A Hypergraph Prognostic Prediction Method and System Based on Spatial Information of Pathological Images
By constructing a hypergraph prediction method based on pathological images, the problems of neglecting spatial context and high-cost acquisition of prior knowledge in tumor prognosis prediction are solved, and survival risk prediction with high accuracy and high interpretability is achieved.
Patent Information
- Application Number
- CN202611056480.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-16
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies neglect the spatial contextual relationships between tissue regions in tumor prognosis prediction, resulting in high costs and poor clinical interpretability in acquiring medically meaningful spatial prior knowledge.
By acquiring whole-slice pathological images, performing image preprocessing and extracting diagnostic prior spatial information, a hypergraph based on spatial prior is constructed. A two-stage spatial hypergraph convolution is then performed in conjunction with the initial node feature matrix. Distributed regional risk mapping is then performed using a multilayer perceptron to achieve survival risk score prediction.
It improved the accuracy of survival risk prediction, achieved highly interpretable pathological spatial validation, reduced the cost of acquiring spatial prior knowledge, and enhanced the clinical interpretability and generalization performance of the model.
Smart Images

Figure CN123120213A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing technology, and in particular to a hypergraph prognostic prediction method and system based on spatial information of pathological images. Background Technology
[0002] Current computational pathology methods primarily employ multi-instance learning frameworks for tumor prognosis prediction. The standard procedure typically involves first segmenting a whole-slice image into tens of thousands of non-overlapping small patches, then using a feature extraction network to extract features from each patch, and finally aggregating all patch features into a slide-level feature representation using attention mechanisms or pooling layers for final survival risk prediction. Current research also explores using graph neural networks to model spatial neighborhood relationships between patches; however, these techniques suffer from the following limitations: 1) Ignoring spatial context between tissue regions: Traditional MIL methods, based on set operations, severely neglect the spatial context between tissue regions, leading to a significant loss of local morphological and global architectural information. However, accurate modeling of the tumor microenvironment (TME) requires a comprehensive analysis of cell morphology and tissue spatial topology.
[0003] 2) Acquiring medically meaningful spatial prior knowledge is extremely costly: Although existing graph computation methods attempt to consider the spatial heterogeneity and architecture of WSI, acquiring such detailed prior knowledge (e.g., the spatial distribution of tissue types) requires expensive and time-consuming manual annotation at the regional level by pathologists. Existing methods have failed to utilize low-cost diagnostic tasks to extract spatial distribution patterns of tumors.
[0004] 3) Poor clinical interpretability: Although existing models can generate attention heatmaps, pathologists find it difficult to intuitively judge whether the areas the model focuses on are consistent with common medical knowledge (i.e., whether the model really focuses on the tumor area to judge the prognosis), and there is a lack of direct comparison and verification methods between diagnostic attention and prognostic attention. Summary of the Invention
[0005] To address the aforementioned technical problems, the present invention aims to provide a hypergraph prognostic prediction method and system based on the spatial information of pathological images, which can improve the accuracy of survival risk prediction while achieving highly interpretable pathological spatial verification.
[0006] The first technical solution adopted in this invention is: a hypergraph prognostic prediction method based on spatial information of pathological images, comprising the following steps: Full-section pathological images were acquired and image preprocessing and diagnostic prior spatial information extraction were performed to obtain an initial node feature matrix and confidence values of the plots predicted as tumor tissue. An initialization constraint mechanism is introduced to divide the confidence value of the map patch being predicted as tumor tissue into several predefined intervals and set overlap between adjacent intervals to construct a hypergraph based on spatial prior. Then, a two-stage spatial hypergraph convolution is performed in combination with the initial node feature matrix to obtain the final hyperedge features. Based on the multilayer perceptron, distributed regional risk mapping is performed. Based on the final hyperedge features, survival risk scores are predicted for the heterogeneous spatial regions corresponding to each hyperedge in the hypergraph. The regional risk vector is obtained, and its arithmetic mean is calculated as the final survival risk, thus realizing hypergraph prognosis prediction.
