A whole-slice pathological image prognosis evaluation method and system based on risk focus identification and tumor microenvironment context aggregation, an electronic device, and a storage medium

Abstract

The application discloses a kind of based on risk focus identification and tumor microenvironment context aggregation whole section pathological image prognosis evaluation method, system, electronic equipment and storage medium, comprising: first to whole section pathological graph block extraction image block feature, pass through disturbance consistency, organization form prototype complete double reliability evaluation, adopt three branch rectification purification feature;Again risk-microenvironment double passage shunt aggregation, respectively generate local high-risk focus, global microenvironment representation and fusion obtain section level prognosis feature;Finally output survival risk score, stratification result and two kinds of interpretable heat map.The application solves the problems of traditional multi-instance learning noise interference, high-risk signal dilution, microenvironment modeling deficiency and poor interpretability, and the prognosis evaluation performance is better than that of existing mainstream model in multi-cancer experiment.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method, system, electronic device, and storage medium for prognostic assessment of whole-slice pathological images based on risk foci identification and tumor microenvironment context aggregation. Background Technology

[0002] Whole-section pathological images contain information closely related to patient prognosis, such as tumor cell morphology, tissue structure, tumor stroma reaction, necrotic areas, lymphocyte infiltration, tumor boundaries, and the tumor microenvironment. Using whole-section pathological images for survival risk assessment helps in patient prognostic stratification, treatment selection, clinical trial risk stratification, and assists in follow-up management.

[0003] Due to the enormous size of whole-slice pathological images, existing methods typically divide them into numerous image patches and use weakly supervised multi-instance learning frameworks for modeling. Typical methods include attention-based multi-instance learning, two-stream multi-instance learning, Transformer aggregation, optimal transport aggregation, and state-space model aggregation. These methods usually first extract image patch features, then weight and summarize the features of all image patches, and finally output a patient-level survival risk score.

[0004] However, existing methods still have the following shortcomings: First, pathological whole-section images often contain tissue folds, abnormal staining, marker marks, scan blur, tissue damage, non-tumor background, and morphologically blurred areas. Existing multi-instance learning methods typically assume that the attention module can automatically ignore these areas, but they lack explicit feature credibility assessment and feature intervention mechanisms. When low-quality or morphologically blurred areas enter the aggregation module, the model may give overconfident or incorrect risk estimates.

[0005] Second, survival risk assessment differs from ordinary classification tasks, as its objectives are typically influenced by multiple pathological factors. Patient prognosis may be driven by the most aggressive local malignant lesions, or it may be influenced by the global tumor microenvironment, immune infiltration, stromal reaction, necrosis, and tissue structure patterns. Single average attention aggregation can easily dilute local high-risk signals, while selecting only the highest-risk image patch makes it difficult to describe the complex microenvironment context.

[0006] Third, existing methods typically only learn the importance of image patches during the aggregation stage, lacking a mechanism to actively correct, retain, or remove ambiguous features before they enter the aggregation process. For image patches at pathological boundaries, such as poorly differentiated regions, tumor-stromal transition zones, and areas with mixed inflammation, simple deletion may result in the loss of valuable information, while direct retention may introduce noise.

[0007] Fourth, while existing survival assessment models can output heatmaps, these heatmaps typically only reflect attention weights and are difficult to distinguish whether the model focuses on local high-risk malignant areas or the global tumor microenvironment pattern, which is not conducive to pathologists understanding the basis of the model's assessment.

