Cancer prognosis model training method and device based on histomics and prediction method

Abstract

The invention discloses a histomics-based cancer prognosis model training method, a histomics-based cancer prognosis model training device and a histomics-based cancer prognosis model prediction method. The histomics-based cancer prognosis model training method comprises the following steps: acquiring a WSI image data set; the WSI image data set is preprocessed; training the classification network based on the pathological basic large model through the WSI image data set; obtaining features extracted in the training process; mapping according to the extracted features and each WSI image to construct a histopathology knowledge graph; obtaining a pre-selected characteristic index group suitable for the Transform model; and training the Transform model according to the pre-selected characteristic index group, so as to obtain the trained Transform model. According to the method, the neural network is used for extracting information such as area, proportion, distribution, form and molecular characterization difference with tissue as a unit, a histomics framework is used for constructing a multivariate significant feature attention model to achieve overall lifetime (OS) risk prediction, and the histopathological image analysis efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method for training a cancer prognostic model based on histomics, a device for training a cancer prognostic model based on histomics, and a method for predicting cancer prognosis based on histomics. Background Technology

[0002] Cancer threatens the lives of millions worldwide and is one of the most serious challenges facing global medicine today, making effective treatment crucial. Prognostic prediction is an effective means of assessing the effectiveness of cancer treatment. Prognostic prediction refers to predicting the disease's progression, possible outcomes, and treatment response by analyzing multi-dimensional information such as the patient's clinical data, biomarkers, and imaging examinations. Tumor histopathology has always held a fundamental position in the field of tumor diagnosis and treatment, reflecting the complex heterogeneity of the tumor environment. For a long time, doctors have accurately determined the type and grade of tumors by carefully observing the morphology, structure, and arrangement of cells in tissue sections, thereby making a preliminary assessment of the patient's prognosis.

[0003] With the application of digital technology in pathological diagnosis, whole-view digital slides (WSI) provide richer information for pathological analysis. However, a single WSI image may contain multiple types of tumors. Taking pancreatic tumors as an example, they can be classified into twelve categories according to pathological type: cancer or high-grade pancreatic intraepithelial neoplasia (PIN), low-grade PIN, normal tissue, acinar atrophy or ductalization, tumor stroma, immune infiltration, lymph nodes, nerves, blood vessels, hematopoietic cell aggregation, necrosis, and others. Therefore, thousands of different features can be extracted from these images. These features are numerous and diverse, containing both valuable information for cancer diagnosis and prognosis, as well as some interfering factors. How to accurately screen specific features for different types of cancer from this massive dataset is a pressing challenge in computational pathology and tumor diagnosis.

[0004] Differences in scanners, staining methods, and slide preparation techniques across laboratories can lead to variations in the image representation of the same tissue across different datasets. This variation results in performance degradation for models applied across datasets, limiting their versatility and applicability. To overcome these challenges, a multi-center co-processing strategy can be employed in model training. Multi-center co-processing aims to leverage pathology data from different centers to learn the most important features of cancer, thereby improving model performance on test data. Through multi-center co-processing, the model can recognize and adapt to differences in data distributions, thus maintaining stable classification performance on WSI datasets generated by different laboratories or devices.

[0005] Therefore, it is desirable to have a technical solution to overcome or at least mitigate one of the aforementioned defects of the prior art.

[0006] Application content

[0007] The purpose of this application is to provide a method for training cancer prognostic models based on tissue omics to overcome or at least mitigate one of the aforementioned deficiencies of the prior art.

[0008] To achieve the above objectives, this application provides a method for training a cancer prognostic model based on histomics, the method comprising:

[0009] Obtain a WSI image dataset, which includes multiple WSI image datasets;

[0010] The WSI image dataset is preprocessed to obtain a preprocessed WSI image dataset;

[0011] The preprocessed WSI image dataset is used to train a classification network based on a large pathological model, thereby obtaining the trained classification network based on a large pathological model.

[0012] Acquire the features extracted during the training of the classification network based on the pathological basis large model;

[0013] A histopathology knowledge graph is constructed by mapping the extracted features and each WSI image.

