A tumor survival prediction system based on quaternion alignment of multi-modal data
By performing rotation alignment and fusion operations in quaternion space, the problem of alignment and fusion of multimodal data is solved, enabling efficient cancer survival prediction and risk stratification, and improving prediction accuracy and interpretability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUN YAT SEN UNIVERSITY CANCER CENTER (CANCER HOSPITAL AFFILIATED TO SUN YAT SEN UNIVERSITY CANCER RESEARCH INSTITUTE OF SUN YAT SEN UNIVERSITY)
- Filing Date
- 2026-04-03
- Publication Date
- 2026-05-29
Smart Images
Figure CN122117337A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical artificial intelligence technology, specifically relating to a tumor survival prediction system based on quaternion-aligned multimodal data. Background Technology
[0002] Cancer survival analysis is one of the most fundamental methods in clinical oncology, quantifying patient survival outcomes and supporting risk stratification. This analysis provides clinicians with crucial insights into disease progression patterns and treatment effectiveness, helping to develop personalized treatment strategies tailored to individual patient characteristics. In recent years, the rise of artificial intelligence (AI) has further propelled the development of survival analysis. Leveraging its powerful capabilities, AI can learn feature representations from multiple clinical data modalities, including histopathological sections, imaging data, and gene expression profiles. However, most current AI-based survival analysis methods typically rely on a single data modality, limiting their ability to capture the complex biological characteristics of cancer progression.
[0003] In clinical practice, making effective diagnostic and treatment decisions often requires integrating multiple data sources to form a comprehensive treatment plan. Multimodal fusion, by integrating multimodal information such as histopathological images and genomic data, can significantly improve prognostic accuracy and the accuracy of clinical outcome prediction. However, due to the inherent structural differences between different modalities, the data alignment and fusion process faces many challenges. For example, histopathological images exhibit complex spatial structures at multiple scales, while genomic data consists of high-dimensional molecular features with drastically different statistical properties. These inherent differences necessitate the design of complex feature space alignment strategies to achieve meaningful data fusion and subsequent analysis.
[0004] To address these challenges, researchers have explored various alignment strategies, including Optimal Transport (OT)-based methods and attention-based methods. OT methods aim to achieve global feature consistency by minimizing the transport cost from one modality to another. However, these methods often overemphasize global alignment, potentially neglecting local biological details crucial for understanding the tumor microenvironment. In contrast, attention-based methods capture fine-grained relationships between histopathological image features and genomic features through cross-attention networks, excelling in modeling local features. However, they still struggle to balance coarse-grained and fine-grained alignment of multimodal features. Maintaining multi-scale feature synergy while considering both global and local information remains a critical issue to be addressed in current multimodal fusion research.
[0005] Given the intrinsic biological link between cell density and spatial distribution and gene expression patterns in histopathological images, spatial transcriptomics studies have further revealed that gene expression profiles can be clearly mapped to specific histological structures. This biological consistency indicates that although the histopathological and genomic modalities differ in their representational forms, they jointly encode the shared tumor microenvironment state within a heterogeneous representational space. Therefore, in the feature fusion process, it is necessary not only to focus on the aggregation of feature amplitudes but also to emphasize the directional alignment of feature vectors between modalities to better reflect their synergistic relationship in biological essence.
[0006] Therefore, how to fully capture local biological details while preserving the global structural consistency of multimodal data, and design efficient and robust feature alignment and fusion strategies, has become a key issue in current deep learning-based multimodal survival analysis research. Summary of the Invention
[0007] In view of the above, the purpose of this invention is to provide a tumor survival prediction system based on quaternion-aligned multimodal data. By integrating histopathological images and genomic data, the system utilizes rotation alignment operations in quaternion space to achieve directional consistency of multimodal features and introduces a learnable rotation cost matrix to balance alignment degree and modal information preservation. Based on this, an attention mechanism is combined to extract deep interaction relationships between modalities, thereby improving the accuracy and robustness of cancer survival prediction and providing technical support for personalized diagnosis and treatment decisions in precision medicine.
