Cancer survival prediction method based on pathological structured reasoning and genomics distillation

Abstract

The invention discloses a cancer survival prediction method based on pathological structured reasoning and genomics distillation, and belongs to the field of computer vision. The method comprises the following steps of: partitioning a full-slice pathology image to extract features, and grouping and coding genomics data according to functions to construct initial pathology and gene nodes; gradually screening related pathological areas by taking the gene nodes as guidance, and generating refined nodes; a heterogeneous graph structure is adaptively constructed based on feature affinity, and message fusion and prediction are carried out by a teacher network; the student network is only based on pathological nodes, the missing information is made up by generating pseudo gene nodes, multi-level knowledge distillation is carried out through attention, structure, representation and prediction signals of the teacher network, and accurate prediction in a single mode is achieved. According to the method, the problem of knowledge migration under structural information loss and gene deletion in multi-modal fusion is solved, accurate survival prediction can still be realized when only pathological images are used, and the method has important clinical application value.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and relates to a cancer survival prediction method, particularly a cancer survival prediction method based on pathological structured reasoning and genomic distillation. Background Technology

[0002] Cancer survival prediction is a crucial step in precision medicine and personalized treatment planning, aiming to accurately predict a patient's survival time or recurrence risk based on their clinical and biological data. Traditional survival analysis relies primarily on limited macroscopic indicators such as clinical stage, age, and gender, with room for improvement in predictive accuracy. With the development of high-throughput sequencing technology and digital pathology, survival prediction using whole-section pathological images and genomic data has become a cutting-edge research direction. Whole-section pathological images can capture histological information such as tumor micromorphology, cellular spatial structure, and tumor microenvironment at the gigapixel level; while genomic data (such as gene expression and mutation information) reveals the molecular mechanisms driving tumor development and progression. Therefore, fusing these two heterogeneous modalities of data holds promise for constructing more comprehensive and accurate prognostic models, providing strong support for clinical decision-making.

[0003] Based on the different input data, existing deep learning-based cancer survival prediction methods can be divided into two main categories: unimodal prediction methods and multimodal fusion prediction methods. Unimodal methods utilize only a single data source. For example, methods based on whole-section pathology images primarily learn prognostic-related morphological features from pathological images to predict survival; while genomics-based methods focus on analyzing multiple omics data to identify survival-related molecular markers for prediction. Multimodal methods comprehensively utilize the complementary prognostic information from both whole-section pathology images and genomics, employing strategies such as feature stitching, collaborative attention mechanisms, and optimal transport to capture deep associations between different modalities, achieving superior predictive performance. These studies have greatly advanced the development of computational pathology and cancer prognostic theory, and have demonstrated enormous potential in clinical translational research.

[0004] However, current methods for predicting cancer survival still face the following serious challenges: (1) Structural information loss in multimodal data fusion: Most existing methods are based on a flattened structure processing assumption, and perform indiscriminate fusion of biological information from two modalities. This approach ignores the spatial topology of tissues and the functional interactions between genes in whole-section pathological images, as well as the inherent structure and information priority within each modality, resulting in the discarding of important intramodal structures. In addition, current methods often fail to effectively separate key pathological-gene correlations from background noise during the fusion process, leading to significant information loss in the final representation.

[0005] (2) Difficulty in knowledge transfer due to missing genomic modalities: In actual clinical applications, genome sequencing is costly, complex, and time-consuming, resulting in a common problem of "missing modalities," where patients only have pathological images but lack corresponding genomic data. In this scenario, the performance of existing multimodal models will drop sharply or even become unusable. Although techniques such as knowledge distillation have been attempted to compensate for missing modalities, traditional distillation methods are difficult to effectively transfer the complex reasoning logic learned by multimodal models to single-modal models that rely solely on pathological images because there is no direct and explicit correspondence between pathological images and genomic data.

[0006] To address the aforementioned issues, there is an urgent need for a cancer survival prediction method that can achieve structured information fusion and efficient cross-modal knowledge transfer, in order to improve the applicability and prediction accuracy of the model in real clinical settings. Summary of the Invention

[0007] In view of this, the present invention provides a cancer survival prediction method based on pathological structured reasoning and genomic distillation to solve the problems of structural information loss and knowledge transfer difficulties under the absence of gene modalities in the existing technology of multimodal data fusion.

