Cervical lesion intercellular relation modeling and analysis system based on graph neural network

Abstract

The invention discloses a cervical lesion intercellular relation modeling and analysis system based on a graph neural network, and the system comprises a medical image collection module which is used for collecting a digital image of a cervical tissue pathological section or a cervical TCT slide; the cell detection and segmentation module is used for extracting spatial position information and morphological characteristics of cells; the cell feature extraction module is used for extracting and fusing the spatial position, morphology, texture and biological marker features of the cells; the cell relation graph construction module is used for constructing a heterogeneous cell relation graph with cells as nodes and inter-cell relations as edges; the graph neural network analysis module is used for carrying out feature learning and modeling on the heterogeneous cell relation graph; the intelligent auxiliary diagnosis module is used for generating auxiliary diagnosis suggestions; and the data management and automatic control module is used for realizing automatic control and case data management of the whole process of the data. The intelligent and automatic level of cervical lesion cell analysis can be comprehensively improved, and the accuracy and efficiency of diagnosis are improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of medical image analysis and bioinformatics, and in particular to a cervical lesion cell-cell relationship modeling and analysis system based on graph neural networks. Background Art

[0002] In recent years, with the increasing demand for screening for cervical cancer and its early lesions, the application of digital pathology image analysis in cervical pathology diagnosis has become increasingly widespread. Traditionally, the diagnosis of cervical lesions relies primarily on the observation and judgment of pathologists on slides. This method is labor-intensive, subjective, and highly dependent on the physician's expertise and experience, making it difficult to consistently guarantee diagnostic efficiency and accuracy. To improve automation, current research has gradually introduced medical image processing and deep learning technologies to achieve automated analysis of cervical tissue slides, significantly promoting the development of intelligent pathology images.

[0003] Currently, common automated cervical lesion analysis systems typically use convolutional neural networks to detect and classify single cells or local areas, and have achieved certain results in cell segmentation, morphological feature extraction, and partial lesion identification. However, these methods mostly use single cells as analysis units and lack systematic modeling of the spatial distribution and interrelationships between cells, making it difficult to fully reveal the pathological characteristics of cell clusters in tissue structures. In addition, traditional methods have limitations in feature fusion, usually focusing only on morphological or single-channel features, ignoring the synergistic effects of multidimensional features such as spatial position, texture, and molecular markers, thereby affecting the accuracy of abnormal area identification and risk grading.

[0004] On the other hand, existing systems are still imperfect in areas such as abnormal region aggregation, cell cluster identification, and spatial statistical analysis. In particular, it is difficult to deeply model complex intercellular relationships and abnormal communication patterns. For diseases such as cervical lesions, which are characterized by high spatial heterogeneity and significant cellular interactions, relying solely on single-point or local features is difficult to fully reflect their pathological changes. At the same time, some existing technologies are not well designed in terms of automated processes, data management, and the generation of visual diagnostic recommendations. The lack of a closed loop from data collection, processing, analysis, to the output of diagnostic recommendations affects the system's usability and promotional effectiveness in actual clinical applications.

[0005] Therefore, how to provide a cervical lesion cell relationship modeling and analysis system based on graph neural network is an urgent problem that needs to be solved by technical personnel in this field. Summary of the Invention

[0006] One purpose of the present invention is to propose a method for modeling and analyzing the relationship between cervical lesion cells based on graph neural networks. By constructing a heterogeneous cell relationship graph and building a dynamic multimodal graph neural network, the present invention can conduct intelligent analysis and auxiliary diagnosis of the spatial distribution of cervical lesion cells and their mutual relationships, which can significantly improve the accuracy and intelligence level of cervical lesion screening and pathological diagnosis.

[0007] According to an embodiment of the present invention, a system for modeling and analyzing the relationship between cervical lesion cells based on a graph neural network includes the following steps:

[0008] Medical image acquisition module, used to obtain multi-channel digital image data of pathological sections of cervical tissue;

[0009] Cell detection and segmentation module, used to detect and segment individual cell regions and extract the spatial location information and morphological characteristics of cells;

[0010] The cell feature extraction module is used to extract texture and biological marker features, and fuse the spatial location information and morphological features of cells to construct a multi-dimensional cell feature set;

[0011] The cell relationship graph construction module is used to construct a heterogeneous cell relationship graph with cells as nodes and inter-cell relationships as edges;

[0012] Graph neural network analysis module, used for feature learning and modeling of heterogeneous cell relationship graphs;

[0013] Intelligent auxiliary diagnosis module, used to complete cervical lesion risk classification, lesion area spatial annotation, cell clustering and abnormal area analysis, and generate auxiliary diagnosis suggestions;

[0014] The data management and automation control module is used to realize full-process automation control of data and case data management.