[0008] Therefore, how to propose a method, system, electronic device and storage medium for prognostic assessment of whole-slice pathological images based on risk foci identification and tumor microenvironment context aggregation, and overcome the shortcomings of existing technologies, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0009] In view of this, the present invention provides a method, system, electronic device, and storage medium for prognostic assessment of whole-slice pathological images based on risk foci identification and tumor microenvironment context aggregation. Before feature aggregation, the reliability of image patch features is assessed and morphological prototypes are rectified. During the aggregation stage, local high-risk lesions and the global tumor microenvironment are modeled separately, thereby improving the robustness and interpretability of survival risk assessment. This addresses the problems of low-quality image patch interference, insufficient processing of blurred morphological features, dilution of local high-risk signals, insufficient global microenvironment modeling, and weak interpretability of assessment results in existing survival assessment methods. To achieve the above objectives, the present invention adopts the following technical solution: A method for prognostic assessment of whole-section pathological images based on risk foci identification and tumor microenvironment context aggregation includes: The whole-section pathological image to be analyzed is obtained. The whole-section pathological image is preprocessed to obtain multiple tissue image blocks. The pathological image feature extraction model is used to extract features from each tissue image block to obtain an image block-level feature set. Perturbation consistency assessment is performed on each image block-level pathological feature, a tissue morphology prototype library is constructed, and the morphological attribution of image block features is assessed based on the closeness between image block features and tissue morphology prototypes. By fusing the perturbation consistency assessment results and the morphological attribution assessment results, image block-level feature confidence scores or feature processing labels are obtained. Based on the feature confidence scores or feature processing labels, multi-branch feature rectification is performed on the image block features to obtain the purified pathological feature set. The purified set of pathological features is input into the risk-microenvironment triage and aggregation module to generate local high-risk lesion representation and global microenvironment representation. The local high-risk lesion representation and global microenvironment representation are then fused to obtain the whole-slice-level prognostic representation. Input the whole-slice-level prognostic representation into the risk prediction head, and output continuous survival risk score, high and low risk stratification results, heat map of key high-risk areas and heat map of tumor microenvironment pattern.

[0010] Optionally, the pathological image feature extraction model includes: a convolutional neural network, a visual Transformer network, or a pathological basic model encoder pre-trained on natural or pathological images.

[0011] Optionally, the tissue morphology prototype library contains multiple representative tissue morphology prototypes to represent pathological morphological patterns in the training set.

[0012] Optionally, the step of performing multi-branch feature rectification on image patch features based on feature confidence scores or feature processing labels includes: retaining high-confidence features; correcting medium-confidence or morphologically ambiguous features using tissue morphology prototypes; and reducing the weight of low-confidence features or removing them from subsequent aggregation.

[0013] Optionally, the risk-microenvironment diversion and aggregation module includes local risk pathways and global microenvironment pathways.

[0014] Optionally, in the local risk pathway, each purified feature in the purified pathological feature set is scored for prognostic risk using an instance risk scoring network, and one or more purified features with the most prognostic significance are selected or weighted and summarized based on the risk scores to generate a representation of local high-risk lesions.

[0015] Optionally, within the global microenvironment pathway, multiple learnable context queries or pattern queries are used to extract different tumor microenvironment patterns from the purified set of pathological features to generate a global microenvironment representation.

[0016] Optionally, a whole-section pathological image prognostic assessment system based on risk foci identification and tumor microenvironment context aggregation includes: The acquisition and processing module is used to acquire the whole-slice pathological images to be analyzed, preprocess the whole-slice pathological images to obtain multiple tissue image blocks, and use the pathological image feature extraction model to extract features from each tissue image block to obtain an image block-level feature set. Tissue morphology prototype library module: used to perform perturbation consistency assessment on each image block-level pathological feature, construct a tissue morphology prototype library, and perform morphological attribution assessment on image block features based on the closeness between image block features and tissue morphology prototypes. Feature credibility fusion module: used to fuse perturbation consistency assessment results and morphological attribution assessment results to obtain image block-level feature credibility score or feature processing label. Based on the feature credibility score or feature processing label, multi-branch feature rectification is performed on the image block features to obtain the purified pathological feature set. Multi-branch feature rectification module: used to input the purified pathological feature set into the risk-microenvironment diversion and aggregation module to generate local high-risk lesion representation and global microenvironment representation, and to fuse the local high-risk lesion representation and global microenvironment representation to obtain the whole-slice level prognostic representation; Prognostic assessment module: This module takes a full-slice-level prognostic representation as input into the risk prediction head and outputs a continuous survival risk score, high- and low-risk stratification results, a heatmap of key high-risk areas, and a heatmap of the tumor microenvironment pattern.

[0017] Optionally, a computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the computer program to implement a whole-section pathological image prognostic assessment method based on risk foci identification and tumor microenvironment context aggregation.