[0007] Furthermore, the step of acquiring whole-slice pathological images and performing image preprocessing and diagnostic prior spatial information extraction to obtain an initial node feature matrix and a confidence value for predicting a patch as tumor tissue specifically includes: Full-section pathological images were acquired and foreground region segmentation was performed to obtain several pathological image blocks. Based on a pre-trained pathology model, initial feature extraction is performed on several pathological image blocks to construct an initial node feature matrix. Several pathological image patches are input into an attention-based binary classification diagnostic model to calculate attention scores, thereby obtaining the confidence value that the patch is predicted to be tumor tissue.
[0008] Furthermore, the step of introducing an initialization constraint mechanism, dividing the confidence value of a patch being predicted as tumor tissue into several predefined intervals, setting overlap between adjacent intervals to construct a hypergraph based on spatial priors, and performing a two-stage spatial hypergraph convolution in conjunction with the initial node feature matrix to obtain the final hyperedge features, specifically includes: Based on the confidence value of a patch being predicted as tumor tissue, a hypergraph based on spatial prior is constructed, and the correlation matrix is output. An initialization constraint mechanism is introduced to perform guided feature initialization. Non-zero values in the association matrix are replaced by diagnostic attention values to obtain a weighted association matrix, and the initial hyperedge features are calculated. The initial node feature matrix and initial hyperedge features are input into a spatial hypergraph attention network. An attention-guided message aggregation mechanism is executed, and the message passing path is weighted according to the confidence value. Two-stage spatial hypergraph convolutions are performed from node to hyperedge and from hyperedge to node to obtain the final hyperedge features.
[0009] Furthermore, the step of constructing a hypergraph based on spatial priors and outputting the association matrix based on the confidence value of a patch being predicted as tumor tissue specifically includes: Each pathological image patch is treated as a node in a hypergraph, and the confidence value of a patch being predicted as tumor tissue is divided into predefined intervals. Patches with similar scores are grouped into corresponding hyperedges. Set a preset overlap ratio for adjacent score intervals, construct a hypergraph based on spatial priors, and output the correlation matrix.
[0010] Furthermore, the expression for initializing the guiding feature is as follows: ; ; In the above formula, This represents the aggregate weight from the node to the hyperedge. The diagonal matrix representing the hyperedge weights. Represents the initial node feature matrix. Indicates the initial hyperedge features. Represents the weighted correlation matrix. express function.
[0011] Furthermore, the specific expression for the two-stage spatial hypergraph convolution is as follows: ; ; In the above formula, This represents the propagation weight matrix from the node to the hyperedge, calculated by the attention function. This represents the propagation weight matrix from the hyperedge to the node, calculated by the attention function. In a hypergraph attention network, the first... Hyperedge features of the layer Indicates the previous layer (i.e., the first layer) The node characteristics of the layer. In a hypergraph attention network, the first... Layer node characteristics.
[0012] Furthermore, the specific expression for predicting risk scores for different tissue regions is as follows: ; In the above formula, This indicates the risk scores for different organizational regions. This represents a multilayer perceptron. Indicates the final hyperedge feature. This represents the risk prediction function.
[0013] The second technical solution adopted in this invention is: a hypergraph prognostic prediction system based on spatial information of pathological images, comprising: The first module is used to acquire whole-slice pathological images and perform image preprocessing and diagnostic prior spatial information extraction to obtain an initial node feature matrix and a confidence value of the plot being predicted as tumor tissue. The second module is used to introduce an initialization constraint mechanism, which divides the confidence value of the map patch being predicted as tumor tissue into several predefined intervals, sets overlap between adjacent intervals, constructs a hypergraph based on spatial prior, and performs two-stage spatial hypergraph convolution in combination with the initial node feature matrix to obtain the final hyperedge features. The third module is used to perform distributed regional risk mapping based on multilayer perceptron. Based on the final hyperedge features, it performs survival risk score prediction on the heterogeneous spatial regions corresponding to each hyperedge in the hypergraph, obtains the regional risk vector, and calculates its arithmetic mean as the final survival risk, thereby realizing hypergraph prognosis prediction.
