing unit divides the information of the intelligent analysis results of digital pathology sections, it parses and evaluates the information of the intelligent analysis results of digital pathology sections, determines the importance level of the information in the intelligent analysis results of digital pathology sections, and disassembles and combines the information according to the importance level, divides the information of the same importance level together, and obtains the information division result. Among them, the importance level is divided into: first importance, second importance and third importance, the first importance refers to the non-general expression information in the intelligent analysis results of digital pathology sections, the second importance refers to the expression information that expresses the judgment in the general expression information of the intelligent analysis results of digital pathology sections, and the third importance refers to the remaining expression information in the general expression information of the intelligent analysis results of digital pathology sections.

[0062] In the above technical solution, different importance levels correspond to multiple different encryption methods, and at the same importance level, different information types correspond to different encryption methods. When performing encryption processing, the encryption method is matched according to the information classification results and the information type to obtain the encryption method corresponding to the corresponding information in the digital pathology section intelligent analysis results.

[0063] In the above technical solution, when the transfer and delivery unit transmits the encrypted information of the intelligent analysis results of the digital pathology slices, multi-link transmission is performed in combination with the information division result, and when transmitting, the encrypted information of the intelligent analysis results of the digital pathology slices to be transmitted is verified, including: temporarily storing the encrypted information of the intelligent analysis results of the digital pathology slices to be transmitted; obtaining the receiving feedback of the data feedback module based on the encrypted information of the intelligent analysis results of the digital pathology slices; verifying whether the encrypted information of the intelligent analysis results of the digital pathology slices before and after the transmission is consistent according to the received feedback, when the encrypted information of the intelligent analysis results of the digital pathology slices before and after the transmission is consistent, clearing the temporarily stored encrypted information of the intelligent analysis results of the digital pathology slices, and completing the transmission of the encrypted information of the intelligent analysis results of the digital pathology slices; when the encrypted information of the intelligent analysis results of the digital pathology slices before and after the transmission is inconsistent, retrieving the temporarily stored encrypted information of the intelligent analysis results of the digital pathology slices, and retransmitting the encrypted information of the intelligent analysis results of the digital pathology slices until the encrypted information of the intelligent analysis results of the digital pathology slices before and after the transmission is consistent, and when the encrypted information of the intelligent analysis results of the digital pathology slices before and after the transmission is consistent, clearing the temporarily stored encrypted information of the intelligent analysis results of the digital pathology slices.

[0064] The above technical solution realizes the transmission of the intelligent analysis results of digital pathology slices through the information transmission module, so that different processing can be performed remotely, eliminating the limitations of spatial distance, and can also ensure the security of the intelligent analysis results of digital pathology slices. By matching the corresponding encryption method according to the information division result for encryption processing, different encryption methods are used for different information, thereby improving the security of the intelligent analysis results of digital pathology slices. When the transfer and transmission unit transmits the encrypted information of the intelligent analysis results of digital pathology slices, the encrypted information of the intelligent analysis results of digital pathology slices to be transmitted is transmitted and verified, thereby ensuring the consistency of the encrypted information of the intelligent analysis results of digital pathology slices before and after transmission, avoiding the encrypted information of the intelligent analysis results of digital pathology slices from being maliciously tampered with to affect the accuracy of the intelligent analysis results of digital pathology slices, thereby ensuring the accuracy of the pathological image analysis system.

[0065] like Figure 3 As shown, in one embodiment provided by the present invention, the data feedback module generates feedback information according to the intelligent analysis result of the digital pathological slice, including:

[0066] S1. Determine the target object of the digital pathology slice according to the intelligent analysis result of the digital pathology slice;

[0067] S2, retrieve the feedback template according to the target object to obtain the target feedback template;

[0068] S3, parsing the intelligent analysis results of digital pathology slices, and filling the parsing information of the intelligent analysis results of digital pathology slices into the target feedback template to obtain feedback information in the initial state;

[0069] S4. Perform key marking and image annotation on the feedback information in the initial state to obtain the feedback information in the final state.

[0070] In the above technical solution, the target object refers to the organ or tissue contained in the digital pathological section.