[0015] Optionally, modules can be connected using the following methods:

[0016] S1, collecting multi-channel digital images of cervical tissue pathological sections, and performing preprocessing and cutting operations to generate valid sub-images;

[0017] S2. The valid sub-image is passed through the Swin Transformer cell detection network to generate a set of candidate cell regions, and the deformable mask convolutional neural network is used to perform pixel-level segmentation on the candidate cell regions to generate a cell segmentation mask;

[0018] S3. Extract the spatial position, morphological features, texture features, and biological marker features of cells through cell segmentation masks, and construct a multidimensional feature set of cells;

[0019] S4. The multidimensional feature set of cells is used as node features, and bimodal association edges are constructed based on the spatial distance between cells and the similarity of feature vectors to generate a heterogeneous cell relationship graph;

[0020] S5. Perform feature learning on heterogeneous cell relationship graphs through dynamic multimodal graph neural networks to generate high-order features and clustering information of cell nodes;

[0021] S6. Perform risk classification, spatial aggregation, and visual annotation on the high-order features and clustering information of cell nodes, and generate auxiliary diagnosis suggestions.

[0022] Optionally, the S1 specifically includes:

[0023] S11. Use a digital pathology scanner to perform high-resolution digital acquisition of cervical tissue pathology sections to obtain original pathology images I. raw ;

[0024] S12, original pathological image I raw Perform color normalization, denoising and contrast enhancement to obtain the preprocessed image I prep ;

[0025] S13, pre-processed image I prep Use sliding window method to cut and divide into N sub-images I i , the sub-image contains biomarker channel information;

[0026] S14. Perform quality assessment on each sub-image, and only retain valid sub-images that meet preset quality standards.

[0027] Optionally, the S2 specifically includes:

[0028] S21, the effective sub-image I i Input into the cell detection network based on Swin Transformer to generate a set of candidate cell regions R i ={r i1 ,r i2 ,…,r im}, where r ij represents the jth candidate cell region in the i-th sub-image, and m is the number of candidate cell regions;

[0029] The cell detection network uses Swin-T as the backbone network, utilizes a sliding window multi-head self-attention mechanism to extract multi-scale cell features, embeds a dynamic deformable convolution module in the feature pyramid output layer, adaptively adjusts the convolution kernel weights according to cell morphology, and optimizes the detection results through a geometric constraint loss function.

[0030] The geometric constraint loss function includes classification loss and Hausdorff distance loss, wherein the classification loss uses the Focal Loss function to calculate the difference between the candidate cell region classification result and the true annotation label; the Hausdorff distance loss calculates the spatial distribution of the candidate region center point and the true cell center point;

[0031] S22, using deformable mask convolutional neural network to map candidate cell regions r ij Perform pixel-level segmentation to generate cell segmentation mask M ij , where M ij (x,y)∈{0,1},M ij (x,y) indicates whether the pixel (x,y) belongs to the segmented cell area;

[0032] The deformable mask convolutional neural network replaces the standard convolutional layer with a deformable convolutional layer in the mask prediction branch, dynamically learns the convolution kernel offset according to the cell boundary shape, and introduces a boundary-aware loss function to jointly optimize the segmentation results;

[0033] The boundary-aware loss function includes Dice loss and weighted cross entropy loss, where Dice loss calculates the similarity between the predicted segmentation result and the true segmentation label, and weighted cross entropy loss gives higher weights to pixels in the segmentation edge area;

[0034] S23, using cell segmentation mask M ij Extract the spatial position of the cell (x ij ,y ij ) and morphological characteristics F ij , the morphological characteristics include area, perimeter and roundness;

[0035] S24, output the spatial position of each cell (x ij ,y ij ) and morphological characteristics F ij .

[0036] Optionally, the S3 specifically includes:

[0037] S31, based on cell segmentation mask M ij And the pixel values ​​in the corresponding cell area, the texture features T of the cell area are calculated using the gray level co-occurrence matrix method ij , the texture features include energy, contrast and entropy;

[0038] S32. Extracting biological marker features of cell regions based on multi-channel image information B ij Specifically, the intensity mean of the biomarker channel at each pixel in the cell area is calculated using the following formula:

[0039]

[0040] Among them, A ij represents the pixel (x, y) set corresponding to the jth cell region in the i-th sub-image, I bio (x,y) represents the intensity value of the biomarker channel at pixel (x,y);

[0041] S33, the spatial position of the cell region (x ij ,y ij ), morphological characteristics F ij , texture feature T ij and biomarker signature B ij Fusion is performed to construct the cell multidimensional feature vector V ij =[x ij ,y ij ,F ij ,T ij ,B ij ];

[0042] S34, all the cell multidimensional feature vectors V ij The multidimensional feature set of cells V = {V ij}.

