Homologous recombination defect analysis system for ovarian cancer based on weakly supervised learning

Abstract

The present invention provides an ovarian cancer homologous recombination defect analysis system based on weak supervised learning, including: a system for cutting a pathological slice image into a plurality of image sub-blocks of specified specifications and performing feature enhancement to generate enhanced image sub-blocks, calling a training method to respectively enhance the image features of each enhanced image sub-block, a physician only needs to provide a weak supervisory signal of correlation prompt, a model can be trained, global information and a plurality of local information of the pathological slice image are obtained, HRD scoring is performed on each local information respectively, a plurality of image features contained in the pathological slice image are analyzed, an HRD distribution heat map is constructed in combination with the global information, a medical marker added by a pathologist in the lesion defect distribution heat map is obtained and analyzed, a defect analysis report of a patient corresponding to the pathological slice image is generated and displayed, so that medical professionals can interpret the results more intuitively, thereby improving the reliability of diagnosis and the accuracy of clinical decision-making.

Description

Technical Field

[0001] The present invention relates to the medical field, and in particular to an ovarian cancer homologous recombination defect analysis system based on weakly supervised learning. Background Art

[0002] Ovarian cancer (EOC) is one of the leading causes of death in patients with gynecological malignancies, and the incidence of EOC has been gradually increasing. Most EOCs are high-grade serous carcinomas (HGSC), and the rest are low-grade serous carcinomas, clear cell carcinomas, endometrioid adenocarcinomas, and mucinous adenocarcinomas. By analyzing the molecular characteristics of ovarian tumors in the Cancer Genome Atlas (TCGA) database, it was found that different ovarian tumor subtypes have different molecular characteristics.

[0003] Currently, pathologists can predict the prognosis of patients by analyzing the molecular characteristics of different ovarian tumor subtypes. However, this technology requires amplification and sequencing of the patient's DNA, which consumes a lot of medical resources and manpower. In addition, the data used by traditional models to train the models requires a lot of precise labeling work. Now, an efficient and fast analysis method is urgently needed to replace the old method.

[0004] Therefore, the present invention provides an ovarian cancer homologous recombination defect analysis system based on weakly supervised learning. Summary of the invention

[0005] The ovarian cancer homologous recombination defect analysis system of the present invention based on weakly supervised learning improves the accuracy of HRD prediction, can help medical professionals estimate the prognosis of patients in a short time, and enables medical professionals to interpret the results more intuitively, thereby improving the reliability of diagnosis and the accuracy of clinical decision-making.

[0006] The present invention provides an ovarian cancer homologous recombination defect analysis system based on weakly supervised learning, comprising:

[0007] A slice enhancement module is used to cut the pathological slice image into a number of image sub-blocks of specified specifications, and perform feature enhancement on each of the image sub-blocks to generate enhanced image sub-blocks;

[0008] A feature training module, used to respectively obtain the WSI features corresponding to each of the enhanced image sub-blocks, and to call a training method to respectively train each of the WSI features to obtain the global information and a plurality of local information of the pathological slice image;

[0009] A defect analysis module, used to perform HRD scoring on each of the local information, analyze several pathological factors contained in the pathological slice image according to the scoring results, and establish a pathological defect distribution heat map in combination with the global information;

[0010] The marking processing module is used to obtain the medical marks added by the pathologist in the lesion defect distribution heat map, respectively obtain the semantic information corresponding to each of the medical marks, generate a defect analysis report for the patient corresponding to the pathological section image, and display it.

[0011] In one practicable manner,

[0012] Also includes:

[0013] According to the patient informed consent agreement, the pathological sample of the corresponding patient is retrieved and sliced ​​to obtain the pathological slice of the patient;

[0014] The pathological slice is transmitted to a preset shooting area for shooting to obtain a pathological slice image of the corresponding patient.

[0015] In one practicable manner,

[0016] The slice enhancement module comprises:

[0017] An image cutting unit, used for adjusting the image input parameters of a preset ViT framework according to the image specifications of the pathological slice image, inputting the pathological slice image into the adjusted preset ViT framework for cutting, and obtaining a plurality of image sub-blocks;

[0018] An image encoding unit, configured to encode each of the image sub-blocks respectively to obtain a compressed sub-block corresponding to each of the image sub-blocks, establish a redundant element of the corresponding image sub-block according to a pixel difference between the image sub-block and the corresponding compressed sub-block, and map the redundant element to the corresponding image sub-block to obtain a redundant distribution feature of each of the image sub-blocks;