[0071] In the above technical solution, different target objects correspond to different feedback templates, and the feedback template is a feedback information framework constructed in advance for the target object.

[0072] In the above technical solution, after obtaining the feedback information in the initial state, artificial intelligence analysis is also performed on the filling information, including: determining the type of the filling information, when the type of the filling information is a chart, performing clarity analysis on the filling information, and judging whether the filling information needs to be processed again in combination with the clarity judgment standard, obtaining the first artificial intelligence analysis result, and then using the artificial intelligence chart processing model to perform secondary processing on the filling information according to the first artificial intelligence analysis result, when the type of the filling information is text, locking the filling sentence based on the filling position, performing preliminary analysis on the filling sentence, judging whether the filling sentence is a complete sentence in the target feedback template, obtaining the second artificial intelligence analysis result, and then According to the second artificial intelligence analysis result, artificial intelligence analysis and adjustment are performed on the fill-in language. When the fill-in sentence is a complete sentence in the target feedback template, the artificial intelligence language analysis model is used to analyze and adjust the fill-in language to ensure that the adjusted fill-in language is fluent and can fully express the information to be expressed. When the fill-in sentence is not a complete sentence in the target feedback template, the position of the fill-in sentence is locked in the target feedback template, and the fill-in language is expanded in combination with the preceding and following sentences of the fill-in sentence to obtain the target analysis sentence. Then, the artificial intelligence language analysis model is used to analyze and adjust the target analysis sentence to ensure that the adjusted target analysis sentence is fluent and can fully express the information to be expressed. Among them, the artificial intelligence chart processing model is a model for image processing of charts, and chart optimization is performed according to the chart. The artificial intelligence language analysis model adjusts the expression of the target analysis sentence or the target analysis sentence through semantic analysis.

[0073] In the above technical solution, key annotation is to highlight the text content of the feedback information, such as: circle annotation, different color marking, etc., and image annotation refers to annotation on the image of the digital pathology section, such as: enlargement and highlighting of some areas, marking of key positions, etc.

[0074] The above technical solution can present standardized feedback information to the pathologist through the data feedback module, so that the pathologist can make a diagnosis based on the feedback, without the need for the pathologist to spend a lot of time analyzing the digital pathology sections, effectively reducing the workload of the pathologist, and the digital pathology sections of different target objects correspond to different feedback templates, so that different information feedback can be performed according to the target object in the feedback information, so that the pathology image analysis system is no longer limited to processing the digital pathology sections of a single object, thereby improving the scope of use of the pathology image analysis system, and by focusing on the feedback information in the initial state and annotating the image, the feedback information has obvious prominent features, so that the key content in the feedback information can attract the attention of the pathologist more, so that the pathologist can efficiently obtain effective feature information from the feedback information, providing convenience for the pathologist to obtain effective information from the feedback information, and then providing convenience for the pathologist to diagnose and analyze the patient's condition, and at the same time, it can also facilitate the pathologist to review the digital pathology sections based on the feedback information to ensure the accuracy of the intelligent analysis results of the digital pathology sections.

[0075] In one embodiment provided by the present invention, when the model processing module uses weakly supervised learning to perform image processing and intelligent analysis on digital pathology sections, image label enhancement processing is performed on the digital pathology sections to generate instance-level labels of the digital pathology sections, and then a supervised method is used to perform intelligent analysis based on the instance-level labels of the digital pathology sections to obtain intelligent analysis results of the digital pathology sections.

[0076] In the above technical solution, when image label enhancement processing is performed on digital pathology sections, instance-level labels are generated based on original image labels of the digital pathology sections.

[0077] In the above technical solution, when a supervised method is used for intelligent analysis based on the instance-level labels of the digital pathology slice, if at least one of the instance-level labels of the digital pathology slice is abnormal, there is at least one lesion area in the digital pathology slice; if all the instance-level labels of the digital pathology slice are normal, there is no lesion area in the digital pathology slice.