[0014] Based on the histopathology knowledge graph and each WSI image, a pre-selected set of feature indicators suitable for the Transformer model is obtained;

[0015] The Transformer model is trained based on the pre-selected feature index group to obtain the trained Transformer model.

[0016] Optionally, preprocessing the WSI image dataset to obtain a preprocessed WSI image dataset includes:

[0017] The following processing is performed on each WSI image data:

[0018] WSI data is acquired and foreground selection is performed on the WSI image based on digital image processing to obtain the foreground region;

[0019] In the found foreground region, find the largest bounding box to segment the foreground and determine it as the valid region;

[0020] The effective domain is mapped to the WSI image to obtain the computational domain, and the coordinates of the tiles to be segmented are determined.

[0021] The tiles are cut according to their coordinates, and the size of the cut tiles is 256×256.

[0022] Optionally, training the classification network based on the pathological baseline model using the WSI image dataset to obtain the trained classification network based on the pathological baseline model includes:

[0023] Read the tile images at 20x, 10x, and 5x pixel sizes after they have been cut;

[0024] The read tiles image is color normalized.

[0025] The normalized tiles image is flipped, rotated, affine transformed, and random Gaussian noise is added to obtain the incremented tiles image. The incremented tiles image and the normalized tiles image together constitute the training data.

[0026] The training data is read using an 80% overlap sliding window strategy and fed into a classification network based on a large pathological model for training.

[0027] Optionally, the step of using an 80% overlap sliding window strategy to read training data and feeding it into a classification network based on a large pathological model for training includes:

[0028] Perform the following operations for each round of training:

[0029] The training results of the basic pathology model are fed into the output network to achieve classification, and the cross-entropy function is used to measure the multi-classification error between the predicted results and the true labels.

[0030] If the error value is less than that of the previous model, the model parameters are saved. Training ends when the number of training rounds reaches a preset value or the model converges.

[0031] Optionally, the step of mapping and constructing a histopathological knowledge graph based on the extracted features and each WSI image includes:

[0032] A histopathological atlas was constructed based on the features extracted from the model and the classification results.

[0033] Optionally, obtaining a pre-selected set of feature indicators suitable for the Transformer model based on the histopathology knowledge graph and each WSI image includes:

[0034] The Mann-Whitney U test was used to identify features of death within a fixed time period. Features with a p-value less than 0.05 were selected as features significantly associated with overall survival, and features with repeated selections were merged.

[0035] The Pearson test was used to assess the linear correlation between features, and features with an R value greater than 0.95 were excluded.

[0036] After the linear correlation test, the Cox regression with minimum absolute contraction and selection operator is used for feature selection. After feature screening, the retained features form a pre-selected feature index group.

[0037] Optionally, the survival composite loss function of the Transformer model is as follows:

[0038] l sur-comp =l prop-risk +l Time =L sur-comp (r,t,e,timelist,P);

[0039] Where r represents the risk score vector, t represents the time vector, and e represents the event indication vector; l sur-comp Represents the survival composite loss function; l prop-risk For proportional risk loss function; l Time The loss function for classifying time-dependent events.

[0040] Optionally, the proportional risk loss function formula is as follows:

[0041] Where π is the sequence of events ordered by time t; risk π(i) Risk scoring; event π(i) For event indicators.

[0042] This application also provides a tissue-based cancer prognostic model training device, the tissue-based cancer prognostic model training device comprising:

[0043] WSI image dataset acquisition module, the WSI image dataset acquisition module is used to acquire WSI image dataset, the WSI image dataset includes multiple WSI image data;

[0044] The preprocessing module is used to preprocess the WSI image dataset to obtain a preprocessed WSI image dataset.

[0045] A classification network training module is used to train a classification network based on a pathological basic model using the WSI image dataset, thereby obtaining the trained classification network based on the pathological basic model.

[0046] A feature acquisition module is used to acquire features extracted during the training of the classification network based on the pathological basis model.

[0047] The atlas construction module is used to construct a histopathological knowledge atlas by mapping the extracted features and each WSI image.

[0048] The feature index selection module is used to obtain a pre-selected set of feature indices suitable for the Transformer model based on the histopathology knowledge graph and each WSI image.

[0049] The Transformer model training module is used to train the Transformer model based on the pre-selected feature index group, thereby obtaining the trained Transformer model.