[0019] An image fitting unit is used to respectively obtain the actual grayscale value corresponding to each pixel point in the image sub-block, respectively perform grayscale sampling on each image sub-block to establish a grayscale matrix of the corresponding image sub-block, respectively perform fitting training on each matrix element in the grayscale matrix to obtain an ideal grayscale matrix corresponding to each image sub-block, and analyze the grayscale difference between the actual grayscale value and the ideal grayscale value corresponding to each pixel point in the corresponding image sub-block based on the ideal grayscale matrix to establish a grayscale deficiency feature of each image sub-block;

[0020] The image enhancement unit is used to establish an anti-redundancy layer and a grayscale compensation layer corresponding to the corresponding image sub-block in the preset ViT framework according to the redundant distribution characteristics and grayscale deficiency characteristics corresponding to the same image sub-block, and cover the anti-redundancy layer and the grayscale compensation layer in the corresponding image sub-block to generate an enhanced image sub-block.

[0021] In one practicable manner,

[0022] The image fitting unit comprises:

[0023] A grayscale sampling subunit, used to respectively obtain the actual grayscale value corresponding to each pixel point in the image subblock, respectively sample the four corner pixel points in each of the image subblocks, and perform a plurality of random samplings on each of the image subblocks to obtain a plurality of sampled grayscale values;

[0024] A matrix construction subunit, for establishing a frame prototype based on the pixel position of each sampled gray value in the image sub-block, rectangularizing the frame prototype to obtain a matrix frame, and inputting the sampled gray value of the corresponding image sub-block into the matrix frame to generate a gray matrix of the corresponding image sub-block;

[0025] A matrix training unit, used to establish a plurality of transverse fitting curves and longitudinal fitting curves according to element differences between adjacent matrix elements in the grayscale matrix, perform depth training on curve points in each of the transverse fitting curves and each of the longitudinal fitting curves using the least squares method to obtain a depth fitting parameter corresponding to each matrix element, and train the corresponding matrix elements using the depth fitting parameters to obtain an ideal grayscale matrix corresponding to each of the image sub-blocks;

[0026] The feature generation unit is used to analyze the grayscale difference between the actual grayscale value and the ideal grayscale value corresponding to each pixel point in the corresponding image sub-block based on the ideal grayscale matrix, and establish the grayscale deficiency feature of each image sub-block.

[0027] In one practicable manner,

[0028] The feature training module comprises:

[0029] A feature analysis unit, used to establish a plurality of homologous recombination defect presentation samples according to the presentation characteristics of the homologous recombination defect, use the homologous recombination defect presentation samples to respectively capture the WSI information contained in each of the enhanced image sub-blocks, establish the WSI features corresponding to each of the enhanced image sub-blocks, and establish the overall WSI features of the pathological slice image;

[0030] A feature training unit, configured to analyze a maximum lesion threshold of the pathological slice image according to the overall WSI feature, and to perform scaling training on a feature range corresponding to each WSI feature according to the maximum lesion threshold, so as to obtain a lesion area range corresponding to each enhanced image sub-block;

[0031] A local analysis unit is used to mark a corresponding weak label for each of the enhanced image sub-blocks according to the WSI features, identify the lesion area in each pixel point in the corresponding enhanced image sub-block according to the weak label and lesion area range corresponding to the same enhanced image sub-block, and determine the lesion probability corresponding to each pixel point in the enhanced image sub-block, and establish local information corresponding to each of the enhanced image sub-blocks;

[0032] A global construction unit is used to construct the global information of the pathological slice at a source position of the enhanced image sub-block in the pathological slice image according to the local information.

[0033] In one practicable manner,

[0034] The defect analysis module comprises:

[0035] A scoring execution unit, configured to perform HRD scoring on each of the local information respectively, and when the HRD score value of the local information is greater than or equal to 30, establish a lesion area contour of the corresponding enhanced image sub-block according to each of the local information;

[0036] A lesion establishing unit, used for reorganizing different lesion contour areas to obtain a global lesion contour of the pathological slice image, obtaining a lesion image of the pathological slice image, acquiring a plurality of image key points in the lesion image, and establishing a plurality of lesion factors according to the image key points;

[0037] The heat map creation unit is used to mark each of the pathological factors on the pathological slice image respectively to generate a lesion defect distribution heat map.

[0038] In one practicable manner,

[0039] The marking processing module comprises:

[0040] A semantic processing unit, used to obtain the medical markers added by the pathologist in the lesion defect distribution heat map, and obtain semantic information corresponding to each of the medical markers;

[0041] The heat map analysis unit is used to establish and display the patient's defect analysis report based on the lesion variable corresponding to each lesion factor in the lesion defect distribution heat map in combination with the corresponding semantic information.