[0078] The model processing module of the above technical solution realizes the conversion of supervised methods in the process of image processing and intelligent analysis of digital pathology sections, so that intelligent analysis can be performed on digital pathology sections under any circumstances, improving the robustness and generalization ability of the model processing module, thereby improving the applicability and scalability of the pathology image analysis system, and by performing image label enhancement processing on digital pathology sections, intelligent analysis can be performed even in the absence of sufficient supervisory information, without the need for manual data labeling, reducing dependence on data labeling, saving data labeling time, and reducing data labeling cost consumption.

[0079] like Figure 4 As shown, in one embodiment provided by the present invention, image label enhancement processing is performed on digital pathological sections, including:

[0080] A1. Perform instance segmentation on digital pathology slices to obtain multiple inheritance instances;

[0081] A2. Train the MIL model based on the inherited instance and obtain the trained MIL model;

[0082] A3. Use the trained MIL model to output the results for the inheritance instance, and build a fully supervised instance dataset based on the output results;

[0083] A4. Use the cMIL model combined with the fully supervised instance dataset to perform image segmentation and annotation on digital pathology sections to obtain instance-level labels for digital pathology sections.

[0084] In the above technical solution, the output result is that the trained MIL model performs instance prediction on the inheritance instance.

[0085] The above technical solution adopts the MIL model and cMIL model to improve the efficiency of the model processing module, so that the intelligent analysis results of digital pathology sections can be obtained in a shorter time, thereby improving the efficiency of the pathology image analysis system, enabling pathologists to quickly obtain feedback information, providing convenience for pathologists, and using the trained MIL model to output results for inherited instances, reducing the errors of the MIL model in processing inherited instances, ensuring the accuracy of the output results, and thus ensuring the accuracy of the pathology image analysis system.

[0086] like Figure 5 As shown, in one embodiment provided by the present invention, instance segmentation is performed on digital pathological sections, including:

[0087] B1. Perform preliminary recognition of image instances on digital pathology sections to obtain the overall size of instances in the digital pathology sections;

[0088] B2. Obtain the image size of digital pathology sections;

[0089] B3, determining a segmentation scale factor according to the image size of the digital pathology slice and the overall size of the instance in the digital pathology slice;

[0090] B4. Segment the digital pathology slice into multiple inheritance instances according to the segmentation ratio factor.

[0091] In the above technical solution, the segmentation scale factor is determined according to the image size of the digital pathology slice and the overall size of the instance in the digital pathology slice, and its size is equal to the size of the first processed digital pathology slice image divided by the instance size, that is, N=M / m, wherein N is the instance segmentation scale factor, M is the size of the first processed digital pathology slice image, and m is the instance size.

[0092] In the above technical solution, the digital pathology slice is segmented into N*N inherited instances according to the segmentation scale factor.

[0093] In the above technical solution, the inherited instances have the same size, so that each selected instance with the same label can be regarded as a representative of the corresponding image.

[0094] The above technical solution divides the data pathology slices into blocks by performing instance segmentation on the digital pathology slices, which provides convenience for subsequent analysis. In addition, segmentation is performed based on image instances to reduce the redundancy of irrelevant information, which can effectively improve the efficiency of subsequent analysis and obtain intelligent analysis results of digital pathology slices in a shorter time.

[0095] In one embodiment provided by the present invention, the MIL model includes: forward propagation and backward propagation. In the forward propagation, the MIL model is trained according to the instance selection criteria to obtain the training results of the inherited instance; in the backward propagation, the image-level label of the inherited instance is obtained, the cross-entropy loss is calculated according to the image-level label of the inherited instance and the training results of the inherited instance, and the parameters of the MIL model are updated according to the cross-entropy loss.

[0096] In the above technical solution, ResNet-50 is selected as the classifier in the MIL model.

[0097] In the above technical solution, when training the MIL model, the parameters of the MIL model are updated multiple times, and the optimal values ​​of the parameters of the MIL model are determined according to the cross entropy loss, and the trained MIL model is determined based on the optimal values ​​of the parameters of the MIL model.

[0098] The above technical solution updates and optimizes the parameters in the MIL model through back propagation, so that more accurate output results can be obtained for inheritance instances through MIL model training, thereby improving the accuracy of the MIL model and providing guarantees for the use of the MIL model.