[0018] Optionally, a computer-readable storage medium stores a computer program that, when executed by a processor, implements a method for prognostic assessment of whole-slice pathological images based on risk foci identification and tumor microenvironment context aggregation.

[0019] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method, system, electronic device and storage medium for prognostic assessment of whole-section pathological images based on risk foci identification and tumor microenvironment context aggregation, which has the following beneficial effects: This invention identifies low-quality, abnormally stained, blurred, tissue-folded, and morphologically distributed features before image patch features enter the aggregation module by perturbation consistency assessment and morphological attribution assessment, thereby reducing the interference of noise features on survival risk estimation from the source.

[0020] This invention performs prototype correction on moderately reliable or morphologically ambiguous features using a tissue morphology prototype library, thereby reducing the shift of abnormal features without simply deleting potentially effective regions, and making boundary state regions such as poorly differentiated regions, tumor-stromal transition zones, and inflammatory mixed regions more robustly handled.

[0021] This invention highlights high-risk lesions with the greatest prognostic significance through local risk pathways, avoiding the average aggregation and dilution of local invasive signals; and aggregates immune infiltration, necrosis, interstitial reactions and tissue structure patterns through global microenvironment pathways, avoiding insufficient context due to relying solely on a single highest-risk region.

[0022] This invention improves the robustness, interpretability, and clinical readability of survival risk assessment by using risk-microenvironment decentralization and aggregation, enabling whole-slice-level prognostic representation to simultaneously include local high-risk driving factors and global tumor microenvironment factors.

[0023] This invention generates heatmaps of key high-risk areas, heatmaps of tumor microenvironment patterns, and feature processing distribution maps, which can distinguish whether the model focuses on local malignant lesions, global microenvironment patterns, or low-confidence areas that have been suppressed / corrected, thus facilitating review by pathologists. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of a whole-slice pathological image prognostic assessment method based on risk foci identification and tumor microenvironment context aggregation, provided by the present invention.

[0026] Figure 2 A schematic diagram of the risk-microenvironment diversion and aggregation module provided for this invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] This invention discloses a method for prognostic assessment of whole-section pathological images based on risk foci identification and tumor microenvironment context aggregation, such as... Figure 1 As shown, it includes: The whole-section pathological image to be analyzed is obtained. The whole-section pathological image is preprocessed to obtain multiple tissue image blocks. The pathological image feature extraction model is used to extract features from each tissue image block to obtain an image block-level feature set. Perturbation consistency assessment is performed on each image block-level pathological feature, a tissue morphology prototype library is constructed, and the morphological attribution of image block features is assessed based on the closeness between image block features and tissue morphology prototypes. By fusing the perturbation consistency assessment results and the morphological attribution assessment results, image block-level feature confidence scores or feature processing labels are obtained. Based on the feature confidence scores or feature processing labels, multi-branch feature rectification is performed on the image block features to obtain the purified pathological feature set. The purified set of pathological features is input into the risk-microenvironment triage and aggregation module to generate local high-risk lesion representation and global microenvironment representation. The local high-risk lesion representation and global microenvironment representation are then fused to obtain the whole-slice-level prognostic representation. Input the whole-slice-level prognostic representation into the risk prediction head, and output continuous survival risk score, high and low risk stratification results, heat map of key high-risk areas and heat map of tumor microenvironment pattern.

[0029] In a specific implementation, a method for prognostic assessment of whole-section pathological images based on risk foci identification and tumor microenvironment context aggregation includes the following steps: S1, acquire the whole-slice pathological image to be analyzed, and perform tissue region detection, background filtering, color normalization, and image block segmentation on the whole-slice pathological image to obtain multiple tissue image blocks. Specifically: Let the input whole-slice pathological image, tissue detection function, and color normalization function be as follows: ; ; ; Tissue detection, color normalization, and image patch cropping can be represented as: ; ; ; in, To effectively organize the mask, Indicates the first A tissue image patch, For the spatial location of the image patch, This refers to the image block size. , The parameters for the organization detection function and the color normalization function can include preset thresholds, model weights, or color normalization template parameters.