[0050] This application also provides a cancer prognostic prediction method based on histomics, the method comprising:

[0051] Obtain the classification network based on the pathological basis large model and the trained Transformer model after training using the tissue omics-based cancer prognostic model training method described above.

[0052] Obtain the WSI image to be classified;

[0053] The WSI image to be classified is input into the classification network based on the pathological basis large model to obtain the classification result;

[0054] Extract the feature index group of the WSI image to be classified based on the pre-selected feature index group;

[0055] The feature index set is input into the trained Transformer model to obtain the prediction results.

[0056] This application's histomics-based cancer prognostic model training method uses neural networks to extract information such as area, proportion, distribution, morphology, and molecular characterization differences at the tissue level. It employs a histomics framework to construct a multivariate salient feature attention model to predict overall survival (OS) risk, improving the efficiency of histopathological image analysis and providing a powerful tool for cancer diagnosis and valuable tumor microenvironment information for cancer research. Simultaneously, the multi-center collaborative strategy enhances the robustness of WSI data results generated from different laboratories and using different equipment. Attached Figure Description

[0057] Figure 1 This is a flowchart of the method of the present invention;

[0058] Figure 2 This is a flowchart of the WSI data preprocessing process of the present invention;

[0059] Figure 3 This is a flowchart of tissue classification for transfer learning in the large pathological model of the present invention.

[0060] Figure 4 A visualization of the tissue classification results of the transfer learning of the pathological big model of the present invention;

[0061] Figure 5 This is a schematic diagram illustrating the tissue quantitative feature extraction, screening, and prediction of the present invention;

[0062] Figure 6 This invention provides an evaluation metric for fine-grained cancer tissue classification prediction in a binary classification task.

[0063] Figure 7 This refers to the overall performance of the Hscore in predicting total survival at 12 and 24 months. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some, but not all, embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0065] In the description of this application, it should be understood that the terms "center", "longitudinal", "lateral", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting the scope of protection of this application.

[0066] like Figure 1 The training methods for cancer prognostic models based on histomics shown include:

[0067] Obtain a WSI image dataset, which includes multiple WSI image datasets;

[0068] The WSI image dataset is preprocessed to obtain a preprocessed WSI image dataset;

[0069] The WSI image dataset is used to train a classification network based on a large pathological model, thereby obtaining the trained classification network based on a large pathological model.

[0070] Acquire the features extracted during the training of the classification network based on the pathological basis large model;

[0071] A histopathology knowledge graph is constructed by mapping the extracted features and each WSI image.

[0072] Based on the histopathology knowledge graph and each WSI image, a pre-selected set of feature indicators suitable for the Transformer model is obtained;

[0073] The Transformer model is trained based on the pre-selected feature index group to obtain the trained Transformer model.

[0074] See Figure 2 In this embodiment, the preprocessing of the WSI image dataset to obtain a preprocessed WSI image dataset includes:

[0075] Step 21: Acquire WSI data and perform foreground selection on the WSI image based on digital image processing to obtain the foreground region (dilation and erosion, opening and closing operations, thresholding);

[0076] Step 22: Find the largest bounding box in the found foreground region to segment the foreground and determine it as the valid region;

[0077] Step 23: Match the effective domain with the WSI image to obtain the computational domain, and determine the coordinates of the tiles to be segmented;

[0078] Step 24: Cut the tiles according to their coordinates, and cut the tiles to a size of 256×256.

[0079] Step 25: Repeat steps 21-24 above until all WSI images have been processed.

[0080] In this embodiment, the background detection algorithm uses the Otsu method and the threshold method.

[0081] In this embodiment, the classification network process based on the large pathological model is as follows: Figure 3 As shown, the step of training a classification network based on a large pathological model using the WSI image dataset to obtain the trained classification network based on a large pathological model includes:

[0082] Read tiles images with 20x (0.5 μm / pixel), 10x (1 μm / pixel), and 5x (2 μm / pixel) pixels after cutting;

[0083] The read tiles image is color normalized.

[0084] The normalized tiles images are processed by flipping, rotating, affine transformation and adding random Gaussian noise to obtain the incremented tiles images. The incremented tiles images and the normalized tiles images together form the training data.