[0099] In an embodiment provided by the present invention, when training a MIL model according to an instance selection standard, two instance selection standards are used to train the MIL models respectively, one is the Max-Max instance selection standard, and the other is the Max-Min instance selection standard. By training the MIL model according to the instance selection standard, two trained MIL models are obtained, and the initial configuration of the MIL model when trained with different instance selection standards is the same.

[0100] In the above technical solution, the Max-Max instance selection criterion is to select the instance with the highest response in the cancer (CA) image as the representative, and to select the instance with the highest CA response in the non-cancer (NC) image as the NC example.

[0101] In the above technical solution, the Max-Min instance selection criterion is to select the instance with the highest response in the CA image.

[0102] The above technical solution improves the performance of the MIL model by adopting two instance selection criteria to train the MIL model respectively, so that the MIL model can highlight CA instances and also reduce the misclassification of CA instances.

[0103] In one embodiment provided by the present invention, when a trained MIL model is used to output results for an inheritance instance, two trained MIL models are used to screen and predict the inheritance instance respectively to obtain labels of the inheritance instance, and then the labels of the inheritance instance and the inheritance instance are matched, combined according to the matching results, and the combined inheritance instances are combined together to obtain a fully supervised instance data set.

[0104] In the above technical solution, each inheritance instance in the fully supervised instance dataset exists with a matching label.

[0105] The above technical solution matches the inheritance instance with the output result of the trained MIL model for the inheritance instance by matching the label of the inheritance instance, so that the sub-data in the fully supervised instance dataset can remain consistent, and the combined inheritance instances are combined to fuse the features of the two instance selection criteria to reduce the problem of distribution bias and obtain a more balanced instance-level dataset.

[0106] like Figure 6 As shown, in one embodiment provided by the present invention, the cMIL model is a MIL model combining two instance selection criteria, and the cMIL model is used in combination with a fully supervised instance dataset to perform image segmentation and annotation on digital pathology sections, including:

[0107] C1. Train the cMIL model to obtain a cMIL training model;

[0108] C2, segmenting the digital pathology slices into instances to obtain digital pathology slice segmentation instances;

[0109] C3, using the cMIL training model to classify digital pathology slide segmentation examples;

[0110] C4. Annotate the digital pathology slice segmentation instances according to the classification results and the fully supervised instance dataset, enhance the instance labels of the digital pathology slices, and obtain the instance-level labels of the digital pathology slices.

[0111] In the above technical solution, the cMIL model is a MIL model that combines the Max-Max instance selection standard and the Max-Min instance selection standard.

[0112] In the above technical solution, the instance-level label of the digital pathology section is N*N high-quality instance labels based on the image of the digital pathology section.

[0113] The above technical solution enables instance-level label acquisition for any digital pathology slice, eliminates the limitations of the pathology image analysis system, and improves the versatility of the pathology image analysis system. In addition, the cMIL model is combined with a fully supervised instance dataset to perform image segmentation and annotation on digital pathology slices, so that training and verification can be performed on a large-scale fully supervised instance dataset, thereby improving the generalization ability of the model processing module and making the pathology image analysis system more robust. In turn, the data feedback module can provide high-quality feedback information to pathologists, providing a guarantee for pathologists to make diagnoses.

[0114] Those skilled in the art should understand that the first and second in the present invention merely refer to different application stages.

[0115] Other embodiments of the present disclosure will be readily apparent to those skilled in the art after considering the specification and practicing the disclosure disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are to be considered as exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0116] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A pathological image analysis system based on weakly supervised learning, characterized in that: include: Model processing module, information delivery module and data feedback module; The model processing module is used to perform image processing and intelligent analysis on digital pathology sections using weakly supervised learning to obtain intelligent analysis results of digital pathology sections; The information transmission module is used to encrypt and transmit the results of the intelligent analysis of digital pathological sections to the data feedback module; The data feedback module is used to generate feedback information based on the intelligent analysis results of the digital pathology sections and present the feedback information to the pathologist.