[0030] S2, The pathological image feature extraction model is used to extract features from each tissue image patch, resulting in a patch-level pathological feature set. Specifically, the pathological image feature extraction model can be a convolutional neural network, visual Transformer network, or pathological basic model encoder pre-trained on natural or pathological images, such as ResNet, DenseNet, ViT, SwinTransformer, CTransPath, UNI, or CONCH. In practice, an encoder pre-trained on H&E pathological images through contrastive learning, mask image modeling, or multi-instance learning can also be used. Specifically: Pathological image feature extraction model Map each tissue image patch to a feature vector: ; ; in, It can be ResNet, DenseNet, ViT, SwinTransformer, CTransPath, UNI, CONCH, or other encoders pre-trained on pathological images. For feature dimension, To effectively organize the number of image blocks.

[0031] S3, perform a perturbation consistency assessment on each image patch-level pathological feature to determine whether the feature is stable under slight perturbations. Specifically: Apply a slight perturbation to each image patch For example, color perturbation, rotation, scaling, cropping, or blurring perturbation yields: ; The perturbation consistency score can be defined as: ; in, For the number of perturbations, For temperature parameters, The larger the value, the more stable the characteristic.

[0032] S4, Construct a tissue morphology prototype library. This library contains multiple representative tissue morphology prototypes to represent common pathological morphological patterns in the training set, such as tumors, stroma, necrosis, inflammation, normal tissue, or other pathological morphological patterns. Specifically: The organizational morphology prototype library is represented as follows: ; in, This represents the number of morphological prototypes. Each prototype can be obtained through training set feature clustering, category prototype statistics, or learnable parameter updates, and is used to represent common morphological patterns such as tumors, stroma, necrosis, inflammation, and normal tissues.

[0033] S5, based on the similarity between image patch features and tissue morphology prototypes, performs morphological attribution assessment on image patch features to determine whether they deviate from common tissue morphology distributions. Specifically: Calculate the distance between image patch features and the prototype: ; or: ; The probability of morphological attribution is defined as: ; like Smaller and A larger value indicates that the image patch features are close to the distribution of common tissue morphology; conversely, a smaller value indicates that its morphological classification is unclear or deviates from the training distribution. Temperature parameter used in morphological assignment probability calculation. For the first The minimum distance between a feature of an image patch and the prototype of all tissue morphology. For the first The maximum probability of each image patch being assigned to all tissue morphology prototypes.

[0034] S6, by fusing the perturbation consistency assessment results and the morphological attribution assessment results, obtains the image block-level feature confidence score or feature processing label. Specifically: By fusing the perturbation consistency and morphological attribution results, a feature reliability score is obtained: ; in, This is the fusion coefficient. Based on the threshold... and (satisfy For example, it is advisable , Features can also be categorized into high-confidence, medium-confidence, and low-confidence classes (either pre-defined based on the confidence score distribution of the training or validation set or determined by quantiles), and processing labels can be generated accordingly. .

[0035] S7. Based on the feature confidence score or feature processing label, perform multi-branch feature rectification on the image patch features. High-confidence features are directly retained; medium-confidence or morphologically ambiguous features are corrected using tissue morphology prototypes; low-confidence features have their weight reduced or are removed from subsequent aggregation, resulting in a purified set of pathological features. Specifically: Perform multi-branch feature rectification. High-confidence features are directly retained: ; Moderately reliable or morphologically ambiguous features are corrected using tissue morphology prototypes: ; Low-confidence features should have their weights reduced or be removed from the set: ; The set of pathological features after purification is as follows: ; in, can be Decide, Low-confidence suppression weights.

[0036] S8 inputs the purified set of pathological features into the risk-microenvironment triage and aggregation module. For example... Figure 2 As shown, the risk-microenvironment triage and aggregation module includes local risk pathways and global microenvironment pathways. Specifically: Input the purified set of pathological features into the risk-microenvironment triage and aggregation module: ; in, As a local risk pathway, It serves as a global microenvironment pathway.