[0008] The technical solution adopted in this invention is as follows: This invention provides a cancer survival prediction method based on pathological structured reasoning and genomic distillation, comprising the following steps: The whole-slice pathological images were divided into blocks and the feature vectors of each image block were extracted to obtain the initial pathological node features; the genomic data were grouped according to biological function and the features of each group were extracted to obtain the initial gene node features. Guided by the initial gene nodes, we gradually screen and aggregate pathological regions related to gene function to generate refined pathological node features. Based on the feature affinity between the refined pathological node features and the initial gene node features, cross-modal edges and intramodal edges are adaptively constructed to form a heterogeneous graph structure; The teacher network performs message propagation and feature fusion on the heterogeneous graph structure to obtain multimodal fusion features, and outputs the survival prediction results of the teacher network. The student network takes only the initial pathological node features as input and generates pseudo-gene node features by mapping the pathological features to the latent space to make up for the missing gene modality information. Using the attention distribution, node embedding, edge weights, and prediction distribution of the teacher network as supervision signals, multi-level knowledge distillation is performed on the student network, and the survival prediction results of the student network are output.

[0009] Furthermore, the whole-slice pathological image is divided into blocks and the feature vectors of each image block are extracted to obtain the initial pathological node features. This includes: dividing the whole-slice pathological image into 512×512 pixel image blocks, using a pre-trained ResNet-50 model to extract a 1024-dimensional feature vector from each image block, and then further processing it into initial pathological node features with a dimension of 256.

[0010] Furthermore, the genomic data were grouped according to biological function and features of each group were extracted to obtain initial gene node features, including: Genomic data were grouped into six biological functions: tumor suppressor, tumorigenesis, protein kinase, cell differentiation, transcription, cytokines, and growth factors. Each group was encoded using a self-normalized neural network (SNN) to obtain initial gene node features.

[0011] Furthermore, guided by the initial gene nodes, pathological regions related to gene function are progressively screened and aggregated to generate refined pathological node features, including: Guided by the initial gene node, the first layer performs a global search to establish a rough mapping between gene features and pathological features; Each layer employs a sparse attention mechanism, which calculates attention only on a subset of highly relevant image patches selected from the previous layer, and multiplies the attention score of the current layer with the weight of the previous layer element by element to preserve persistently highly relevant regions. The aggregated data yielded a refined set of pathological node features.

[0012] Furthermore, based on the feature affinity between the refined pathological node features and the initial gene node features, cross-modal edges and intra-modal edges are adaptively constructed to form a heterogeneous graph structure, including: The affinity matrix is ​​obtained by calculating the normalized inner product of node features. Edges are established by taking the node pairs corresponding to the positive affinity scores, and gene-pathology cross-modal edges, gene-gene modal inner edges, and pathology-pathology modal inner edges are formed simultaneously to form a heterogeneous graph structure.

[0013] Furthermore, the teacher network performs message propagation and feature fusion on the heterogeneous graph structure to obtain multimodal fusion features, and outputs the survival prediction results of the teacher network, including: The heterogeneous graph is encoded through a heterogeneous graph transformer layer, and node features are updated through a message passing mechanism. The updated gene and pathology node representations are fused using cross-attention to obtain multimodal fusion features, which are then input into the prediction head to output survival prediction results.

[0014] Furthermore, the student network uses only the initial pathological node features as input, and generates pseudo-gene node features by mapping the pathological features to the latent space to compensate for missing gene modality information, including: Global pooling of initial pathological node features; After two layers of nonlinear transformation and projection onto a 512-dimensional latent space, pseudo-gene nodes are obtained; Using real gene nodes in the teacher network as the supervision target, semantic alignment is achieved by minimizing the embedding differences between the two.