[0008] To achieve the above-mentioned objectives, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a tumor survival prediction system based on quaternion-aligned multimodal data, comprising: The quaternion encoding module is used to map multimodal tumor feature data to corresponding quaternions; The quaternion rotation alignment module is used to calculate the difference quaternion based on the quaternions of each modal feature, and extract the rotation axis and rotation angle from the difference quaternion to construct the rotation quaternion. The rotation quaternion is used to rotate the quaternion of at least one modal feature to achieve the alignment of feature directions between modalities. The multimodal feature fusion module is used to fuse the quaternions of aligned modal features to extract intermodal interaction relationships and generate a global representation. The risk prediction output module is used to output tumor survival risk prediction results based on the global characterization.
[0009] Preferably, the multimodal tumor feature data includes at least pathological image features and genomic features. The pathological image features are obtained by feature extraction from the histopathological images of tumor patients, and the genomic features are obtained by dimensionality reduction and encoding of a set of tumor-related genes.
[0010] Preferably, the quaternion encoding module employs a quaternion encoder to map each modal feature to the quaternion space and encode them into corresponding quaternions.
[0011] Preferably, in the quaternion rotation alignment module, the step of calculating difference quaternions based on quaternions of each modal feature, and extracting rotation axes and rotation angles from the difference quaternions to construct rotation quaternions, includes: Select the quaternion of any modal feature as the benchmark, and calculate the difference quaternion between the quaternions of other modal features and the quaternion of the benchmark modal feature; The rotation angle is extracted from the real part of the differential quaternion, the rotation axis is extracted from the imaginary part, and a learnable rotation cost matrix is introduced to correct the rotation angle. Construct a rotation quaternion based on the rotation axis and the corrected rotation angle.
[0012] Preferably, the multimodal feature fusion module employs an attention-based feature fusion network to map the quaternions of aligned modal features to query matrices, key matrices, and value matrices, respectively. It captures the feature interaction relationships between modalities through attention calculation and uses a gated attention mechanism to weight the modalities based on their importance contributions, ultimately fusing the modalities to generate a global representation.
[0013] Preferably, the risk prediction output module adopts a fully connected layer, which includes at least two linear layers: the first linear layer is used to perform dimensionality reduction processing on the global representation of the input, and the second linear layer is used to receive the features after dimensionality reduction in the first layer and further calculate and output the patient's survival risk probability.
[0014] Preferably, a joint loss function is used to train a tumor survival prediction model that includes a quaternion encoding module, a quaternion rotation alignment module, a multimodal feature fusion module, and a risk prediction output module. The joint loss function includes a survival analysis loss for supervising survival prediction and a rotation loss for constraining the consistency of quaternion rotation alignment.
[0015] Secondly, the present invention provides a tumor survival prediction method based on quaternion-aligned multimodal data, implemented using the aforementioned tumor survival prediction system based on quaternion-aligned multimodal data, comprising the following steps: The quaternion encoding module is used to map multimodal tumor feature data into corresponding quaternions; The quaternion rotation alignment module calculates the difference quaternion based on the quaternion of each modal feature, and extracts the rotation axis and rotation angle from the difference quaternion to construct the rotation quaternion. The rotation quaternion is then used to rotate the quaternion of at least one modal feature to achieve the alignment of feature directions between modalities. The quaternions of aligned modal features are fused using a multimodal feature fusion module to extract intermodal interaction relationships and generate a global representation. The risk prediction output module outputs tumor survival risk prediction results based on the global characterization.
[0016] Thirdly, an electronic device provided by an embodiment of the present invention includes a memory and one or more processors. The memory is used to store a computer program, and the processor is used to implement the above-described tumor survival prediction method based on quaternion-aligned multimodal data when executing the computer program.
[0017] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a computer, implements the aforementioned tumor survival prediction method based on quaternion-aligned multimodal data.
[0018] Compared with the prior art, the beneficial effects of the present invention include at least the following: (1) Multimodal Feature Alignment and Fusion: This invention achieves precise alignment of histopathological image features and genomic features in the instance-level vector direction by using rotation alignment operations in quaternion space, while preserving the inherent amplitude of each modality feature. This effectively solves the problem of inconsistent feature directions caused by the heterogeneity of multimodal data. Furthermore, a rotation loss function and a rotation cost matrix are introduced to optimize the rotation process, reduce information distortion between modalities, and preserve modality specificity. Based on this, deep interaction relationships between modalities are further extracted to generate a biologically meaningful global representation, significantly improving the quality of multimodal fusion and the accuracy of downstream survival prediction.