[0043] Optionally, the S4 specifically includes:

[0044] S41, the cell multidimensional feature set V = {V ij} as node features, initialize the heterogeneous cell relationship graph G = (N, E) containing nodes and edges, where N is the node set, E is the edge set, and each node n ij ∈N corresponds to a cell and its feature vector V ij ;

[0045] S42, constructing a bimodal association edge according to the spatial coordinates and feature vector similarity of each cell, wherein the bimodal association edge includes a spatial association edge and a feature similarity edge;

[0046] For any two nodes n ij and n kl , calculate the Euclidean distance d based on the spatial coordinates (ij)(kl) , if the Euclidean distance d (ij)(kl) Less than or equal to the set distance threshold d th , then establish a spatial correlation edge;

[0047] For any two nodes n ij and n kl , if the feature vector similarity ‖V ij -V kl ‖ is less than or equal to the set similarity threshold δ v, then a feature association edge is established, where ‖·‖ represents the L2 norm;

[0048] S43. For node n with established edge ij and n kl , calculate the edge weight w through the composite weight function (ij)(kl) :

[0049]

[0050] Among them, α (ij)(kl) represents the self-attention coefficient, exp represents the exponential function with the natural constant e as the base, σ d Represents the normalization coefficient of Euclidean distance, σ v Represents the normalization coefficient of the multidimensional feature vector;

[0051] S44. Integrate all nodes N and spatially associated edges E s , feature-related edge E f and edge weights W to generate a weighted heterogeneous cell relationship graph G = (N,E s ∪E f ,W).

[0052] Optionally, the S5 specifically includes:

[0053] S51, the heterogeneous cell relationship diagram G = (N,E s ∪E f ,W) is input into a dynamic multimodal graph neural network, which updates the edge weights in real time through a learnable Softmax function, and dynamically fuses spatial position, morphological features, texture features, and biological marker features to generate modality gating weights;

[0054] S52. Perform heterogeneous message passing through three layers of graph convolution operations with integrated multi-head self-attention mechanisms. In the message generation phase, the message function is customized based on spatial proximity or feature similarity. In the aggregation phase, the multi-head self-attention mechanism is used to weighted sum the features of neighboring nodes.

[0055] S53, using the spectral clustering algorithm to obtain the cluster labels and cluster centers of the cell nodes based on the high-order features, and using the local outlier factor method to calculate the abnormal probability score;

[0056] S54 , outputting a pathological feature recognition result, wherein the pathological feature recognition result includes a cell distribution topology map, a clustering structure, and a spatial heat map of abnormal cell areas.

[0057] Optionally, the S6 specifically includes:

[0058] S61, assign a lesion risk level label L to each cell node based on the high-order features, cluster labels, abnormal probability scores and pathological feature recognition results of each cell node ij ∈{0,1,2,…,K}, K is the number of risk level categories;

[0059] S62, label the risk level L ij Cell nodes exceeding the set threshold are aggregated using a density estimation algorithm to obtain the spatial region of abnormal cells and the area of ​​abnormal regions;

[0060] S63, in the original image I prep Visually mark abnormal cells, draw the boundaries of abnormal areas and label risk levels;

[0061] S64, inputting the high-order features of the cell nodes, cluster labels, cluster centers, and abnormal area into the downstream hybrid expert neural network for analysis, and generating auxiliary diagnosis suggestions, the suggestions including description of the lesion extent, spatial location of the abnormal area, and cell cluster characteristics;

[0062] The hybrid expert neural network introduces a multi-channel feature adaptation mechanism and a dynamic expert gating module, implements expert branches for feature information from different sources, conducts targeted feature learning and weighted integration, and can adaptively adjust the output results of each expert branch according to the distribution of input features.

[0063] Optionally, the data management and automation control module calls the medical image acquisition, cell detection and segmentation, feature extraction, relationship graph construction, graph neural network analysis and auxiliary diagnosis display modules in sequence according to the preset process sequence through the task scheduler, performs data verification on the processing results of each step, stores case-related data by category and module, and indexes and associates them through unique identifiers, while recording the operation log of each analysis.

[0064] The beneficial effects of the present invention are as follows: by constructing a heterogeneous cell relationship graph that integrates multiple types of nodes and multiple types of edges, integrating the spatial position, morphological characteristics, texture characteristics and biological marker characteristics of cells, the joint expression of multi-dimensional information of cells is achieved, and the complex relationship between cervical lesion cells is deeply learned and dynamically modeled through a dynamic multimodal graph neural network. The system can not only automatically complete the lesion risk level classification, abnormal cell area aggregation, cell clustering and cluster feature analysis, but also rely on downstream auxiliary decision models to generate personalized auxiliary diagnosis suggestions. This system improves the accuracy of pathological feature recognition and abnormal cell detection capabilities, and can realize the visualization of results and automated data management, thereby improving the intelligence level of pathological diagnosis and the efficiency of clinical application. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] 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:

[0066] Figure 1 This is a flowchart of a method for modeling and analyzing the relationship between cells of cervical lesions based on a graph neural network proposed by the present invention;

[0067] Figure 2 It is a flow chart of cell detection, segmentation and feature extraction in the present invention;