[0037] S9, in the local risk pathway, uses an instance risk scoring network to assign a prognostic risk score to each purified feature in the purified pathological feature set. Based on the risk scores, one or more purified features with the highest prognostic significance are selected or weighted to generate a representation of a local high-risk lesion. The output here is not the unprocessed original image patch features, but rather a local high-risk lesion-level representation obtained from the purified pathological feature set after risk scoring, Top-K selection, or attention weighting. Specifically: For each cleanup feature, the local risk pathway calculates an instance risk score: ; in, This is the transpose of the learnable weight vector of the instance risk scoring network. It is a nonlinear characteristic transformation function. This is a bias term.

[0038] And based on the Top-K selection set or attention weights, the representation of local high-risk lesions is obtained: ; ; ; This indicates a high-risk lesion area, rather than the features of the original, unpurified image patch.

[0039] S10, within the global microenvironment pathway, utilizes multiple learnable context queries or pattern queries to extract different tumor microenvironment patterns from the purified pathological feature set, generating a global microenvironment representation. Specifically: Global microenvironment pathway settings A learnable context query ,calculate: ; in, This is a learnable feature transformation matrix used to project sanitized features into an attention space consistent with the context query; Indicates in all Normalization is performed on each purification feature.

[0040] The corresponding microenvironment model is represented as: ; The entire microenvironment representation is obtained by splicing or pooling: .

[0041] S11 fuses the representation of local high-risk lesions with the global microenvironment representation to obtain a full-slice-level prognostic representation. Specifically: Integrating local high-risk lesion representation and global microenvironment representation: ; in, This represents the prognosis at the whole slice level. This indicates vector concatenation. , These are the fusion bias vector and the fusion weight matrix, respectively. Presentation layer normalization operation.

[0042] S12 takes the full-slice-level prognostic representation as input into the risk prediction head and outputs a continuous survival risk score, high- and low-risk stratification results, a heatmap of key high-risk areas, and a heatmap of the tumor microenvironment pattern. Specifically: The risk prediction head outputs a continuous survival risk score: ; in, , , These are the transpose of the learnable weight vector of the risk prediction head, the nonlinear mapping function, and the bias term, respectively.

[0043] Generate high- and low-risk stratification based on thresholds determined from the training or validation set: ; in, The threshold for high and low risk stratification is preferably determined based on the median of the risk scores in the training or validation set, a preset quantile, or clinical criteria, and should remain fixed during the testing or deployment phase.

[0044] If Cox survival loss is used for training, the following can be used: ; in, For survival time, This indicates the outcome of the event. A heat map of key high-risk areas can be generated by... or The microenvironment pattern heatmap obtained through mapping can be derived from... The feature processing distribution map obtained by mapping can be derived from... The mapping is obtained.

[0045] The global mapping can be written as: ; in, This represents a set of image block-level pathological features generated by a pathological image feature extraction model. This indicates the retention, correction, or suppression operations performed based on the perturbation consistency score, morphological prototype attribution probability, and confidence score. Local high-risk lesion representations are generated using an instance risk scoring network; Aggregate tumor microenvironment patterns through multiple learnable context queries; Local high-risk information and global micro-environment information are fused into a full-slice-level prognostic representation; Output continuous survival risk score and risk stratification results.

[0046] In a specific implementation, a whole-slice pathological image prognostic assessment system based on risk foci identification and tumor microenvironment context aggregation includes: The acquisition and processing module is used to acquire the whole-slice pathological images to be analyzed, preprocess the whole-slice pathological images to obtain multiple tissue image blocks, and use the pathological image feature extraction model to extract features from each tissue image block to obtain an image block-level feature set. Tissue morphology prototype library module: used to perform perturbation consistency assessment on each image block-level pathological feature, construct a tissue morphology prototype library, and perform morphological attribution assessment on image block features based on the closeness between image block features and tissue morphology prototypes. Feature credibility fusion module: used to fuse perturbation consistency assessment results and morphological attribution assessment results to obtain image block-level feature credibility score or feature processing label. Based on the feature credibility score or feature processing label, multi-branch feature rectification is performed on the image block features to obtain the purified pathological feature set. Multi-branch feature rectification module: used to input the purified pathological feature set into the risk-microenvironment diversion and aggregation module to generate local high-risk lesion representation and global microenvironment representation, and to fuse the local high-risk lesion representation and global microenvironment representation to obtain the whole-slice level prognostic representation; Prognostic assessment module: This module takes a full-slice-level prognostic representation as input into the risk prediction head and outputs a continuous survival risk score, high- and low-risk stratification results, a heatmap of key high-risk areas, and a heatmap of the tumor microenvironment pattern.