[0042] In one practicable manner,

[0043] The global construction unit comprises:

[0044] An image recombining subunit, configured to establish a recombined pathological slice image according to the local information and the source position of the enhanced image sub-block in the pathological slice image;

[0045] A loss analysis subunit, used for calculating the combined loss weights between adjacent enhanced image sub-blocks in the reconstructed pathological slice image according to formula (1);

[0046]

[0047] Wherein, Q represents the combined loss weight of the reconstructed pathological slice image, m represents the number of enhanced image sub-blocks, Indicates the first label value corresponding to the weak label;

[0048] The information adjustment subunit is used to establish the global information of the pathological slice image according to the combined loss weight and the local information corresponding to each enhanced image sub-block.

[0049] In one practicable manner,

[0050] The process of marking a corresponding weak label for each enhanced image sub-block according to the WSI feature comprises:

[0051] Determine the sub-block importance level of the corresponding enhanced image sub-block in the pathological slice image according to the WSI feature, and match the corresponding enhanced image sub-block with a corresponding image definition label according to the sub-block importance level;

[0052] Matching a corresponding weak label for each enhanced image sub-block according to formula (2);

[0053]

[0054] Represented as the weak label matched by the i-th enhanced image sub-block, i The second label value of the image definition label corresponding to the i-th enhanced image sub-block, M n represents the lesion probability of the WSI feature corresponding to the i-th enhanced image sub-block.

[0055] In one practicable manner,

[0056] Also includes:

[0057] The enhanced image sub-blocks with consistent weak labels are clustered to generate image classes with the same label and displayed.

[0058] The achievable beneficial effects of the above technical solution are as follows: in order to assist pathologists in clarifying the HRD mutation status of patients, the pathological slice image is first cut, and in order to improve the annotation efficiency of pathologists, only the relevant areas need to be annotated; in order to improve the accuracy of weakly supervised learning, each image sub-block is enhanced accordingly, and then the corresponding WSI features are obtained, and the WSI features are further trained accordingly through different training methods, thereby determining the global information and local information of the pathological slice image, and further preliminarily determining several pathological factors contained in the pathological slice image by performing HRD scoring on the local information, and then providing a heat map output, so that medical professionals can interpret the results more intuitively, thereby improving the reliability of diagnosis and the accuracy of clinical decision-making, and further generating the HRD analysis report of the patient according to the marks added by the pathologist on the heat map. Such technical means help pathologists improve the reliability of diagnosis and the accuracy of clinical decision-making, which is of great significance to personalized and precise oncology, improves the accuracy of HRD prediction, deepens our understanding of tumor biology, and ultimately improves the treatment outcomes of patients.

[0059] 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, claims, and drawings.

[0060] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] 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:

[0062] Figure 1 It is a schematic diagram of the composition of the ovarian cancer homologous recombination defect analysis system based on weakly supervised learning in an embodiment of the present invention;

[0063] Figure 2 Schematic diagram of the composition of the slice enhancement module of the ovarian cancer homologous recombination defect analysis system based on weakly supervised learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] 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.

[0065] Example 1

[0066] This embodiment provides an ovarian cancer homologous recombination defect analysis system based on weakly supervised learning, such as Figure 1 As shown, including:

[0067] A slice enhancement module is used to cut the pathological slice image into a number of image sub-blocks of specified specifications, and perform feature enhancement on each of the image sub-blocks to generate enhanced image sub-blocks;

[0068] A feature training module, used to respectively obtain the WSI features corresponding to each of the enhanced image sub-blocks, and to call a training method to respectively train each of the WSI features to obtain the global information and a plurality of local information of the pathological slice image;

[0069] A defect analysis module, used to perform HRD scoring on each of the local information, analyze several pathological factors contained in the pathological slice image according to the scoring results, and establish a pathological defect distribution heat map in combination with the global information;

[0070] The marking processing module is used to obtain the medical marks added by the pathologist in the lesion defect distribution heat map, respectively obtain the semantic information corresponding to each of the medical marks, generate a defect analysis report for the patient corresponding to the pathological section image, and display it.

[0071] In this example, the specified specification is 9*9;

[0072] In this example, the WSI feature represents the feature extracted from the pathological section and used to describe the tissue structure and cell morphology of ovarian cells;

[0073] In this example, the global information refers to the information presented by the entire pathological slice image, and the local information refers to the information presented by an image sub-block in the pathological slice image;

[0074] In this example, pathological factors include: the histopathological structure of ovarian cancer presents solid, papillary, labyrinthine, and glandular growth patterns, and the size of the nucleus is more than three times that of the normal cell nucleus;

[0075] In this example, the medical markers refer to the markers added by the pathologist to the lesion defect distribution heat map.