[0014] The beneficial effects of the method and system of this invention are as follows: This invention obtains an initial node feature matrix and a confidence value for a patch to be predicted as tumor tissue by acquiring a whole-slice pathological image and performing image preprocessing and diagnostic prior spatial information extraction. This eliminates the need for expensive pixel-level annotation and extracts highly correlated tumor spatial heterogeneity patterns through a weakly supervised model. Furthermore, an initialization constraint mechanism is introduced to divide the confidence value for a patch to be predicted as tumor tissue into several predefined intervals, with overlap set between adjacent intervals to construct a hypergraph based on spatial priors. Finally, a two-stage spatial hypergraph convolution is performed using the initial node feature matrix to obtain the final hyperedge feature matrix. This approach directly transforms diagnostic confidence into spatial prior knowledge to guide prognostic prediction. Hypergraph networks, combined with diagnostic prior spatial information, can better capture the complex high-order pathological spatial relationships within the tumor microenvironment (TME). Finally, distributed regional risk mapping is performed based on a multilayer perceptron. Based on the final hyperedge features, survival risk scores are predicted for the heterogeneous spatial regions corresponding to each hyperedge in the hypergraph, resulting in regional risk vectors. The arithmetic mean of these vectors is then calculated as the final survival risk, achieving hypergraph prognostic prediction. This approach improves the accuracy of survival risk prediction while providing highly interpretable pathological spatial validation. Attached Figure Description
[0015] Figure 1 This is a flowchart of the steps of a hypergraph prognostic prediction method based on spatial information of pathological images according to the present invention; Figure 2 This is a structural block diagram of a hypergraph prognostic prediction system based on spatial information of pathological images according to the present invention; Figure 3 This is a schematic diagram of the hypergraph prognostic prediction process based on the spatial information of pathological images provided in a specific embodiment of the present invention; Figure 4 This is a visualization diagram of spatial attention comparison for diagnosis and prognosis provided in a specific embodiment of the present invention. Detailed Implementation
[0016] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.
[0017] Reference Figure 1 and Figure 3 This invention provides a hypergraph prognostic prediction method based on spatial information of pathological images, the method comprising the following steps: S100. Obtain the whole-slice pathological image and perform image preprocessing and diagnostic prior spatial information extraction to obtain the initial node feature matrix and the confidence value of the patch being predicted as tumor tissue. S110. Obtain the whole-slice pathological image and perform foreground region segmentation to obtain several pathological image blocks; In this embodiment, the foreground region of the whole slice image (WSI) is segmented following the CLAM method. At a magnification of 20x, the WSI is segmented into non-overlapping patches of size 256×256 using a sliding window strategy. The pathology-based model (i.e., UNIv1) is used as the patch encoder to obtain the feature representation of each patch.
[0018] In some specific embodiments, background segmentation and slicing are performed first. Following the CLAM method, the foreground region of the whole slice image (WSI) is segmented. At a magnification of 20x, the WSI is segmented into non-overlapping patches of size 256×256 using a sliding window strategy.
[0019] Then, feature encoding is performed, using the pre-trained pathology-based model UNIv1 as a patch encoder to obtain the initial feature representation of each patch and construct the initial node feature matrix. .
[0020] S120. Based on the pre-trained pathology basic model, perform initial feature extraction processing on several pathological image blocks to construct an initial node feature matrix; S130. Input several pathological image blocks into an attention-based binary classification diagnostic model to calculate attention scores and obtain the confidence value that the image blocks are predicted to be tumor tissue.