[0085] The training data is read using an 80% overlap sliding window strategy and fed into a classification network based on a large pathological model for training.

[0086] In this embodiment, the step of using an 80% overlapping sliding window strategy to read training data and inputting it into a classification network based on a large pathological model for training includes:

[0087] Perform the following operations for each round of training:

[0088] The training results of the basic pathology model are fed into the output network to achieve classification, and the cross-entropy function is used to measure the multi-classification error between the predicted results and the true labels.

[0089] If the error value is less than that of the previous model, the model parameters are saved. Training ends when the number of training rounds reaches a preset value or the model converges.

[0090] In this embodiment, the step of mapping and constructing a histopathological knowledge graph based on the extracted features and each WSI image includes:

[0091] Histopathological atlases were constructed through scanning and classification.

[0092] In this embodiment, the present application uses a toolkit to extract statistical features, morphological features, and texture features from WSI images and histopathological atlases, performs mathematical statistics on the features, selects appropriate features, and saves them as corresponding categories.

[0093] Specifically, obtaining a pre-selected set of feature indicators suitable for the Transformer model based on the histopathology knowledge graph and each WSI image includes:

[0094] The Mann-Whitney U test was used to identify features of death within a fixed time period. Features with a p-value less than 0.05 were selected as features significantly associated with overall survival, and features with repeated selections were merged.

[0095] The Pearson test was used to assess the linear correlation between features, and features with an R value greater than 0.95 were excluded.

[0096] After the linear correlation test, feature selection was performed using LASSO Cox regression with minimum absolute shrinkage and selection operator (LASSO). After feature selection, the retained features were used for modeling.

[0097] In this embodiment, the survival composite loss function of the Transformer model is as follows:

[0098] l sur-comp =l prop-risk +l Time =L sur-comp (r,t,e,timelist,P); where r

[0099] Let t represent the risk score vector, t represent the time vector, and e represent the event indicator vector; l sur-comp Represents the survival composite loss function; l prop-risk For proportional risk loss function; l Time The loss function for classifying time-dependent events.

[0100] In this embodiment, the proportional risk loss function formula is as follows:

[0101] Where π is the sequence of events ordered by time t, and t π(1) ≤t π(2) ≤…≤t π(N) risk π(i)Risk scoring; event π(i) For event indicators.

[0102]

[0103] in, σ(x) is the number of valid samples, and σ(x) is the sigmoid function.

[0104]

[0105] Where P is the total number of time steps, and p is the p-th time step.

[0106] In this embodiment, the classification network based on the pathological macro-model includes a pathological macro-model feature extractor and a pyramid attention network. A pre-trained macro-model network assists in extracting features from pathological image patches. An attention mechanism is employed to fuse feature information from different magnifications of the same target anchor point; multi-head attention is used to output the probability of variant tissues.

[0107] In this embodiment, multi-head attention is used to output multiple types of probability values:

[0108]

[0109] Where Q, K, and V represent the query, key, and value of the query, respectively, and they are derived from the encoding of the features. n_head represents the dimension of each attention head.

[0110] MultiAttention(Q,K,V)=Contect(head1,head2,…,head n W M ;

[0111] in, This represents the multilayer perceptron used for output.

[0112] A trained classification network based on a large pathological model is used for end-to-end classification of foreground regions in whole-slice images (WSI) via non-overlapping sliding. The output provides a probability estimate of all categories for the tissue (e.g., 12 categories in pancreatic cancer: cancer or high-grade PIN, low-grade PIN, normal tissue, acinar atrophy, tumor stroma, immune infiltration, lymph nodes, nerves, blood vessels, hematopoietic cell aggregation, necrosis, and others). The category of the target image patch is determined by the predicted activation function, with the highest-scoring category selected. Features extracted by the classifier are mapped to WSI images to construct a histopathological knowledge graph. The prediction results are visualized as follows: Figure 4 As shown. After feature extraction and retraining, the features extracted by the classifier match the gold standard region delineated by pathologists.