[0076] The working principle and beneficial effects of the above technical solution are as follows: in order to assist pathologists in clarifying the HRD mutation status of patients, the pathological slice image is first cut, and each image sub-block is enhanced accordingly to improve the accuracy of weakly supervised learning. In order to improve the annotation efficiency of pathologists, only the relevant area needs to be annotated as the learning supervision signal of the model, and then the corresponding WSI features are obtained, and the WSI features are further trained accordingly through different training methods, thereby determining the global information and local information of the pathological slice image, and further preliminarily determining several pathological factors contained in the pathological slice image by performing HRD scoring on the local information, and then providing a heat map output, so that medical professionals can interpret the results more intuitively, thereby improving the reliability of diagnosis and the accuracy of clinical decision-making, and further generating a defect analysis report for the patient based on the marks added by the pathologist on the heat map. Such technical means help pathologists improve the reliability of diagnosis and the accuracy of clinical decision-making, which is of great significance to personalized and precision oncology, improves the accuracy of HRD prediction, deepens our understanding of tumor biology, and ultimately improves the treatment outcomes of patients.

[0077] Example 2

[0078] Based on Example 1, the ovarian cancer homologous recombination defect analysis system based on weakly supervised learning further includes:

[0079] According to the patient informed consent agreement, the pathological sample of the corresponding patient is retrieved and sliced ​​to obtain the pathological slice of the patient;

[0080] The pathological slice is transmitted to a preset shooting area for shooting to obtain a pathological slice image of the corresponding patient.

[0081] The working principle and beneficial effects of the above technical solution: In order to achieve humanistic care, pathological sections are photographed with the patient's knowledge to lay the basis for subsequent analysis.

[0082] Example 3

[0083] Based on Example 1, the ovarian cancer homologous recombination defect analysis system based on weakly supervised learning is as follows: Figure 2 As shown, the slice enhancement module includes:

[0084] An image cutting unit, used for adjusting the image input parameters of a preset ViT framework according to the image specifications of the pathological slice image, inputting the pathological slice image into the adjusted preset ViT framework for cutting, and obtaining a plurality of image sub-blocks;

[0085] An image encoding unit, configured to encode each of the image sub-blocks respectively to obtain a compressed sub-block corresponding to each of the image sub-blocks, establish a redundant element of the corresponding image sub-block according to a pixel difference between the image sub-block and the corresponding compressed sub-block, and map the redundant element to the corresponding image sub-block to obtain a redundant distribution feature of each of the image sub-blocks;

[0086] An image fitting unit is used to respectively obtain the actual grayscale value corresponding to each pixel point in the image sub-block, respectively perform grayscale sampling on each image sub-block to establish a grayscale matrix of the corresponding image sub-block, respectively perform fitting training on each matrix element in the grayscale matrix to obtain an ideal grayscale matrix corresponding to each image sub-block, and analyze the grayscale difference between the actual grayscale value and the ideal grayscale value corresponding to each pixel point in the corresponding image sub-block based on the ideal grayscale matrix to establish a grayscale deficiency feature of each image sub-block;

[0087] The image enhancement unit is used to establish an anti-redundancy layer and a grayscale compensation layer corresponding to the corresponding image sub-block in the preset ViT framework according to the redundant distribution characteristics and grayscale deficiency characteristics corresponding to the same image sub-block, and cover the anti-redundancy layer and the grayscale compensation layer in the corresponding image sub-block to generate an enhanced image sub-block.

[0088] In this example, the preset ViT framework represents a technical framework for image segmentation;

[0089] In this example, the image input parameter represents a parameter used to adapt to the image specifications of the pathological cut image, thereby adjusting the storage space of the image processing area of ​​the preset ViT framework;

[0090] In this example, the compressed sub-block represents the image generated by eliminating redundant data in the image sub-block by encoding;

[0091] In this example, the redundant element represents non-critical information in the image sub-block, that is, information in the image sub-block that does not belong to the pathological factor;

[0092] In this example, the grayscale matrix represents a matrix obtained by arranging the actual grayscale values ​​obtained by sampling;

[0093] In this example, the grayscale compensation layer represents a layer established in the preset ViT framework and used to perform grayscale compensation on the image sub-block. When performing compensation, the layer is covered on the corresponding image sub-block;

[0094] In this example, the anti-redundancy layer refers to a layer established in the preset ViT framework and used to eliminate redundancy of image sub-blocks. When eliminating redundancy, the layer is covered on the corresponding image sub-block.