[0068] Figure 3 This is a flow chart for constructing a heterogeneous cell relationship diagram in the present invention;

[0069] Figure 4 This is a flow chart of intelligent analysis of cervical lesion cells in the present invention. DETAILED DESCRIPTION

[0070] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0071] refer to Figure 1-4 A graph neural network-based modeling and analysis system for the relationship between cervical lesion cells includes the following steps:

[0072] Medical image acquisition module, used to obtain multi-channel digital image data of pathological sections of cervical tissue;

[0073] Cell detection and segmentation module, used to detect and segment individual cell regions and extract the spatial location information and morphological characteristics of cells;

[0074] The cell feature extraction module is used to extract texture and biological marker features, and fuse the spatial location information and morphological features of cells to construct a multi-dimensional cell feature set;

[0075] The cell relationship graph construction module is used to construct a heterogeneous cell relationship graph with cells as nodes and inter-cell relationships as edges;

[0076] Graph neural network analysis module, used for feature learning and modeling of heterogeneous cell relationship graphs;

[0077] Intelligent auxiliary diagnosis module, used to complete cervical lesion risk classification, lesion area spatial annotation, cell clustering and abnormal area analysis, and generate auxiliary diagnosis suggestions;

[0078] The data management and automation control module is used to realize full-process automation control of data and case data management.

[0079] In this embodiment, the modules are connected through the following methods:

[0080] S1, collecting multi-channel digital images of cervical tissue pathological sections, and performing preprocessing and cutting operations to generate valid sub-images;

[0081] S2. The valid sub-image is passed through the Swin Transformer cell detection network to generate a set of candidate cell regions, and the deformable mask convolutional neural network is used to perform pixel-level segmentation on the candidate cell regions to generate a cell segmentation mask;

[0082] S3. Extract the spatial position, morphological features, texture features, and biological marker features of cells through cell segmentation masks, and construct a multidimensional feature set of cells;

[0083] S4. The multidimensional feature set of cells is used as node features, and bimodal association edges are constructed based on the spatial distance between cells and the similarity of feature vectors to generate a heterogeneous cell relationship graph;

[0084] S5. Perform feature learning on heterogeneous cell relationship graphs through dynamic multimodal graph neural networks to generate high-order features and clustering information of cell nodes;

[0085] S6. Perform risk classification, spatial aggregation, and visual annotation on the high-order features and clustering information of cell nodes, and generate auxiliary diagnosis suggestions.

[0086] In this embodiment, S1 specifically includes:

[0087] S11. Use a digital pathology scanner to perform high-resolution digital acquisition of cervical tissue pathology sections to obtain original pathology images I. raw ;

[0088] S12, original pathological image I raw Perform color normalization, denoising and contrast enhancement to obtain the preprocessed image I prep ;

[0089] S13, pre-processed image I prep Use sliding window method to cut and divide into N sub-images I i , the sub-image contains biomarker channel information;

[0090] S14. Perform quality assessment on each sub-image, and only retain valid sub-images that meet preset quality standards.

[0091] In this embodiment, S2 specifically includes:

[0092] S21, the effective sub-image I iInput into the cell detection network based on Swin Transformer to generate a set of candidate cell regions R i ={r i1 ,r i2 ,…,r im}, where r ij represents the jth candidate cell region in the i-th sub-image, and m is the number of candidate cell regions;

[0093] The cell detection network uses Swin-T as the backbone network, utilizes a sliding window multi-head self-attention mechanism to extract multi-scale cell features, embeds a dynamic deformable convolution module in the feature pyramid output layer, adaptively adjusts the convolution kernel weights according to cell morphology, and optimizes the detection results through a geometric constraint loss function.

[0094] The geometric constraint loss function includes classification loss and Hausdorff distance loss, wherein the classification loss uses the Focal Loss function to calculate the difference between the candidate cell region classification result and the true annotation label; the Hausdorff distance loss calculates the spatial distribution of the candidate region center point and the true cell center point;

[0095] S22, using deformable mask convolutional neural network to map candidate cell regions r ij Perform pixel-level segmentation to generate cell segmentation mask M ij , where M ij (x,y)∈{0,1},M ij (x,y) indicates whether the pixel (x,y) belongs to the segmented cell area;

[0096] The deformable mask convolutional neural network replaces the standard convolutional layer with a deformable convolutional layer in the mask prediction branch, dynamically learns the convolution kernel offset according to the cell boundary shape, and introduces a boundary-aware loss function to jointly optimize the segmentation results;

[0097] The boundary-aware loss function includes Dice loss and weighted cross entropy loss, where Dice loss calculates the similarity between the predicted segmentation result and the true segmentation label, and weighted cross entropy loss gives higher weights to pixels in the segmentation edge area;

[0098] S23, using cell segmentation mask M ij Extract the spatial position of the cell (x ij ,y ij ) and morphological characteristics F ij , the morphological characteristics include area, perimeter and roundness;

[0099] S24, output the spatial position of each cell (x ij ,y ij ) and morphological characteristics Fij .