[0019] (2) Enhanced biological interpretability: The aligned multimodal feature representation generated by this invention has good interpretability, revealing the biological interaction between histopathological features and genetic features, providing clinicians with understandable predictive evidence. Extensive validation on multiple public and private nasopharyngeal carcinoma datasets has demonstrated the significant improvement in prognostic performance and model stability of this invention, effectively supporting the identification and risk stratification of high-risk patient groups for various cancer types, and providing a reliable basis for the formulation of personalized monitoring and treatment strategies in precision medicine. Attached Figure Description
[0020] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 This is a schematic diagram of the structure of a tumor survival prediction system based on quaternion-aligned multimodal data provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the framework of the tumor survival prediction model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the framework of the quaternion encoding module and the quaternion rotation alignment module provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of the framework of the multimodal feature fusion module provided in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the tumor survival prediction method based on quaternion-aligned multimodal data provided in this embodiment of the invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.
[0023] The inventive concept of this invention is as follows: Addressing the problems of insufficient capture of tumor biological complexity, inadequate feature alignment between modalities, and lack of biological interpretability in existing survival prediction methods due to single-modality analysis, this invention provides a tumor survival prediction system based on quaternion-aligned multimodal data. By mapping histopathological image features and genomic features to a quaternion space, quaternion rotation alignment operations are used to achieve consistent vector direction alignment of multimodal features at the instance level. A learnable rotation cost matrix is introduced to balance the degree of alignment with the preservation of modality-specific information, thereby achieving refined alignment while retaining the biological specificity of the modality. Furthermore, this is further fused to generate a unified global representation with biological significance, ultimately achieving accurate cancer survival risk prediction and interpretable patient risk stratification.
[0024] like Figure 1As shown, this embodiment provides a tumor survival prediction system 100 based on quaternion-aligned multimodal data, including: a quaternion encoding module 110, a quaternion rotation alignment module 120, a multimodal feature fusion module 130, and a risk prediction output module 140. A tumor survival prediction model is constructed based on these four modules, and the overall model architecture is as follows. Figure 2 As shown.
[0025] In this embodiment, the quaternion encoding module 110 is used to map the multimodal tumor feature data to corresponding quaternions. Specifically, multimodal tumor feature data is first acquired. This invention is trained based on the TCGA public dataset. For each tumor patient, the system uses two types of multimodal data: pathological image data: the entire pathological slide is divided into several patches, and the ResNet-50 model is used to extract features from each patch to obtain high-dimensional pathological image features. Genomic data: Key tumor-related genes (such as oncogenes, transcription factors, tumor suppressor genes, cytokines, and growth factors) were selected, and gene expression features were extracted using a fully connected neural network (SNN) to obtain high-dimensional genomic features. All features are normalized to ensure a consistent distribution of the input data, facilitating subsequent feature fusion.
[0026] like Figure 3 As shown, the system will use pathological image features and genomic characteristics The inputs are fed into the quaternion encoder, mapped to the quaternion space, and the corresponding quaternion representations are obtained. and The quaternion encoder specifically comprises multiple parallel linear mapping branches and nonlinear transformation modules. Its core objective is to embed real-valued features into the quaternion space for structured representation. For the input features, the quaternion encoder generates the real part of the quaternion through four independent linear transformations. and three imaginary parts The four components are then concatenated along the feature dimension to obtain a quaternion representation. , The quaternion representation, representing the feature dimension, is further processed by a non-linear activation function and regularization to form the final encoding result. In a multi-layered structure, the above mappings can be cascaded and stacked to achieve feature abstraction from low to high levels. In this way, the original features are projected into multiple subspaces, thus preserving input information to some extent. Simultaneously, the subspaces corresponding to different components provide a representational basis for potential semantic partitioning, enabling the model to learn differentiated representations between components. Furthermore, this quaternion representation provides a structured foundation for subsequent cross-component interactions based on quaternion operations, thereby supporting further association modeling and cross-modal alignment.