[0021] In this embodiment, an attention-based diagnostic model (CLAM) is trained using data with only slide-level labels to perform a binary classification task (tumor vs. normal). The trained diagnostic model is used to calculate an attention score for each tile, which indicates the confidence level that the tile is predicted as tumor tissue. In this way, each WSI is decoupled into a confidence matrix using a multi-instance learning (MIL)-based binary classification model, thereby systematically quantifying and characterizing tissue heterogeneity within WSIs.
[0022] In some specific embodiments, the diagnostic model is first trained using data with only slide-level labels (normal / tumor) to train an attention-based binary classification diagnostic model.
[0023] The spatial heterogeneity matrix is then extracted, and the WSI (Warranty Indicator) for predicting the prognosis is input into the diagnostic model to obtain the attention score for each patch. This score indicates the confidence level that the patch is predicted as tumor tissue. This confidence matrix systematically quantifies and characterizes the tissue heterogeneity within the WSI, serving as the core spatial information prior for guiding subsequent hypergraph construction.
[0024] S200. An initialization constraint mechanism is introduced to divide the confidence value of the map patch being predicted as tumor tissue into several predefined intervals and set overlap between adjacent intervals to construct a hypergraph based on spatial prior. Then, a two-stage spatial hypergraph convolution is performed in combination with the initial node feature matrix to obtain the final hyperedge features. S210. Based on the confidence value of a patch being predicted as tumor tissue, construct a hypergraph based on spatial prior and output the correlation matrix. Specifically, each pathological image patch is regarded as a node in a hypergraph, and the confidence value of the patch being predicted as tumor tissue is divided into predefined intervals. Patches with similar scores are grouped into corresponding hyperedges. A preset ratio overlap is set for adjacent score intervals to construct a hypergraph based on spatial prior and output the correlation matrix.
[0025] In this embodiment, each image patch is considered a node in a hypergraph. Patches with similar scores are grouped into hyperedges based on the normalized diagnostic confidence score. The diagnostic scores are divided into predefined intervals, with a certain percentage (e.g., 50%) of overlap maintained between adjacent intervals to ensure spatial continuity. The hypergraph obtained through these steps captures two key pieces of information: first, the diagnostic information obtained through confidence-based clustering; and second, the organizational structure reflected in the hyperedge connection patterns.
[0026] In some specific embodiments, in order to capture organizational structure and spatial continuity, this embodiment constructs a hypergraph based on the acquired spatial prior information. The process specifically includes: 1) Node mapping: Treat each tile in WSI as a node in the hypergraph. .
[0027] 2) Hyperedge Clustering (Spatial Regionalization): Normalized diagnostic confidence scores are divided into predefined intervals (e.g., (0, 0.1]). Patches with similar scores are grouped to corresponding hyperedges. middle.
[0028] 3) Spatial continuity guarantee: Maintain a certain proportion (e.g., 50%) of overlap between adjacent fraction intervals to ensure spatial continuity.
[0029] 4) Construct an association matrix using the above mechanism. ( For the number of nodes, (This refers to the number of hyperedges). This hypergraph simultaneously captures diagnostic information based on confidence clustering and the organizational spatial architecture reflected through hyperedge connection patterns.
[0030] S220. Introduce an initialization constraint mechanism to perform guided feature initialization. Replace the non-zero values in the association matrix with diagnostic attention values to obtain a weighted association matrix and calculate the initial hyperedge features. S230. Input the initial node feature matrix and the initial hyperedge features into a space-based hypergraph attention network, execute the attention-guided message aggregation mechanism, constrain the message passing path with weights based on the confidence value, and perform two-stage spatial hypergraph convolution from node to hyperedge and from hyperedge to node to obtain the final hyperedge features.
[0031] In this embodiment, a spatial hypergraph attention network (HGAT) is used to learn the structural representation of WSI. Before the first layer of message passing, guided by diagnostic information from a prior diagnostic model, the initial feature matrix of the hyperedge is obtained by aggregating information from the node representations. The node features, hyperedge features, and correlation matrix are then input into two hypergraph attention convolutional layers to propagate information across the hyperedges. This process employs a two-stage (node-hyperedge-node) spatial hypergraph convolution with learnable differentiable functions.