[0100] In this embodiment, S3 specifically includes:

[0101] S31, based on cell segmentation mask M ij And the pixel values ​​in the corresponding cell area, the texture features T of the cell area are calculated using the gray level co-occurrence matrix method ij , the texture features include energy, contrast and entropy;

[0102] S32. Extracting biological marker features of cell regions based on multi-channel image information B ij Specifically, the intensity mean of the biomarker channel at each pixel in the cell area is calculated using the following formula:

[0103]

[0104] Among them, A ij represents the pixel (x, y) set corresponding to the jth cell region in the i-th sub-image, I bio (x,y) represents the intensity value of the biomarker channel at pixel (x,y);

[0105] S33, the spatial position of the cell region (x ij ,y ij ), morphological characteristics F ij , texture feature T ij and biomarker signature B ij Fusion is performed to construct the cell multidimensional feature vector V ij =[x ij ,y ij ,F ij ,T ij ,B ij ];

[0106] S34, all the cell multidimensional feature vectors V ij The multidimensional feature set of cells V = {V ij}.

[0107] In this embodiment, the S4 specifically includes:

[0108] S41, the cell multidimensional feature set V = {V ij} as node features, initialize the heterogeneous cell relationship graph G = (N, E) containing nodes and edges, where N is the node set, E is the edge set, and each node n ij ∈N corresponds to a cell and its feature vector V ij ;

[0109] S42, constructing a bimodal association edge according to the spatial coordinates and feature vector similarity of each cell, wherein the bimodal association edge includes a spatial association edge and a feature similarity edge;

[0110] For any two nodes n ij and n kl , calculate the Euclidean distance d based on the spatial coordinates (ij)(kl) , if the Euclidean distance d (ij)(kl) Less than or equal to the set distance threshold d th , then establish a spatial correlation edge;

[0111] For any two nodes n ij and n kl , if the feature vector similarity ‖V ij -V kl ‖ is less than or equal to the set similarity threshold δ v , then a feature association edge is established, where ‖·‖ represents the L2 norm;

[0112] S43. For node n with established edge ij and n kl , calculate the edge weight w through the composite weight function (ij)(kl) :

[0113]

[0114] Among them, α (ij)(kl) represents the self-attention coefficient, exp represents the exponential function with the natural constant e as the base, σ d Represents the normalization coefficient of Euclidean distance, σ v Represents the normalization coefficient of the multidimensional feature vector;

[0115] S44. Integrate all nodes N and spatially associated edges E s , feature-related edge E f and edge weights W to generate a weighted heterogeneous cell relationship graph G = (N,E s ∪E f ,W).

[0116] In this embodiment, the S5 specifically includes:

[0117] S51, the heterogeneous cell relationship diagram G = (N,E s ∪E f ,W) is input into a dynamic multimodal graph neural network, which updates the edge weights in real time through a learnable Softmax function, and dynamically fuses spatial position, morphological features, texture features, and biological marker features to generate modality gating weights;

[0118] S52. Perform heterogeneous message passing through three layers of graph convolution operations with integrated multi-head self-attention mechanisms. In the message generation phase, the message function is customized based on spatial proximity or feature similarity. In the aggregation phase, the multi-head self-attention mechanism is used to weighted sum the features of neighboring nodes.

[0119] S53, using the spectral clustering algorithm to obtain the cluster labels and cluster centers of the cell nodes based on the high-order features, and using the local outlier factor method to calculate the abnormal probability score;

[0120] S54 , outputting a pathological feature recognition result, wherein the pathological feature recognition result includes a cell distribution topology map, a clustering structure, and a spatial heat map of abnormal cell areas.

[0121] In this embodiment, S6 specifically includes:

[0122] S61, assign a lesion risk level label L to each cell node based on the high-order features, cluster labels, abnormal probability scores and pathological feature recognition results of each cell node ij ∈{0,1,2,…,K}, K is the number of risk level categories;

[0123] S62, label the risk level L ij Cell nodes exceeding the set threshold are aggregated using a density estimation algorithm to obtain the spatial region of abnormal cells and the area of ​​abnormal regions;

[0124] S63, in the original image I prep Visually mark abnormal cells, draw the boundaries of abnormal areas and label risk levels;

[0125] S64, inputting the high-order features of the cell nodes, cluster labels, cluster centers, and abnormal area into the downstream hybrid expert neural network for analysis, and generating auxiliary diagnosis suggestions, the suggestions including description of the lesion extent, spatial location of the abnormal area, and cell cluster characteristics;

[0126] The hybrid expert neural network introduces a multi-channel feature adaptation mechanism and a dynamic expert gating module, implements expert branches for feature information from different sources, conducts targeted feature learning and weighted integration, and can adaptively adjust the output results of each expert branch according to the distribution of input features.