[0015] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention introduces a structured heterogeneous graph fusion mechanism, which represents genes and refined pathological features together as nodes in a heterogeneous graph. Specifically, firstly, through a progressive gene-guided node refinement module, gene function information is used as prior knowledge to iteratively screen and aggregate the pathological features most relevant to specific gene functions from a massive amount of whole-slice pathological images, thereby effectively reducing background redundancy and focusing on key regions. On this basis, through a dynamic edge construction strategy based on feature affinity, pathological-gene associations within pathological modalities (such as spatial topology), within gene modalities (such as functional interactions), and across modalities are explicitly modeled. This structured fusion method not only preserves the intrinsic structural information of each modality but also profoundly reveals the complex interactions between them, thereby avoiding the information loss caused by traditional flattened structure fusion methods.

[0016] (2) This invention introduces a multi-level graph knowledge distillation strategy to construct a bridge for transferring multimodal knowledge from the teacher network to the student network, enabling the student network, which only uses pathological images, to inherit the complex reasoning ability of the multimodal teacher network. This strategy first trains the student network to generate "pseudo-gene" representations from pathological images through a gene feature reasoning module to solve the multimodal semantic gap problem. Then, by aligning with the teacher's internal attention distribution and nodes, the student network can inherit the complex intramodal and cross-modal dependency reasoning logic of the teacher network. In common clinical scenarios with only pathological image input, the student network can still reproduce or even surpass the predictive performance of the multimodal model, significantly improving the model's practicality and robustness when gene data is missing.

[0017] Experimental results show that the cancer survival prediction method based on pathological structured reasoning and genomic distillation proposed in this invention has an accurate predictive effect on cancer survival risk. Attached Figure Description

[0018] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1This is a flowchart of a cancer survival prediction method based on pathological structured reasoning and genomic distillation in an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the model structure of a cancer survival prediction method based on pathological structured reasoning and genomic distillation in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the progressive gene-guided node refinement module in an embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0023] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] like Figure 1 As shown, an embodiment of the present invention provides a cancer survival prediction method based on pathological structured reasoning and genomic distillation, which specifically includes the following steps: S1. The whole-slice pathological image is divided into blocks and the feature vector of each image block is extracted to obtain the initial pathological node features; the genomic data is grouped according to biological function and the features of each group are extracted to obtain the initial gene node features.

[0025] Specifically, whole-slice pathological image datasets were obtained from The Cancer Genome Atlas project, including datasets for bladder urothelial carcinoma (BLCA), lung adenocarcinoma (LUAD), endometrial cancer (UCEC), breast cancer (BRCA), glioblastoma, and low-grade glioma (GBMLGG). These images typically have a resolution of billions of pixels, making them unsuitable for direct input into deep learning models. To facilitate subsequent analysis, this invention uses whole-slice pathological images magnified 20 times and segmented them into non-overlapping 512×512 pixel image patches. A pre-trained ResNet-50 model is used to extract a 1024-dimensional feature vector from each image patch. Each 1024-dimensional feature vector is further processed into a set of feature representations: (1) in, The number of image patches extracted from whole-section pathological images. The dimension of the feature vector (here, ); Simultaneously, genomic data corresponding to the whole-slice pathological image dataset was obtained from The Cancer Genome Atlas project. This genomic data includes molecular features from multiple omics perspectives, such as gene expression profiles from the transcriptome (RNA sequencing, RNASeq) and copy number variations (CNVs) from the genome. This invention preprocessed and screened the raw genomic data, retaining only genes relevant to subsequent analysis. The selected genes were then divided into six biologically significant functional groups: tumor suppressor, tumorigenesis, protein kinase, cell differentiation, transcription, and cytokines and growth factors. Each functional genome was encoded by a dedicated self-normalized neural network (SNN), resulting in a genomic feature representation. (2) in, This indicates the number of functional genomes.

[0026] S2, guided by the initial gene node, gradually screens and aggregates pathological regions related to gene function to generate refined pathological node features.