2. The pathological image analysis system according to claim 1, characterized in that: The information delivery module includes: a monitoring confirmation unit, an encryption processing unit and a transfer delivery unit; The monitoring confirmation unit is used to monitor the results of the intelligent analysis of digital pathological sections with respect to the model processing module, and obtain the results of the intelligent analysis of digital pathological sections; The encryption processing unit is used to divide information for the intelligent analysis results of digital pathology slices, and perform encryption processing according to the information division results matching the corresponding encryption method to obtain encrypted information of the intelligent analysis results of digital pathology slices; The transfer unit is used to transmit encrypted information of the intelligent analysis result of the digital pathology slice.

3. The pathological image analysis system according to claim 1, characterized in that: The data feedback module generates feedback information according to the intelligent analysis results of digital pathological sections, including: Determine the target object of the digital pathology slice according to the intelligent analysis result of the digital pathology slice; Retrieve the feedback template according to the target object to obtain the target feedback template; Parse the intelligent analysis results of digital pathology slices, and fill the parsing information of the intelligent analysis results of digital pathology slices into the target feedback template to obtain feedback information in the initial state; The feedback information in the initial state is marked with key points and image annotated to obtain the feedback information in the final state.

4. The pathological image analysis system according to claim 1, characterized in that: When the model processing module uses weakly supervised learning to perform image processing and intelligent analysis on digital pathology sections, image label enhancement processing is performed on the digital pathology sections to generate instance-level labels of the digital pathology sections, and then a supervised method is used to perform intelligent analysis based on the instance-level labels of the digital pathology sections to obtain intelligent analysis results of the digital pathology sections.

5. The pathological image analysis system according to claim 4, characterized in that: Image label enhancement processing for digital pathology slides, including: Perform instance segmentation on digital pathology sections to obtain multiple inherited instances; Train the MIL model based on the inherited instance and obtain the trained MIL model; Use the trained MIL model to output the results for the inheritance instance, and build a fully supervised instance dataset based on the output results; The cMIL model is combined with a fully supervised instance dataset to perform image segmentation and annotation on digital pathology sections, and instance-level labels of digital pathology sections are obtained.

6. The pathological image analysis system according to claim 5, characterized in that: Perform instance segmentation on digital pathology slides, including: Perform preliminary image instance recognition on digital pathology sections to obtain the overall size of the instances in the digital pathology sections; Get the image size of digital pathology slides; Determine a segmentation scale factor according to an image size of the digital pathology slice and an overall size of instances in the digital pathology slice; The digital pathology slices are segmented into multiple inheritance instances according to the segmentation scale factor.

7. The pathological image analysis system according to claim 5, characterized in that: The MIL model includes: forward propagation and back propagation. In the forward propagation, the MIL model is trained according to the instance selection standard to obtain the training result of the inherited instance; in the back propagation, the image-level label of the inherited instance is obtained, the cross-entropy loss is calculated according to the image-level label of the inherited instance and the training result of the inherited instance, and the parameters of the MIL model are updated according to the cross-entropy loss.

8. The pathological image analysis system according to claim 7, characterized in that: When training the MIL model according to the instance selection standard, two instance selection standards are used to train the MIL model respectively, one is the Max-Max instance selection standard, and the other is the Max-Min instance selection standard. By training the MIL model according to the instance selection standard, two trained MIL models are obtained. The initial configuration of the MIL model when trained with different instance selection standards is the same.

9. The pathological image analysis system according to claim 8, characterized in that: When the trained MIL model is used to output the results for the inheritance instance, two trained MIL models are used to screen and predict the inheritance instance respectively to obtain the label of the inheritance instance, and then the label of the inheritance instance is matched with the inheritance instance, and the matching results are combined, and the combined inheritance instances are combined together to obtain a fully supervised instance data set.

10. The pathological image analysis system according to claim 5, characterized in that: The cMIL model is a MIL model that combines two instance selection criteria. The cMIL model is used in combination with a fully supervised instance dataset to perform image segmentation and annotation on digital pathology sections, including: Training the cMIL model to obtain a cMIL training model; Segment the digital pathology slices into instances to obtain digital pathology slice segmentation instances; Use the cMIL training model to classify digital pathology slide segmentation examples; According to the classification results and the fully supervised instance dataset, the digital pathology slice segmentation instances are annotated, the instance labels of the digital pathology slices are enhanced, and the instance-level labels of the digital pathology slices are obtained.

Images