[0127] In this embodiment, the data management and automation control module calls the medical image acquisition, cell detection and segmentation, feature extraction, relationship graph construction, graph neural network analysis and auxiliary diagnosis display modules in sequence according to the preset process sequence through the task scheduler, performs data verification on the processing results of each step, stores case-related data by category and module, and indexes and associates them through unique identifiers, while recording the operation log of each analysis.

[0128] Example 1:

[0129] In order to verify the feasibility of the present invention in implementation, the present invention was applied to the cervical cancer screening project of the pathology department of a tertiary hospital. From January 2024 to December 2024, the hospital collected a total of 1,800 cervical tissue pathology sections from physical examinations and outpatients in the hospital and surrounding areas. The samples covered a variety of different pathological types, including normal, cervical intraepithelial neoplasia (CIN1, CIN2, CIN3) and cervical cancer. Due to the large number of cases and the complexity of pathological types, this project belongs to a large-scale and high-load practical application scenario, which puts higher requirements on the processing power, stability and operating efficiency of the automated analysis system. All slices are digitally acquired at high resolution using a multi-channel digital pathology scanner to obtain original digital images containing biomarker channel information. The hospital pathology information system assigns a unique case number to each patient and records the patient's basic information, pathological diagnosis results and follow-up data in detail.

[0130] In practical applications, the system of the present invention first automatically receives the digitized slice image, performs color normalization, denoising and contrast enhancement on the original image, and then uses a sliding window method to cut and obtain high-quality sub-images. The system can automatically perform cell detection and segmentation on each sub-image, and accurately extract the spatial position and morphological characteristics of individual cells, such as area, perimeter and roundness. Based on the cell segmentation mask, the system further extracts the texture features of the cells, including energy, contrast and entropy, and extracts biological marker features at the same time. After all cell features are fused, a multi-dimensional feature set of cells is automatically generated. Based on the spatial distance and feature vector differences between cells, the system constructs a relationship graph with cells as nodes, and uses a heterogeneous multi-headed graph neural network for feature learning and deep modeling, thereby achieving the classification of diseased cells, spatial annotation of abnormal areas and generation of auxiliary diagnosis suggestions.

[0131] In order to objectively evaluate the performance of the system of the present invention, manual diagnosis of the same period was selected as a control. From the dimensions of different pathological types, different tissue regions and different doctor teams, the cell detection accuracy, cell segmentation F1 value, lesion grading accuracy, abnormal area detection recall rate, diagnostic consistency Kappa value, average single case analysis time, automatic report completeness rate and case batch export efficiency indicators were statistically analyzed. Among them, the cell segmentation F1 value is used to measure the score of the system of the present invention and manual diagnosis in cell segmentation accuracy. The higher the value, the closer the segmentation effect is to the standard answer marked by manual annotation; the abnormal area detection recall rate is a measure of the system's ability to identify all real abnormal areas when detecting abnormal areas in cervical lesion sections, such as suspicious lesions and diseased cell clusters; the diagnostic consistency Kappa value refers to the consistency between the system of the present invention and manual diagnosis in cervical lesion grading or abnormal area judgment. The higher the Kappa value, the closer the judgment results between the system and manual diagnosis, and the stronger the clinical usability and reliability of the system. Table 1 shows the comparative evaluation results of the system of the present invention and manual diagnosis.

[0132] Table 1 Comparative evaluation results between the system of the present invention and manual diagnosis

[0133]

[0134] As can be seen from the above table, the system of the present invention is superior to manual diagnosis in terms of cell detection accuracy, cell segmentation F1 value, lesion grading accuracy and abnormal area detection recall rate. In this scenario, the average single-case analysis time of the system of the present invention is only 38.6 seconds, which is much lower than the 275.4 seconds of manual diagnosis, greatly improving the diagnostic efficiency. The completeness rate of automatically generated auxiliary diagnosis reports is as high as 99.2%. The system also supports efficient batch export and long-term management of case data, which greatly improves the convenience of hospital case archiving and follow-up. In addition, the diagnostic results output by the system are compared with the gold standard diagnosis of multiple director-level pathologists. The diagnostic consistency Kappa value is as high as 0.927, which further demonstrates the reliability and practical value of the system in clinical applications. Under different pathological classifications, image difficulty and doctor team conditions, the system has shown good robustness and adaptability, which can effectively reduce human errors and improve the accuracy of pathological grading and abnormal area identification.

[0135] During actual application, hospital pathology staff reported that the automated process and visual interface of the system significantly reduced repetitive labor and improved the identification rate of difficult cases. The system supports long-term management, flexible querying, and batch export of case data, meeting the needs of clinical research and management. In summary, the system of the present invention significantly improves the accuracy and efficiency of cervical lesion pathology diagnosis, providing strong technical support for large-scale screening and early diagnosis.