[0027] Specifically, genomic features are divided into six groups according to function, which can be directly used as six genomic nodes. In contrast, pathological images contain a large number of image patches. Treating all image patches as independent nodes directly would lead to structural imbalance and information redundancy, making it difficult to separate key pathological-gene correlations from background noise. Therefore, this invention uses the initial gene nodes as a guide and employs a multi-layer sparse attention mechanism to iteratively screen and aggregate refined pathological nodes that are highly correlated with the functional groups of each gene from a massive number of initial pathological nodes, thereby effectively reducing information redundancy and separating key pathological-gene correlations. like Figure 3 As shown, the first layer performs a global search to establish genomic features. and pathological features A rough mapping between them; Subsequently, a sparse attention mechanism is applied to each layer to narrow the scope of analysis, focusing only on the most valuable subset of image patches filtered out by the previous layer. To ensure that only positions with consistently strong relevance are reinforced, while positions that are only focused on in a single layer are suppressed, the attention weights of layer l>1 are calculated by the element-wise product of the attention score of the current layer and the weight of the previous layer:

[0028]

[0029] (3) in, For the first Attention graph of the layer Indicates the first The selected layer %Index to the image patch For the corresponding attention weights associated with these indices, Indicates only Above calculation , ⊙ represents element-wise multiplication. and Initialize as a matrix of all 1s; After the final refinement layer, a highly focused attention map is generated. This graph encodes weights for aggregating raw image patch features into a biologically meaningful and compact pathological representation; specifically, for For each of the gene functional groups, the corresponding image patch features are calculated. A weighted sum is used to obtain a condensed feature, and the condensed features obtained from different gene functional groups are used as pathological nodes of the isomorphism map. : (4) The number of layers α in the progressive gene-guided node refinement module is adaptively adjusted according to the complexity of different datasets. For some datasets, α is set to 3, while for other more complex datasets, α is set to 5. For example, for the bladder urothelial carcinoma (BLCA), lung adenocarcinoma (LUAD), and endometrial cancer (UCEC) datasets, α is set to... For the Breast Cancer (BRCA) and Glioblastoma and Low-Grade Glioblastoma (GBMLGG) datasets, a deeper configuration is used, setting... .

[0030] S3, based on the feature affinity between the refined pathological node features and the initial gene node features, adaptively constructs cross-modal edges and intramodal edges to form a heterogeneous graph structure.

[0031] Specifically, by adaptively constructing the following three types of edges (cross-modal edges between gene-pathology nodes, intramodal edges between gene-gene nodes, and intramodal edges between pathology nodes), a heterogeneous graph structure that can simultaneously represent intramodal functional interactions and intermodal associations is constructed.

[0032] like Figure 2 As shown, three affinity matrices are defined to capture inter-modal and intra-modal relationships: gene-pathology. Pathology - Pathological and gene-gene The affinity matrix is ​​obtained by calculating the normalized inner product of node features: (5) in, and They represent the modes respectively. and The The and the first The feature vector of each node; Based on the calculated affinity matrix, this invention defines a specific edge set for each relationship type. To ensure that only meaningful connections enter the graph structure, potential edges are pruned by retaining only edge pairs with positive affinity scores. This step filters out weak connections or non-existent relations, resulting in a more focused and robust graph topology. Formally, the edge set for each relation type is defined as: (6) S4, the teacher network performs message propagation and feature fusion on the heterogeneous graph structure to obtain multimodal fusion features, and outputs the survival prediction results of the teacher network.

[0033] Specifically, heterogeneous graphs After construction, it is encoded through a multi-layered Heterogeneous Graph Transformer (HGT) layer, and message passing operations are performed: each node in each layer All methods aggregate information from adjacent nodes within and between modes to update and enrich their own representations. (7) in, and They are nodes In the Layer and first The layer's characteristics, where N(v) represents the neighborhood of node v; After the message is transmitted, the updated genome node is obtained. and pathological node representation Subsequently, a multi-head attention mechanism is used to integrate information from the two modalities to generate fused features. : (8) Where (·)z represents the first z. The output vector of the cross-attention mechanism. Finally, the obtained fusion features The data is fed into the classification head to obtain the final teacher network survival prediction results.

[0034] S5, the student network takes only the initial pathological node features as input and generates pseudo-gene node features by mapping the pathological features to the latent space, which is used to make up for the missing gene modality information.

[0035] By mapping the global pooling features of the initial pathological nodes to a structured latent space, pseudo-gene nodes are inferred and generated that semantically aim to approximate the real gene nodes of the teacher network. By minimizing the distance between the generated pseudo-gene nodes and the real gene nodes of the teacher network, students are able to infer molecular features from visual data alone, without requiring genomic data.