[0047] In a specific embodiment, a computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a method for prognostic assessment of whole-slice pathological images based on risk foci identification and tumor microenvironment context aggregation.

[0048] In a specific embodiment, a computer-readable storage medium stores a computer program that, when executed by a processor, implements a method for prognostic assessment of whole-slice pathological images based on risk foci identification and tumor microenvironment context aggregation.

[0049] In a specific embodiment, the input is a whole-section pathological image stained with H&E obtained from patient diagnosis or surgery. The system first generates an effective tissue region and segments it into multiple tissue image blocks, then uses a pathological image feature extraction model to obtain image block-level pathological features. Each image block is slightly perturbed and re-encoded to calculate perturbation consistency; simultaneously, a tissue morphology prototype library is used to assess whether it resembles common morphological patterns such as tumors, stroma, necrosis, inflammation, or normal tissue.

[0050] Image patch features with both high perturbation consistency and clear morphological attribution are directly retained; when image patches are at morphological boundaries or in a moderately confident state, they are corrected using the closest tissue morphological prototype; when image patches exhibit low confidence or significantly deviate from common morphological distributions, their weight is reduced or they are removed. The purified set of pathological features is fed into the local risk pathway and the global microenvironment pathway, respectively. The former locates the most prognostic high-risk lesions, while the latter aggregates immune infiltration, necrosis, interstitial reactions, and tissue structure patterns, finally outputting a continuous survival risk score and high-low risk stratification.

[0051] Table 1 illustrates the performance differences between the model of this invention and other comparative methods such as ABMIL, CLAM, DSMIL, TransMIL, DTFD-MIL, HIPT, MambaMIL, OTSurv, and Panther on five prognostic risk assessment tasks: BRCA, LUSC, BLCA, STAD, and CRC. The values ​​in the table can be interpreted as the mean ± standard deviation of the C-index or similar survival assessment consistency index. The Mean represents the average performance across the five tasks. The Mean of the model of this invention is 0.691, higher than all the comparative methods, indicating that feature credibility rectification and risk-microenvironment diversion aggregation can improve the performance of prognostic risk assessment.

[0052] Table 1 Comparison Results of Survival Risk Assessment Performance

[0053] Table 2 illustrates the incremental contributions of each module. The basic aggregation model performs only routine aggregation of image patch features; with the addition of split aggregation, the model begins to model local high-risk lesions and the global microenvironment separately; with the addition of low-confidence feature suppression, the interference of low-quality or morphologically abnormal regions on the assessment is reduced; the complete model further retains the morphological prototype correction and confidence fusion mechanisms. As shown in Table 2, the Mean of the complete model is 0.691, which is higher than the 0.645 of the basic aggregation model, indicating that each module can synergistically enhance the robustness and interpretability of survival risk assessment.

[0054] Table 2 Ablation Experiment Results

[0055] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0056] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for prognostic assessment of whole-section pathological images based on risk foci identification and tumor microenvironment context aggregation, characterized in that, include: The whole-section pathological image to be analyzed is obtained. The whole-section pathological image is preprocessed to obtain multiple tissue image blocks. The pathological image feature extraction model is used to extract features from each tissue image block to obtain an image block-level feature set. Perturbation consistency assessment is performed on each image block-level pathological feature, a tissue morphology prototype library is constructed, and the morphological attribution of image block features is assessed based on the closeness between image block features and tissue morphology prototypes. By fusing the perturbation consistency assessment results and the morphological attribution assessment results, image block-level feature confidence scores or feature processing labels are obtained. Based on the feature confidence scores or feature processing labels, multi-branch feature rectification is performed on the image block features to obtain the purified pathological feature set. The purified set of pathological features is input into the risk-microenvironment triage and aggregation module to generate local high-risk lesion representation and global microenvironment representation. The local high-risk lesion representation and global microenvironment representation are then fused to obtain the whole-slice-level prognostic representation. Input the whole-slice-level prognostic representation into the risk prediction head, and output continuous survival risk score, high and low risk stratification results, heat map of key high-risk areas and heat map of tumor microenvironment pattern.