[0095] The working principle and beneficial effects of the above technical solution: In order to further improve the effectiveness of defect analysis, the image input parameters of the ViT framework are first adjusted according to the image specifications of the pathological slice image, and then the pathological slice image is cut to ensure the uniformity of the cutting, and a number of image sub-blocks are obtained. Then, the image sub-blocks are encoded to obtain compressed sub-blocks corresponding to each image sub-block, and the redundant elements contained in the image sub-blocks are determined. The distribution of the redundant elements in the image sub-blocks is determined by mapping, so as to establish redundant distribution features, and then the actual grayscale values ​​corresponding to each pixel in the image sub-block are sampled to establish a grayscale matrix, and then an ideal grayscale matrix is ​​obtained by fitting training the grayscale matrix, and then the grayscale deficiency feature of the image sub-block is established according to the grayscale difference between the two, and then the image sub-block is correspondingly eliminated and compensated for the two features in the preset ViT framework, and finally the enhanced image sub-block is obtained. Through this technology, not only the image sub-block can be enhanced, but also the corresponding enhancement can be performed according to the characteristics of different image sub-blocks during the enhancement process, thereby achieving precise enhancement and improving the enhancement effect.

[0096] Example 4

[0097] On the basis of Example 3, the image fitting unit of the ovarian cancer homologous recombination defect analysis system based on weakly supervised learning includes:

[0098] A grayscale sampling subunit, used to respectively obtain the actual grayscale value corresponding to each pixel point in the image subblock, respectively sample the four corner pixel points in each of the image subblocks, and perform a plurality of random samplings on each of the image subblocks to obtain a plurality of sampled grayscale values;

[0099] A matrix construction subunit, for establishing a frame prototype based on the pixel position of each sampled gray value in the image sub-block, rectangularizing the frame prototype to obtain a matrix frame, and inputting the sampled gray value of the corresponding image sub-block into the matrix frame to generate a gray matrix of the corresponding image sub-block;

[0100] A matrix training unit, used to establish a plurality of transverse fitting curves and longitudinal fitting curves according to element differences between adjacent matrix elements in the grayscale matrix, perform depth training on curve points in each of the transverse fitting curves and each of the longitudinal fitting curves using the least squares method to obtain a depth fitting parameter corresponding to each matrix element, and train the corresponding matrix elements using the depth fitting parameters to obtain an ideal grayscale matrix corresponding to each of the image sub-blocks;

[0101] The feature generation unit is used to analyze the grayscale difference between the actual grayscale value and the ideal grayscale value corresponding to each pixel point in the corresponding image sub-block based on the ideal grayscale matrix, and establish the grayscale deficiency feature of each image sub-block.

[0102] In this example, the lateral fitting curve represents a curve established in a row vector according to the element differences between adjacent matrix elements;

[0103] In this example, the longitudinal fitting curve represents a curve established in a column vector according to the element differences between adjacent matrix elements.

[0104] The working principle and beneficial effects of the above technical solution are as follows: In order to improve the effectiveness of sampling and prevent the matrix frame from being disordered, the four corner pixel points of the image sub-block are fixed during sampling, and then random sampling is performed to obtain a number of sampling grayscale values. The frame prototype is established based on the sampling grayscale values ​​and the pixel positions in the image sub-block, and then the matrix frame is generated. The grayscale matrix is ​​established by inputting the sampling grayscale values ​​into the matrix frame, and the matrix elements are deeply trained according to the horizontal fitting curve and the vertical fitting curve between adjacent matrix elements to obtain the corresponding deep fitting matrix, thereby establishing an ideal grayscale matrix. Finally, the ideal grayscale matrix is ​​used to compare the actual grayscale values ​​of each pixel point to generate a grayscale deficiency feature, which serves as a basis for subsequent local and global analysis.

[0105] Example 5

[0106] Based on Example 1, the ovarian cancer homologous recombination defect analysis system based on weakly supervised learning, the feature training module includes:

[0107] A feature analysis unit, used to establish a plurality of homologous recombination defect presentation samples according to the presentation characteristics of the homologous recombination defect, use the homologous recombination defect presentation samples to respectively capture the WSI information contained in each of the enhanced image sub-blocks, establish the WSI features corresponding to each of the enhanced image sub-blocks, and establish the overall WSI features of the pathological slice image;

[0108] A feature training unit, configured to analyze a maximum lesion threshold of the pathological slice image according to the overall WSI feature, and to perform scaling training on a feature range corresponding to each WSI feature according to the maximum lesion threshold, so as to obtain a lesion area range corresponding to each enhanced image sub-block;

[0109] A local analysis unit is used to mark a corresponding weak label for each of the enhanced image sub-blocks according to the WSI features, identify the lesion area in each pixel point in the corresponding enhanced image sub-block according to the weak label and lesion area range corresponding to the same enhanced image sub-block, and determine the lesion probability corresponding to each pixel point in the enhanced image sub-block, and establish local information corresponding to each of the enhanced image sub-blocks;

[0110] A global construction unit is used to construct the global information of the pathological slice at a source position of the enhanced image sub-block in the pathological slice image according to the local information.