[0032] In some specific embodiments, a spatially based Hypergraph Attention Network (HGAT) is used for structural representation learning to propagate information across hyperedges, specifically including: 1) Guided feature initialization: Before the first layer of message passing, the association matrix is replaced with diagnostic attention values. The non-zero values in the matrix are used to obtain the weighted correlation matrix. Calculate the initial hyperedge features. Its expression is: ; ; in, It is a diagonal matrix indicating the weights of the hyperedges. This represents the aggregation weight from the node to the hyperedge. The guiding feature initialization utilizes diagnostic spatial information to guide the aggregation of initial features.
[0033] 2) Two-stage spatial hypergraph convolution: Node features and hyperedge features are input into the HGAT layer. A two-stage message passing mechanism is executed, including node-to-hyperedge and hyperedge-to-node communication. ; ; in, and These represent the propagation weight matrices from the node to the hyperedge and from the hyperedge to the node, respectively, calculated by the attention function.
[0034] S300: Based on the multilayer perceptron, distributed regional risk mapping is performed. Based on the final hyperedge features, survival risk scores are predicted for the heterogeneous spatial regions corresponding to each hyperedge in the hypergraph. The regional risk vector is obtained, and its arithmetic mean is calculated as the final survival risk, thus realizing hypergraph prognosis prediction.
[0035] In this embodiment, unlike most existing methods that use pooling to obtain slide-level representations, this embodiment directly utilizes the final hyperedge features to predict risk scores for different tissue regions. The average of these risk scores for different regions is calculated as the final survival risk for each patient.
[0036] In some specific embodiments, the final hyperedge features output by the network are directly utilized. Risk scores for different organizational regions (i.e., each hyperedge) are predicted using a multilayer perceptron (MLP). Its expression is: ; In the above formula, This indicates the risk scores for different organizational regions. This represents a multilayer perceptron. Indicates the final hyperedge feature. This represents the risk prediction function, which represents the operation of mapping high-dimensional features to specific one-dimensional risk values.
[0037] Calculate the average risk score across all regions as the patient's ultimate survival risk. .
[0038] In addition, the error between predicted risk and actual survival time is calculated using the NLLSurvLoss loss function and then backpropagated.
[0039] Finally, it should be noted that the system also includes regional spatial interpretability output. To facilitate verification by pathologists, the system visualizes and reconstructs the decision-making process of the hypergraph, outputting the following comparative atlases, as shown below. Figure 4 As shown: 1) Output the original image and initial mapping: The system displays the original full-slice image (e.g., ...). Figure 4 As shown in (a)), and outputs a tile-level initial diagnostic attention heatmap extracted based on the weakly supervised diagnostic model (as shown in [a]). Figure 4 (as shown in (b)) 2) Output a regionalized diagnostic attention map based on diagnostic confidence (e.g., Figure 4 As shown in (c) (each colored area represents a superedge / a heterogeneous space interval).
[0040] 3) Output based on predicted risk The generated prognostic attention map (e.g.) Figure 4 (as shown in (d)).
[0041] 4) Generate quantitative comparison panels (e.g.) Figure 4 As shown in (e): a region-to-region direct comparison bar chart containing diagnostic and prognostic attention, and a distribution chart of the number of tiles contained in each hyperborder region, enabling pathologists to visually verify whether the decision region conforms to medical consensus.
[0042] In summary, the purpose of this invention is to provide a pathological image hypergraph prognostic prediction method guided by diagnostic prior spatial information. This method aims to address the problems of neglecting spatial contextual relationships between tissue regions, the extremely high cost of acquiring medically meaningful spatial prior knowledge, and poor clinical interpretability in existing technologies. This invention creatively uses the attention matrix output by a low-cost, weakly supervised diagnostic model as spatial information prior, and guides the construction and feature learning of the hypergraph by quantifying spatial heterogeneity within slices. This improves the accuracy of survival risk prediction while achieving highly interpretable pathological spatial validation.