2. The method for prognostic assessment of whole-section pathological images based on risk foci identification and tumor microenvironment context aggregation according to claim 1, characterized in that, The pathological image feature extraction model includes: a convolutional neural network, a visual Transformer network, or a pathological basic model encoder pre-trained on natural or pathological images.

3. The method for prognostic assessment of whole-section pathological images based on risk foci identification and tumor microenvironment context aggregation according to claim 1, characterized in that, The tissue morphology prototype library contains multiple representative tissue morphology prototypes used to represent pathological morphological patterns in the training set.

4. The method for prognostic assessment of whole-section pathological images based on risk foci identification and tumor microenvironment context aggregation according to claim 1, characterized in that, The step of performing multi-branch feature rectification on image patch features based on feature confidence scores or feature processing labels includes: retaining high-confidence features; correcting medium-confidence or morphologically ambiguous features using tissue morphology prototypes; and reducing the weight of low-confidence features or removing them from subsequent aggregation.

5. The method for prognostic assessment of whole-section pathological images based on risk foci identification and tumor microenvironment context aggregation according to claim 1, characterized in that, The risk-microenvironment diversion and aggregation module includes local risk pathways and global microenvironment pathways.

6. The method for prognostic assessment of whole-section pathological images based on risk foci identification and tumor microenvironment context aggregation according to claim 5, characterized in that, In the local risk pathway, each purified feature in the purified pathological feature set is scored for prognostic risk using an instance risk scoring network. Based on the risk scores, one or more purified features with the most prognostic significance are selected or weighted and summarized to generate a representation of local high-risk lesions.

7. The method for prognostic assessment of whole-section pathological images based on risk foci identification and tumor microenvironment context aggregation according to claim 5, characterized in that, In the global microenvironment pathway, multiple learnable context queries or pattern queries are used to extract different tumor microenvironment patterns from the purified pathological feature set and generate a global microenvironment representation.

8. A whole-section pathological image prognostic assessment system based on risk foci identification and tumor microenvironment context aggregation, characterized in that, include: The acquisition and processing module is used to acquire the whole-slice pathological images to be analyzed, preprocess the whole-slice pathological images to obtain multiple tissue image blocks, and use the pathological image feature extraction model to extract features from each tissue image block to obtain an image block-level feature set. Tissue morphology prototype library module: used to perform perturbation consistency assessment on each image block-level pathological feature, construct a tissue morphology prototype library, and perform morphological attribution assessment on image block features based on the closeness between image block features and tissue morphology prototypes. Feature credibility fusion module: used to fuse perturbation consistency assessment results and morphological attribution assessment results to obtain image block-level feature credibility score or feature processing label. Based on the feature credibility score or feature processing label, multi-branch feature rectification is performed on the image block features to obtain the purified pathological feature set. Multi-branch feature rectification module: used to input the purified pathological feature set into the risk-microenvironment diversion and aggregation module to generate local high-risk lesion representation and global microenvironment representation, and to fuse the local high-risk lesion representation and global microenvironment representation to obtain the whole-slice level prognostic representation; Prognostic assessment module: This module takes a full-slice-level prognostic representation as input into the risk prediction head and outputs a continuous survival risk score, high- and low-risk stratification results, a heatmap of key high-risk areas, and a heatmap of the tumor microenvironment pattern.

9. A computer device, comprising: The memory, the processor, and the computer program stored in the memory and executable on the processor are characterized in that the processor executes the computer program to implement the whole-slice pathological image prognostic assessment method based on risk foci identification and tumor microenvironment context aggregation as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements a whole-slice pathological image prognostic assessment method based on risk foci identification and tumor microenvironment context aggregation, as described in any one of claims 1-7.

Images