[0111] In this example, the maximum lesion threshold is 30;

[0112] In this example, the lesion area range indicates the range of the area where the lesion occurs in an enhanced image sub-block;

[0113] In this example, the weak label indicates that the lesion amount of the corresponding enhanced image sub-block is determined according to the WSI feature, thereby assigning a low-precision label to the enhanced image sub-block.

[0114] The working principle and beneficial effects of the above technical solution are as follows: In order to accurately analyze the pathological slice image, a number of defect presentation samples are first established in advance, and the WSI information contained in the enhanced sub-image is captured through the samples, and its corresponding WSI features and the overall WSI features of the pathological slice image are established. The lesion area range corresponding to the enhanced image sub-block is further determined by scaling, and the lesion probability corresponding to each pixel is analyzed by annotating the enhanced image sub-block with corresponding weak labels combined with its corresponding lesion area range, and the local information of the enhanced image sub-block is established. The full set of information of the pathological slice image is further established, and the analysis of the pathological slice image is completed, achieving the purpose of assisting pathologists in diagnosis.

[0115] Example 6

[0116] On the basis of Example 1, the ovarian cancer homologous recombination defect analysis system based on weakly supervised learning, the defect analysis module includes:

[0117] A scoring execution unit, configured to perform HRD scoring on each of the local information respectively, and when the HRD score value of the local information is greater than or equal to 30, establish a lesion area contour of the corresponding enhanced image sub-block according to each of the local information;

[0118] A lesion establishing unit, used for reorganizing different lesion contour areas to obtain a global lesion contour of the pathological slice image, obtaining a lesion image of the pathological slice image, acquiring a plurality of image key points in the lesion image, and establishing a plurality of lesion factors according to the image key points;

[0119] The heat map creation unit is used to mark each of the pathological factors on the pathological slice image respectively to generate a lesion defect distribution heat map.

[0120] In this example, a tumor is considered HRD-positive if the HRD score is high (≥30) or if the tumor carries a pathogenic or likely pathogenic BRCA1 / 2 mutation, and is considered HRD-negative if the HRD score is low (<30) and the tumor does not have a pathogenic or likely pathogenic BRCA1 / 2 mutation.

[0121] The working principle and beneficial effects of the above technical solution are: to preliminarily analyze whether the patient is sick through scoring, and to determine the pathological factors when the patient is sick for reference by the pathologist.

[0122] Example 7

[0123] On the basis of Example 1, the ovarian cancer homologous recombination defect analysis system based on weakly supervised learning, the marker processing module includes:

[0124] A semantic processing unit, used to obtain the medical markers added by the pathologist in the lesion defect distribution heat map, and obtain semantic information corresponding to each of the medical markers;

[0125] The heat map analysis unit is used to establish and display the patient's defect analysis report based on the lesion variable corresponding to each lesion factor in the lesion defect distribution heat map in combination with the corresponding semantic information.

[0126] The working principle and beneficial effects of the above technical solution: In order to make the report more user-friendly and reduce the work pressure of the pathologist, the pathologist adds medical tags to the lesion defect distribution heat map, and then establishes the patient's defect analysis report by analyzing the semantic information of each medical tag, so that the pathologist can understand the patient's condition at a glance.

[0127] Example 8

[0128] On the basis of Example 5, the ovarian cancer homologous recombination defect analysis system based on weakly supervised learning, the global construction unit includes:

[0129] An image recombining subunit, configured to establish a recombined pathological slice image according to the local information and the source position of the enhanced image sub-block in the pathological slice image;

[0130] A loss analysis subunit, used for calculating the combined loss weights between adjacent enhanced image sub-blocks in the reconstructed pathological slice image according to formula (1);

[0131]

[0132] Wherein, Q represents the combined loss weight of the reconstructed pathological slice image, m represents the number of enhanced image sub-blocks, Indicates the first label value corresponding to the weak label;

[0133] The information adjustment subunit is used to establish the global information of the pathological slice image according to the combined loss weight and the local information corresponding to each enhanced image sub-block.

[0134] The working principle and beneficial effects of the above technical solution: Since different enhanced images can only reflect partial information, in order to ensure the validity of global information, the loss weights between different enhanced image sub-blocks are analyzed during image reconstruction, so as to make corresponding adjustments to the local information and generate global information.

[0135] Example 9

[0136] On the basis of Example 5, the weakly supervised learning-based ovarian cancer homologous recombination defect analysis system, according to the WSI features, respectively labels each enhanced image sub-block with a corresponding weak label, including:

[0137] Determine the sub-block importance level of the corresponding enhanced image sub-block in the pathological slice image according to the WSI feature, and match the corresponding enhanced image sub-block with a corresponding image definition label according to the sub-block importance level;

[0138] Matching a corresponding weak label for each enhanced image sub-block according to formula (2);

[0139]

[0140] Represented as the weak label matched by the i-th enhanced image sub-block, i The second label value of the image definition label corresponding to the i-th enhanced image sub-block, M n represents the lesion probability of the WSI feature corresponding to the i-th enhanced image sub-block.