[0043] The embodiments of the present invention have the following technical advantages compared with the prior art: 1) Low-cost acquisition and application of spatial information: This embodiment eliminates the need for expensive pixel-level annotation. Instead, it utilizes readily available slide-level diagnostic labels to extract highly correlated spatial heterogeneity patterns of tumors through a weakly supervised model. This diagnostic confidence is directly converted into spatial prior knowledge to guide prognostic prediction, filling the gap in low-cost spatial prior utilization in existing technologies.
[0044] 2) Enhanced generalization performance and cross-center robustness: Hypergraph networks that incorporate diagnostic prior spatial information can better capture the complex pathological spatial relationships within the tumor microenvironment (TME). In cross-validation and cross-center validation on the LUAD, BRCA, and CRC cancer datasets, the consistency index (C-index) of this embodiment outperforms existing (SOTA) methods.
[0045] 3) Providing clinical-grade regional spatial interpretability: Unlike traditional neural networks that directly extract features, this invention achieves direct spatial comparison between the "diagnostic attention heatmap" and the "prognostic risk heatmap" in corresponding tissue regions through spatial regionalization. This intuitive visualization of prior spatial information enables pathologists to efficiently verify whether the model's decision regions conform to medical consensus, greatly reducing the workload of clinical validation.
[0046] Reference Figure 2 A hypergraph prognostic prediction system based on spatial information of pathological images, comprising: The first module 201 is used to acquire whole-slice pathological images and perform image preprocessing and diagnostic prior spatial information extraction to obtain an initial node feature matrix and a confidence value of the plot being predicted as tumor tissue. The second module 202 is used to introduce an initialization constraint mechanism, which divides the confidence value of the map patch being predicted as tumor tissue into several predefined intervals, sets overlap between adjacent intervals, constructs a hypergraph based on spatial prior, and performs two-stage spatial hypergraph convolution in combination with the initial node feature matrix to obtain the final hyperedge features. The third module 203 is used to perform distributed regional risk mapping based on multilayer perceptron. Based on the final hyperedge features, it performs survival risk score prediction on the heterogeneous spatial regions corresponding to each hyperedge in the hypergraph, obtains the regional risk vector, and calculates its arithmetic mean as the final survival risk, thereby realizing hypergraph prognosis prediction.
[0047] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.
[0048] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.
Claims
1. A hypergraph prognosis prediction method based on pathological image spatial information, characterized in that, Includes the following steps: Full-section pathological images were acquired and image preprocessing and diagnostic prior spatial information extraction were performed to obtain an initial node feature matrix and confidence values of the plots predicted as tumor tissue. An initialization constraint mechanism is introduced to divide the confidence value of the map patch being predicted as tumor tissue into several predefined intervals and set overlap between adjacent intervals to construct a hypergraph based on spatial prior. Then, a two-stage spatial hypergraph convolution is performed in combination with the initial node feature matrix to obtain the final hyperedge features. Based on the multilayer perceptron, distributed regional risk mapping is performed. Based on the final hyperedge features, survival risk scores are predicted for the heterogeneous spatial regions corresponding to each hyperedge in the hypergraph. The regional risk vector is obtained, and its arithmetic mean is calculated as the final survival risk, thus realizing hypergraph prognosis prediction. 2.The method of claim 1, wherein, The step of acquiring whole-slice pathological images and performing image preprocessing and diagnostic prior spatial information extraction to obtain an initial node feature matrix and a confidence value for predicting a patch as tumor tissue specifically includes: Full-section pathological images were acquired and foreground region segmentation was performed to obtain several pathological image blocks. Based on a pre-trained pathology model, initial feature extraction is performed on several pathological image blocks to construct an initial node feature matrix. Several pathological image patches are input into an attention-based binary classification diagnostic model to calculate attention scores, thereby obtaining the confidence value that the patch is predicted to be tumor tissue. 