[0136] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A graph neural network-based modeling and analysis system for the relationship between cervical lesion cells, characterized by: include: Medical image acquisition module, used to obtain multi-channel digital image data of cervical tissue pathology sections or cervical TCT films; Cell detection and segmentation module, used to detect and segment individual cell regions and extract the spatial location information and morphological characteristics of cells; The cell feature extraction module is used to extract texture and biological marker features, and fuse the spatial location information and morphological features of cells to construct a multi-dimensional cell feature set; The cell relationship graph construction module is used to construct a heterogeneous cell relationship graph with cells as nodes and inter-cell relationships as edges; Graph neural network analysis module, used for feature learning and modeling of heterogeneous cell relationship graphs; Intelligent auxiliary diagnosis module, used to complete cervical lesion risk classification, lesion area spatial annotation, cell clustering and abnormal area analysis, and generate auxiliary diagnosis suggestions; The data management and automation control module is used to realize full-process automation control of data and case data management.

2. The cervical lesion cell relationship modeling and analysis system based on graph neural network according to claim 1 is characterized in that: The modules are implemented as follows: S1, collecting multi-channel digital images of cervical tissue pathology sections or cervical TCT slides, and performing preprocessing and cutting operations to generate valid sub-images; S2. The valid sub-image is passed through the Swin Transformer cell detection network to generate a set of candidate cell regions, and the deformable mask convolutional neural network is used to perform pixel-level segmentation on the candidate cell regions to generate a cell segmentation mask; S3. Extract the spatial position, morphological features, texture features, and biological marker features of cells through cell segmentation masks, and construct a multidimensional feature set of cells; S4, using the cell multidimensional feature set as the node feature, adaptively constructing a heterogeneous cell relationship graph according to the type of input medical image; S5. Perform feature learning on heterogeneous cell relationship graphs through dynamic multimodal graph neural networks to generate high-order features and clustering information of cell nodes; S6. Perform risk classification, spatial aggregation, and visual annotation on the high-order features and clustering information of cell nodes, and generate auxiliary diagnosis suggestions.

3. The graph neural network-based cervical lesion cell relationship modeling and analysis system according to claim 2, characterized in that: Said S1 specifically includes: S11, using a digital pathology scanner to perform high-resolution digital acquisition of cervical tissue pathology sections or cervical TCT preparations to obtain original pathology images I raw ; S12, original pathological image I raw Perform color normalization, denoising and contrast enhancement to obtain the preprocessed image I prep ; S13, pre-processed image I prep Use sliding window method to cut and divide into N sub-images I i , the sub-image contains biomarker channel information; S14. Perform quality assessment on each sub-image, and only retain valid sub-images that meet preset quality standards.

4. The graph neural network-based cervical lesion cell relationship modeling and analysis system according to claim 2, characterized in that: The S2 specifically includes: S21, the effective sub-image I i Input into the cell detection network based on Swin Transformer to generate a set of candidate cell regions R i ={r i1 ,r i2 ,…,r im }, where r ij represents the jth candidate cell region in the i-th sub-image, and m is the number of candidate cell regions; The cell detection network uses Swin-T as the backbone network, utilizes a sliding window multi-head self-attention mechanism to extract multi-scale cell features, embeds a dynamic deformable convolution module in the feature pyramid output layer, adaptively adjusts the convolution kernel weights according to cell morphology, and optimizes the detection results through a geometric constraint loss function. The geometric constraint loss function includes classification loss and Hausdorff distance loss, wherein the classification loss uses the FocalLoss function to calculate the difference between the candidate cell region classification result and the true annotation label; the Hausdorff distance loss calculates the spatial distribution of the candidate region center point and the true cell center point; S22, using deformable mask convolutional neural network to map candidate cell regions r ij Perform pixel-level segmentation to generate cell segmentation mask M ij , where M ij (x,y)∈{0,1},M ij (x,y) indicates whether the pixel (x,y) belongs to the segmented cell area; The deformable mask convolutional neural network replaces the standard convolutional layer with a deformable convolutional layer in the mask prediction branch, dynamically learns the convolution kernel offset according to the cell boundary shape, and introduces a boundary-aware loss function to jointly optimize the segmentation results; The boundary-aware loss function includes Dice loss and weighted cross entropy loss, where Dice loss calculates the similarity between the predicted segmentation result and the true segmentation label, and weighted cross entropy loss gives higher weights to pixels in the segmentation edge area.

5. The cervical lesion cell relationship modeling and analysis system based on graph neural network according to claim 2, characterized in that: The S3 specifically includes: S31. Using cell segmentation mask M ij Extract the spatial position of the cell (x ij ,y ij ) and morphological characteristics F ij , the morphological characteristics include area, perimeter and roundness; S32, based on cell segmentation mask M ij And the pixel values ​​in the corresponding cell area, the texture features T of the cell area are calculated using the gray level co-occurrence matrix method ij , the texture features include energy, contrast and entropy; S33. Extract biological marker features of cell regions based on multi-channel image information B ij , specifically, calculating the mean intensity of the biomarker channel at each pixel in the cell area; S34, the spatial position of the cell region (x ij ,y ij ), morphological characteristics F ij , texture feature T ij and biomarker signature B ij Fusion is performed to construct the cell multidimensional feature vector V ij =[x ij ,y ij ,F ij ,T ij ,B ij ]; S35, transform all the cell multidimensional feature vectors V ij The multidimensional feature set of cells V = {V ij }.