[0141] In this example, the image definition label represents a label established according to the importance of an enhanced image sub-block in the pathological slice image.

[0142] The working principle and beneficial effects of the above technical solution are as follows: according to the importance level of the enhanced image sub-block in the pathological slice image and the lesion probability of the enhanced image sub-block, the corresponding weak label is matched for the enhanced image sub-block, thereby ensuring the effectiveness of the weak label allocation.

[0143] Example 10

[0144] Based on Example 9, the ovarian cancer homologous recombination defect analysis system based on weakly supervised learning further includes:

[0145] The enhanced image sub-blocks with consistent weak labels are clustered to generate image classes with the same label and displayed.

[0146] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. An ovarian cancer homologous recombination defect analysis system based on weakly supervised learning, characterized in that: include: A slice enhancement module is used to cut the pathological slice image into a number of image sub-blocks of specified specifications, and perform feature enhancement on each of the image sub-blocks to generate enhanced image sub-blocks; A feature training module, used to respectively obtain the WSI features corresponding to each of the enhanced image sub-blocks, and to call a training method to respectively train each of the WSI features to obtain the global information and a plurality of local information of the pathological slice image; A defect analysis module, used to perform HRD scoring on each of the local information, analyze several pathological factors contained in the pathological slice image according to the scoring results, and establish a pathological defect distribution heat map in combination with the global information; A marking processing module, used to obtain the medical markings added by the pathologist in the lesion defect distribution heat map, respectively obtain the semantic information corresponding to each of the medical markings, generate a defect analysis report for the patient corresponding to the pathological section image, and display it; The slice enhancement module comprises: An image cutting unit, used for adjusting the image input parameters of a preset ViT framework according to the image specifications of the pathological slice image, inputting the pathological slice image into the adjusted preset ViT framework for cutting, and obtaining a plurality of image sub-blocks; An image encoding unit, configured to encode each of the image sub-blocks respectively to obtain a compressed sub-block corresponding to each of the image sub-blocks, establish a redundant element of the corresponding image sub-block according to a pixel difference between the image sub-block and the corresponding compressed sub-block, and map the redundant element to the corresponding image sub-block to obtain a redundant distribution feature of each of the image sub-blocks; An image fitting unit is used to respectively obtain the actual grayscale value corresponding to each pixel point in the image sub-block, respectively perform grayscale sampling on each image sub-block to establish a grayscale matrix of the corresponding image sub-block, respectively perform fitting training on each matrix element in the grayscale matrix to obtain an ideal grayscale matrix corresponding to each image sub-block, and analyze the grayscale difference between the actual grayscale value and the ideal grayscale value corresponding to each pixel point in the corresponding image sub-block based on the ideal grayscale matrix to establish a grayscale deficiency feature of each image sub-block; An image enhancement unit, configured to establish an anti-redundancy layer and a grayscale compensation layer corresponding to a corresponding image sub-block in the preset ViT framework according to redundant distribution characteristics and grayscale deficiency characteristics corresponding to the same image sub-block, and cover the anti-redundancy layer and the grayscale compensation layer in the corresponding image sub-block to generate an enhanced image sub-block; The image fitting unit comprises: A grayscale sampling subunit, used to respectively obtain the actual grayscale value corresponding to each pixel point in the image subblock, respectively sample the four corner pixel points in each of the image subblocks, and perform a plurality of random samplings on each of the image subblocks to obtain a plurality of sampled grayscale values; A matrix construction subunit, for establishing a frame prototype based on the pixel position of each sampled gray value in the image sub-block, rectangularizing the frame prototype to obtain a matrix frame, and inputting the sampled gray value of the corresponding image sub-block into the matrix frame to generate a gray matrix of the corresponding image sub-block; A matrix training unit, used to establish a plurality of transverse fitting curves and longitudinal fitting curves according to element differences between adjacent matrix elements in the grayscale matrix, perform depth training on curve points in each of the transverse fitting curves and each of the longitudinal fitting curves using the least squares method to obtain a depth fitting parameter corresponding to each matrix element, and train the corresponding matrix elements using the depth fitting parameters to obtain an ideal grayscale matrix corresponding to each of the image sub-blocks; The feature generation unit is used to analyze the grayscale difference between the actual grayscale value and the ideal grayscale value corresponding to each pixel point in the corresponding image sub-block based on the ideal grayscale matrix, and establish the grayscale deficiency feature of each image sub-block.