[0113] Spatial heterogeneity includes both intra- and inter-tissue heterogeneity. To quantify tissue spatial heterogeneity, radiomics features, known as histopathological features, need to be extracted from whole-section images of histopathological tissues. Histopathological features originate from different types of tissues and include statistical, morphological, and textural features. Statistical features include tissue area, tissue proportion, typology, inter-tissue proportion, typological proportion, inter-tissue typological proportion, and Shannon entropy. Morphological features include centroid, perimeter, elliptical similarity, minimum closed circle similarity, and convexity. Texture features are extracted using the gray-level co-occurrence matrix (GLCM) and local binary pattern (LBP). Texture features include contrast, homogeneity, stability, and quantified heterogeneity of the tissue. By extracting fine-grained features of different tumor types, precise tumor classification can be achieved.

[0114] Different histopathological features correspond to different prognoses. By capturing information between features using a Transformer-based pathomics Transformer survival network, risk prediction can be effectively performed. The risk prediction process is as follows: Figure 5 As shown.

[0115] To ensure the effectiveness of the model, statistical methods were used to analyze the prediction results, such as... Figure 6 , Figure 7 As shown.

[0116] This application also provides a tissue-based cancer prognostic model training device, the tissue-based cancer prognostic model training device comprising:

[0117] WSI image dataset acquisition module, the WSI image dataset acquisition module is used to acquire WSI image dataset, the WSI image dataset includes multiple WSI image data;

[0118] The preprocessing module is used to preprocess the WSI image dataset to obtain a preprocessed WSI image dataset.

[0119] A classification network training module is used to train a classification network based on a pathological basic model using the WSI image dataset, thereby obtaining the trained classification network based on the pathological basic model.

[0120] A feature acquisition module is used to acquire features extracted during the training of the classification network based on the pathological basis model.

[0121] The atlas construction module is used to construct a histopathological knowledge atlas by mapping the extracted features and each WSI image.

[0122] The feature index selection module is used to obtain a pre-selected set of feature indices suitable for the Transformer model based on the histopathology knowledge graph and each WSI image.

[0123] The Transformer model training module is used to train the Transformer model based on the pre-selected feature index group, thereby obtaining the trained Transformer model.

[0124] This application also provides a cancer prognostic prediction method based on histomics, the method comprising:

[0125] Obtain the classification network based on the pathological basis large model and the trained Transformer model after training using the tissue omics-based cancer prognostic model training method described above.

[0126] Obtain the WSI image to be classified;

[0127] The WSI image to be classified is input into the classification network based on the pathological basis large model to obtain the classification result;

[0128] Extract the feature index group of the WSI image to be classified based on the pre-selected feature index group;

[0129] The feature index set is input into the trained Transformer model to obtain the prediction results.

[0130] In this embodiment, inputting the feature index set into a trained Transformer model to obtain prediction results includes:

[0131] The extracted features are used as network inputs;

[0132] The Transformer's attention unit captures the attention between features for risk prediction;

[0133] Output a prediction of mortality risk.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and not to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for training a cancer prognostic model based on histomics, characterized in that, The training method for the cancer prognostic model based on histomatics includes: Obtain a WSI image dataset, which includes multiple WSI image datasets; The WSI image dataset is preprocessed to obtain a preprocessed WSI image dataset; The preprocessed WSI image dataset is used to train a classification network based on a large pathological model, thereby obtaining the trained classification network based on a large pathological model. Acquire the features extracted during the training of the classification network based on the pathological basis large model; A histopathology knowledge graph is constructed by mapping the extracted features and each WSI image. Based on the histopathology knowledge graph and each WSI image, a pre-selected set of feature indicators suitable for the Transformer model is obtained; The Transformer model is trained based on the pre-selected feature index group to obtain the trained Transformer model.

2. The cancer prognostic prediction method based on histomics as described in claim 1, characterized in that, The step of preprocessing the WSI image dataset to obtain a preprocessed WSI image dataset includes: The following processing is performed on each WSI image data: WSI data is acquired and foreground selection is performed on the WSI image based on digital image processing to obtain the foreground region; In the found foreground region, find the largest bounding box to segment the foreground and determine it as the valid region; The effective domain is mapped to the WSI image to obtain the computational domain, and the coordinates of the tiles to be segmented are determined. The tiles are cut according to their coordinates, and the size of the cut tiles is 256×256.