6. The graph neural network-based cervical lesion cell relationship modeling and analysis system according to claim 2, characterized in that: The S4 specifically includes: S41, the cell multidimensional feature set V = {V ij } as node features, initialize the heterogeneous cell relationship graph G = (N, E) containing nodes and edges, where N is the node set, E is the edge set, and each node n ij ∈N corresponds to a cell and its feature vector V ij ; S42, adaptively selecting a method for constructing inter-cell relationship edges according to the type of collected medical images; When the input is a cervical tissue pathology section image, a bimodal correlation edge is constructed according to the spatial coordinates of each cell and the similarity of the feature vector, wherein the bimodal correlation edge includes a spatial correlation edge and a feature similarity edge; For any two nodes n ij and n kl , calculate the Euclidean distance d based on the spatial coordinates (ij)(kl) , if the Euclidean distance d (ij)(kl) Less than or equal to the set distance threshold d th , then establish a spatial correlation edge; For any two nodes n ij and n kl , if the feature vector similarity ‖V ij -V kl ‖ is less than or equal to the set similarity threshold δ v , then a feature association edge is established, where ‖·‖ represents the L2 norm; When the input is a cervical TCT slide image, a statistical distribution or density estimation method is used to construct spatial correlation edges based on the distribution density, proximity relationship or feature similarity of cells in the slide. The spatial correlation edges represent the cell aggregation relationship in the slide field of view. S43. For node n with established edge ij and n kl , calculate the edge weight w through the composite weight function (ij)(kl) : Among them, α (ij)(kl) represents the self-attention coefficient, exp represents the exponential function with the natural constant e as the base, σ d Represents the normalization coefficient of Euclidean distance, σ v Represents the normalization coefficient of the multidimensional feature vector; S44. Integrate all nodes N and spatially associated edges E s , feature-related edge E f and edge weights W to generate a weighted heterogeneous cell relationship graph G = (N,E s ∪E f ,W).

7. The cervical lesion cell relationship modeling and analysis system based on graph neural network according to claim 2, characterized in that: The S5 specifically includes: S51, the heterogeneous cell relationship diagram G = (N,E s ∪E f ,W) is input into a dynamic multimodal graph neural network, which updates the edge weights in real time through a learnable Softmax function, and dynamically fuses spatial position, morphological features, texture features, and biological marker features to generate modality gating weights; S52. Perform heterogeneous message passing through three layers of graph convolution operations with integrated multi-head self-attention mechanisms. In the message generation phase, the message function is customized based on spatial proximity or feature similarity. In the aggregation phase, the multi-head self-attention mechanism is used to weighted sum the features of neighboring nodes. S53, using the spectral clustering algorithm to obtain the cluster labels and cluster centers of the cell nodes based on the high-order features, and using the local outlier factor method to calculate the abnormal probability score; S54 , outputting a pathological feature recognition result, wherein the pathological feature recognition result includes a cell distribution topology map, a clustering structure, and a spatial heat map of abnormal cell areas.

8. The graph neural network-based cervical lesion cell relationship modeling and analysis system according to claim 2, characterized in that: The S6 specifically includes: S61, assign a lesion risk level label L to each cell node based on the high-order features, cluster labels, abnormal probability scores and pathological feature recognition results of each cell node ij ∈{0,1,2,…,K}, K is the number of risk level categories; S62, label the risk level L ij Cell nodes exceeding the set threshold are aggregated using a density estimation algorithm to obtain the spatial region of abnormal cells and the area of ​​abnormal regions; S63, in the original image I prep Visually mark abnormal cells, draw the boundaries of abnormal areas and label risk levels; S64. The high-order features, cluster labels, cluster centers, and abnormal area of ​​the cell nodes are input into the downstream hybrid expert neural network for analysis, and auxiliary diagnosis suggestions are generated. The suggestions include a description of the degree of lesions, the spatial location of the abnormal area, and the characteristics of the cell cluster.

9. The cervical lesion cell relationship modeling and analysis system based on graph neural network according to claim 1, characterized in that: The data management and automation control module calls the medical image acquisition, cell detection and segmentation, feature extraction, relationship graph construction, graph neural network analysis and auxiliary diagnosis display modules in sequence according to the preset process sequence through the task scheduler, verifies the data of the processing results of each step, stores case-related data by category and module, and indexes and associates them through unique identifiers, while recording the operation log of each analysis.

Images