3.The method of claim 2, wherein, The step of introducing an initialization constraint mechanism, which divides the confidence value of a patch being predicted as tumor tissue into several predefined intervals and sets overlap between adjacent intervals to construct a hypergraph based on spatial prior, and then performs a two-stage spatial hypergraph convolution combined with the initial node feature matrix to obtain the final hyperedge features, specifically includes: Based on the confidence value of a patch being predicted as tumor tissue, a hypergraph based on spatial prior is constructed, and the correlation matrix is output. An initialization constraint mechanism is introduced to perform guided feature initialization. Non-zero values in the association matrix are replaced by diagnostic attention values to obtain a weighted association matrix, and the initial hyperedge features are calculated. The initial node feature matrix and initial hyperedge features are input into a spatial hypergraph attention network. An attention-guided message aggregation mechanism is executed, and the message passing path is weighted according to the confidence value. Two-stage spatial hypergraph convolutions are performed from node to hyperedge and from hyperedge to node to obtain the final hyperedge features. 4.The method of claim 3, wherein, The step of constructing a hypergraph based on spatial priors and outputting the association matrix based on the confidence value of a patch being predicted as tumor tissue specifically includes: Each pathological image patch is treated as a node in a hypergraph, and the confidence value of a patch being predicted as tumor tissue is divided into predefined intervals. Patches with similar scores are grouped into corresponding hyperedges. Set a preset overlap ratio for adjacent score intervals, construct a hypergraph based on spatial priors, and output the correlation matrix.
5. The hypergraph-based prognosis prediction method based on spatial information of pathological images according to claim 4, characterized in that, The expression for initializing the guiding feature is as follows: ; ; In the above formula, denotes the aggregated weight of a node to a hyperedge, denotes a diagonal matrix of hyperedge weights, denotes the initial node feature matrix, denotes the initial hyperedge feature, denotes the weighted affinity matrix, denotes the function. 6.The method of claim 5, wherein, The specific expression for the two-stage spatial hypergraph convolution is as follows: ; ; In the above formula, denotes the propagation weight matrix from nodes to hyper-edges computed by the attention function, denotes the propagation weight matrix from hyper-edges to nodes computed by the attention function, denotes the hyper-edge feature in the layer of the hypergraph attention network, denotes the node feature of the previous layer (i.e., the layer), denotes the node feature in the layer of the hypergraph attention network.
7. The hypergraph-based prognosis prediction method based on spatial information of pathological images according to claim 6, characterized in that, The specific expression for predicting risk scores for different tissue regions is as follows: ; In the above formulae, denotes a risk score for different tissue regions, denotes a multi-layer perceptron, denotes a final hyperedge feature, denotes a risk prediction function.
8. A hypergraph prognostic prediction system based on pathological image spatial information, characterized in that, Includes the following modules: The first module is used to acquire whole-slice pathological images and perform image preprocessing and diagnostic prior spatial information extraction to obtain an initial node feature matrix and a confidence value of the plot being predicted as tumor tissue. The second module is used to introduce an initialization constraint mechanism, which divides the confidence value of the map patch being predicted as tumor tissue into several predefined intervals, sets overlap between adjacent intervals, constructs a hypergraph based on spatial prior, and performs two-stage spatial hypergraph convolution in combination with the initial node feature matrix to obtain the final hyperedge features. The third module is used to perform distributed regional risk mapping based on multilayer perceptron. Based on the final hyperedge features, it performs survival risk score prediction on the heterogeneous spatial regions corresponding to each hyperedge in the hypergraph, obtains the regional risk vector, and calculates its arithmetic mean as the final survival risk, thereby realizing hypergraph prognosis prediction.