[0036] Specifically, the student network first performs global average pooling on the initial pathological node features to obtain the global feature embedding of the pathological image. This is then projected into a structured latent space to infer and generate pseudo-gene tokens that semantically approximate the real gene nodes of the teacher network. This transformation process is achieved through a two-layer multilayer perceptron with nonlinear activation and normalization functions, and the output is reshaped into... A sequence of token vectors: (9) in, , It is the ReLU activation function. Representation layer normalization, and For trainable parameters, , For the corresponding bias term; To ensure that these inferred tokens have semantic meaning, the generation process is supervised using the real gene embedding G of the teacher network as the target, and a gene-level semantic alignment loss is introduced. By minimizing the difference between the generated pseudo-gene nodes and the teacher's real gene nodes: (10) Complete the gene modality completion.

[0037] S6 uses the attention distribution, node embedding, edge weights, and prediction distribution of the teacher network as supervision signals to perform multi-level knowledge distillation on the student network and outputs the survival prediction results of the student network.

[0038] Specifically, based on the aforementioned gene-level semantic alignment, the complex reasoning logic of the teacher network is further distilled using reasoning logic distillation loss, which is composed of... , and composition: Attention Alignment: By aligning with the internal attention distribution of the teacher network, the student network is guided to imitate the visual focus of the teacher network during the refinement of pathological nodes, thereby learning and reproducing the spatial reasoning logic of the teacher network. This is achieved through attention alignment loss in order to imitate the teacher's visual focus. Force the attention of the top k most-watched image patches in both networks to be consistent: (11) in, and The detailed weight distribution of nodes represents the teacher network and the student network, respectively. Structural alignment: By aligning node embeddings and edge weights in heterogeneous graph structures, the student network learns the intramodal and intermodal dependencies encoded by the teacher network, thereby reproducing its relational reasoning logic. This facilitates the transfer of the teacher's relational reasoning ability. Structural alignment loss... Aligning dynamically integrated node embeddings and attention-based edge weights in heterogeneous graph structures: (12) in, , These are the node embeddings of nodes in the teacher network and the student network, respectively. , These represent the corresponding edge weights; Representation-Prediction Alignment: By aligning the final fused feature representations and predicted distributions of the student and teacher networks, this ensures that the student network's decision-making basis for making the final survival risk judgment is consistent with that of the teacher network. To ensure consistency at the outcome level, representation-prediction alignment loss is used. Matching the fusion feature representation between student and teacher networks and softening the survival risk distribution: (13) in, , These represent high-level fusion feature embeddings for the teacher network and the student network, respectively. and These represent the output distributions of their predictions of softened survival risk. Represents cosine similarity. Indicates the Kullback-Leibler divergence; Teachers' Internet Use Survival Prediction Loss Optimize: Given a multimodal input, a teacher network predicts a discrete risk vector. And recursively calculate the survival probability : (14) in, Indicates a deleted state. The observed time of the event or the time of censoring; The student network is trained under the same survival objective, while additionally weighted multiple distillation losses are incorporated to capture the reasoning behavior of the teacher network. The total loss formula is expressed as: (15) The loss weight coefficient is uniformly set to .

[0039] The training of student networks is guided by a multi-level graph knowledge distillation strategy, and the survival prediction results of the student networks are output.

[0040] This invention is divided into two stages: teacher network training and student network training. First, the teacher network is trained independently until its performance converges. Then, the parameters of the trained teacher network are fixed, and its output internal knowledge is used to guide the training of the student network. To ensure the generalization ability of the model and the robustness of the results, the overall training process adopts a five-fold cross-validation strategy. In each fold, the data is divided into a training set and a validation set in a 4:1 ratio.

[0041] This invention is implemented in the PyTorch framework, where the ADAMW optimizer is used during the training phase, and the learning rate is... And the batch size is 1.