3. The cancer prognostic prediction method based on histomics as described in claim 2, characterized in that, The step of training a classification network based on a comprehensive pathological model using the WSI image dataset to obtain the trained classification network based on the comprehensive pathological model includes: Read the tile images at 20x, 10x, and 5x pixel sizes after they have been cut; The read tiles image is color normalized. The normalized tiles image is flipped, rotated, affine transformed, and random Gaussian noise is added to obtain the incremented tiles image. The incremented tiles image and the normalized tiles image together constitute the training data. The training data is read using an 80% overlap sliding window strategy and fed into a classification network based on a large pathological model for training.

4. The cancer prognostic prediction method based on histomics as described in claim 3, characterized in that, The step of using an 80% overlap sliding window strategy to read training data and then feeding it into a classification network based on a large pathological model for training includes: Perform the following operations for each round of training: The training results of the basic pathology model are fed into the output network to achieve classification, and the cross-entropy function is used to measure the multi-classification error between the predicted results and the true labels. If the error value is less than that of the previous model, the model parameters are saved. Training ends when the number of training rounds reaches a preset value or the model converges.

5. The cancer prognostic prediction method based on histomics as described in claim 4, characterized in that, The process of mapping and constructing a histopathological knowledge graph based on the extracted features and each WSI image includes: A histopathological atlas was constructed based on the features extracted from the model and the classification results.

6. The cancer prognostic prediction method based on histomics as described in claim 5, characterized in that, The step of obtaining a pre-selected set of feature indicators suitable for the Transformer model based on the histopathological knowledge graph and each WSI image includes: The Mann-Whitney U test was used to identify features of death within a fixed time period. Features with a p-value less than 0.05 were selected as features significantly associated with overall survival, and features with repeated selections were merged. The Pearson test was used to assess the linear correlation between features, and features with an R value greater than 0.95 were excluded. After the linear correlation test, the Cox regression with minimum absolute contraction and selection operator is used for feature selection. After feature screening, the retained features form a pre-selected feature index group.

7. The cancer prognostic prediction method based on histomics as described in claim 6, characterized in that, The survival composite loss function of the Transformer model is as follows: l sur-comp =l prop-risk +l Time =L sur-comp (r,t,e,timelist,P); Where r represents the risk score vector, t represents the time vector, and e represents the event indication vector; l sur-comp Represents the survival composite loss function; l prop-risk For proportional risk loss function; l Time The loss function for classifying time-dependent events.

8. The cancer prognostic prediction method based on histomics as described in claim 7, characterized in that, The formula for the proportional risk loss function is as follows: Where π is the sequence of events ordered by time t; risk π(i) Risk scoring; event π(i) For event indicators.

9. A training device for a cancer prognostic model based on histomics, characterized in that, The tissue-based cancer prognostic model training device includes: WSI image dataset acquisition module, the WSI image dataset acquisition module is used to acquire WSI image dataset, the WSI image dataset includes multiple WSI image data; The preprocessing module is used to preprocess the WSI image dataset to obtain a preprocessed WSI image dataset. A classification network training module is used to train a classification network based on a pathological basic model using the WSI image dataset, thereby obtaining the trained classification network based on the pathological basic model. A feature acquisition module is used to acquire features extracted during the training of the classification network based on the pathological basis model. The atlas construction module is used to construct a histopathological knowledge atlas by mapping the extracted features and each WSI image. The feature index selection module is used to obtain a pre-selected set of feature indices suitable for the Transformer model based on the histopathology knowledge graph and each WSI image. The Transformer model training module is used to train the Transformer model based on the pre-selected feature index group, thereby obtaining the trained Transformer model.

10. A cancer prognostic prediction method based on tissue omics, characterized in that, The tissue-based cancer prognostic prediction method includes: Obtain a classification network based on a large pathological model and a trained Transformer model trained using the histomics-based cancer prognostic model training method as described in any one of claims 1 to 8. Obtain the WSI image to be classified; The WSI image to be classified is input into the classification network based on the pathological basis large model to obtain the classification result; Extract the feature index group of the WSI image to be classified based on the pre-selected feature index group; The feature index set is input into the trained Transformer model to obtain the prediction results.

Images