2. The ovarian cancer homologous recombination defect analysis system based on weakly supervised learning according to claim 1, characterized in that: Also includes: According to the patient informed consent agreement, the pathological sample of the corresponding patient is retrieved and sliced ​​to obtain the pathological slice of the patient; The pathological slice is transmitted to a preset shooting area for shooting to obtain a pathological slice image of the corresponding patient.

3. The ovarian cancer homologous recombination defect analysis system based on weakly supervised learning according to claim 1, characterized in that: The feature training module comprises: A feature analysis unit, used to establish a plurality of homologous recombination defect presentation samples according to the presentation characteristics of the homologous recombination defect, use the homologous recombination defect presentation samples to respectively capture the WSI information contained in each of the enhanced image sub-blocks, establish the WSI features corresponding to each of the enhanced image sub-blocks, and establish the overall WSI features of the pathological slice image; A feature training unit, configured to analyze a maximum lesion threshold of the pathological slice image according to the overall WSI feature, and to perform scaling training on a feature range corresponding to each WSI feature according to the maximum lesion threshold, so as to obtain a lesion area range corresponding to each enhanced image sub-block; A local analysis unit is used to mark a corresponding weak label for each of the enhanced image sub-blocks according to the WSI features, identify the lesion area in each pixel point in the corresponding enhanced image sub-block according to the weak label and lesion area range corresponding to the same enhanced image sub-block, and determine the lesion probability corresponding to each pixel point in the enhanced image sub-block, and establish local information corresponding to each of the enhanced image sub-blocks; A global construction unit is used to construct the global information of the pathological slice at a source position of the enhanced image sub-block in the pathological slice image according to the local information.

4. The ovarian cancer homologous recombination defect analysis system based on weakly supervised learning according to claim 1, characterized in that: The defect analysis module comprises: A scoring execution unit, configured to perform HRD scoring on each of the local information respectively, to obtain an HRD score value corresponding to each of the local information, and to establish a lesion area contour of a corresponding enhanced image sub-block according to each of the local information when the HRD score value is greater than or equal to 30; A lesion establishing unit, used for reorganizing different lesion contour areas to obtain a global lesion contour of the pathological slice image, obtaining a lesion image of the pathological slice image, acquiring a plurality of image key points in the lesion image, and establishing a plurality of lesion factors according to the image key points; The heat map creation unit is used to mark each of the pathological factors on the pathological slice image respectively to generate a lesion defect distribution heat map.

5. The ovarian cancer homologous recombination defect analysis system based on weakly supervised learning according to claim 1, characterized in that: The marking processing module comprises: A semantic processing unit, used to obtain the medical markers added by the pathologist in the lesion defect distribution heat map, and obtain semantic information corresponding to each of the medical markers; The heat map analysis unit is used to establish and display the patient's defect analysis report based on the lesion variable corresponding to each lesion factor in the lesion defect distribution heat map in combination with the corresponding semantic information.

6. The ovarian cancer homologous recombination defect analysis system based on weakly supervised learning according to claim 3, characterized in that: The global building unit comprises: An image recombining subunit, configured to establish a recombined pathological slice image according to the local information and the source position of the enhanced image subblock in the pathological slice image; A loss analysis subunit, used for calculating the combined loss weights between adjacent enhanced image sub-blocks in the reconstructed pathological slice image according to formula (1); （1） in, represents the combined loss weight of the reconstructed pathological slice image, m represents the number of enhanced image sub-blocks, Indicates the first label value corresponding to the weak label; The information adjustment subunit is used to establish the global information of the pathological slice image according to the combined loss weight and the local information corresponding to each enhanced image sub-block.

7. The ovarian cancer homologous recombination defect analysis system based on weakly supervised learning according to claim 3, characterized in that: The process of marking a corresponding weak label for each enhanced image sub-block according to the WSI feature comprises: Determine the sub-block importance level of the corresponding enhanced image sub-block in the pathological slice image according to the WSI feature, and match the corresponding enhanced image sub-block with a corresponding image definition label according to the sub-block importance level; Matching a corresponding weak label for each enhanced image sub-block according to formula (2); （2） Represented as the weak label matched by the i-th enhanced image sub-block, The second label value representing the image definition label corresponding to the i-th enhanced image sub-block, represents the lesion probability of the WSI feature corresponding to the i-th enhanced image sub-block, and m represents the HRD score value corresponding to the i-th enhanced image sub-block.

8. The ovarian cancer homologous recombination defect analysis system based on weakly supervised learning according to claim 7, characterized in that: Also includes: The enhanced image sub-blocks with consistent weak labels are clustered to generate image classes with the same label and displayed.

Images