[0042] Extensive experiments demonstrate that the method in this invention has a competitive advantage over 16 state-of-the-art survival prediction methods on five public datasets.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cancer survival prediction method based on pathological structured reasoning and genomic distillation, characterized in that, include: The whole-slice pathological image is divided into blocks and the feature vectors of each image block are extracted to obtain the initial pathological node features; Genomic data are grouped according to biological function and features of each group are extracted to obtain initial gene node features; Guided by the initial gene node features, pathological regions related to gene function are screened and aggregated to generate refined pathological node features; Based on the feature affinity between the refined pathological node features and the initial gene node features, cross-modal edges and intramodal edges are adaptively constructed to form a heterogeneous graph structure; The teacher network performs message propagation and feature fusion on the heterogeneous graph structure to obtain multimodal fusion features, and outputs the survival prediction results of the teacher network. The student network takes only the initial pathological node features as input and generates pseudo-gene node features by mapping the pathological features to the latent space to make up for the missing gene modality information. Using the attention distribution, node embedding, edge weights, and prediction distribution of the teacher network as supervision signals, multi-level knowledge distillation is performed on the student network, and the survival prediction results of the student network are output.

2. The cancer survival prediction method based on pathological structured reasoning and genomic distillation according to claim 1, characterized in that, The whole-slice pathological image is divided into blocks and the feature vectors of each image block are extracted to obtain the initial pathological node features. This includes: dividing the whole-slice pathological image into 512×512 pixel image blocks, using a pre-trained ResNet-50 model to extract a 1024-dimensional feature vector from each image block, and then further processing it into initial pathological node features with a dimension of 256.

3. The cancer survival prediction method based on pathological structured reasoning and genomic distillation according to claim 1, characterized in that, Genomic data were grouped according to biological function, and features were extracted from each group to obtain initial gene node features, including: Genomic data were grouped into six biological functions: tumor suppressor, tumorigenesis, protein kinase, cell differentiation, transcription, cytokines, and growth factors. Each group was encoded using a self-normalized neural network (SNN) to obtain initial gene node features.

4. The cancer survival prediction method based on pathological structured reasoning and genomic distillation according to claim 1, characterized in that, Guided by initial gene node features, pathological regions related to gene function are progressively screened and aggregated to generate refined pathological node features, including: Guided by the initial gene node, the first layer performs a global search to establish a rough mapping between gene features and pathological features; Each layer employs a sparse attention mechanism, which calculates attention only on a subset of highly relevant image patches selected from the previous layer, and multiplies the attention score of the current layer with the weight of the previous layer element by element to preserve persistently highly relevant regions. The aggregated data yielded a refined set of pathological node features.

5. The cancer survival prediction method based on pathological structured reasoning and genomic distillation according to claim 1, characterized in that, Based on the feature affinity between the refined pathological node features and the initial gene node features, cross-modal edges and intramodal edges are adaptively constructed to form a heterogeneous graph structure, including: The affinity matrix is ​​obtained by calculating the normalized inner product of node features. Edges are established by taking the node pairs corresponding to the positive affinity scores, and gene-pathology cross-modal edges, gene-gene modal inner edges, and pathology-pathology modal inner edges are formed simultaneously to form a heterogeneous graph structure.

6. The cancer survival prediction method based on pathological structured reasoning and genomic distillation according to claim 1, characterized in that, The teacher network performs message propagation and feature fusion on the heterogeneous graph structure to obtain multimodal fused features, and outputs the survival prediction results of the teacher network, including: The heterogeneous graph is encoded through a heterogeneous graph transformer layer, and node features are updated through a message passing mechanism. The updated gene and pathology node representations are fused using cross-attention to obtain multimodal fusion features, which are then input into the prediction head to output survival prediction results.

7. The cancer survival prediction method based on pathological structured reasoning and genomic distillation according to claim 1, characterized in that, The student network takes only initial pathological node features as input and generates pseudo-gene node features by mapping these features to a latent space. These pseudo-gene node features are used to compensate for missing gene modality information, including: Global pooling of initial pathological node features; After two layers of nonlinear transformation and projection onto a 512-dimensional latent space, pseudo-gene nodes are obtained; Using real gene nodes in the teacher network as the supervision target, semantic alignment is achieved by minimizing the embedding differences between the two.

Images