[0027] In this embodiment, the quaternion rotation alignment module 120 is used to calculate differential quaternions based on the quaternions of each modal feature, and extract the rotation axis and rotation angle from the differential quaternions to construct a rotation quaternion. The rotation quaternion is then used to rotate the quaternion of at least one modal feature to achieve alignment of feature directions between modalities. Specifically, after the multimodal features are projected into the quaternion space by the quaternion encoder, the system uses the quaternion rotation alignment module to achieve spatial alignment of pathological image features and genomic features, ensuring geometric consistency.
[0028] like Figure 3 As shown, firstly, the quaternion representation of pathological image features is calculated. Quaternion representation of genomic features Difference quaternions This represents the directional difference between two modal features in quaternion space, and the rotation axis is extracted from the difference quaternion. and rotation angle The rotation axis characterizes the optimal transformation direction followed during feature alignment, while the rotation angle reflects the magnitude of change required for alignment.
[0029] To optimize the rotation process, a learnable rotation cost matrix is introduced. Each element in the rotation cost matrix Indicates the first The rotation angle information of the first feature dimension is related to the second feature dimension. The contribution weights of each feature dimension to the final rotation angle are used to adjust the alignment cost between different modalities, preserving modality specificity. The rotation cost matrix is adjusted for the rotation angle as follows: Based on the axis of rotation and the corrected rotation angle. Construct rotation quaternions ,in, The real part represents the rotation. Represents the imaginary part of rotation. Denotes the axis of rotation, and satisfies , These represent the components of the rotation axis in the x, y, and z directions, respectively. This is achieved using rotation quaternions. Rotate and align the quaternion representations of pathological image features: , Quaternions representing genomic features after orientation alignment of pathological image features.
[0030] In this embodiment, the multimodal feature fusion module 130 is used to fuse the quaternions of the aligned modal features to extract intermodal interaction relationships and generate a global representation. Specifically, as shown below... Figure 4 As shown, the quaternion representation of the aligned pathological image features Quaternion representation of genomic features The images were flattened separately to obtain aligned pathological image features and aligned genomic features. and The system employs a multimodal feature fusion module for multimodal feature fusion: [The module then...] and Global representations of pathological modalities were generated using the Transformer attention mechanism. and global characterization of genome modalities Then, the gating mechanism is used to calculate the fusion weight, and the multimodal features are weighted and spliced to highlight important features, optimize the interaction effect between modalities, and obtain the final unified global representation.
[0031] In this embodiment, the risk prediction output module 140 is used to output tumor survival risk prediction results based on the global representation. Specifically, the system inputs the fused unified global representation into the fully connected layer and finally outputs the patient's survival risk probability (such as low risk, medium risk, and high risk). This probability result is used by clinicians to assess the patient's risk stratification and treatment plan.
[0032] The training process for the tumor survival prediction model is as follows: (1) Extract pathological images and genomic data from cancer patients to form multimodal data; (2) Generate labels based on the patient's survival status (survival / death), which are 0 or 1 respectively; (3) Divide the data into training set and validation set, and normalize the pathological image data and genomic data. Input a batch of data in each iteration; (4) Use the quaternion encoding module and the quaternion rotation alignment module to align the multimodal data, and output the global representation through the multimodal feature fusion module; (5) The global representation input risk prediction output module is classified and trained, the model parameters are iteratively updated until convergence, and an early stopping mechanism is adopted to terminate the training in advance when the performance of the validation set continuously declines, and the optimal model is saved.
[0033] During model training, a joint loss function is used, including a survival analysis loss for supervising survival prediction and a rotation loss for constraining quaternion rotation alignment consistency, to ensure prediction performance and feature alignment effectiveness.
[0034] Among them, survival analysis loss Constructed based on negative log-likelihood loss: , in, Indicates whether an event has occurred. Indicates the sample observation time. Indicates input features, Represents the survival function. Indicates the number of samples. Indicates the sample index. This represents the risk function.
[0035] Rotational loss This is used to constrain the alignment of different modalities (pathology, genes) in quaternion space. First, it uses two quaternion features... and Dot product calculation of rotation angle Introducing a learnable cost matrix The rotation angle is weighted to obtain Ultimate rotational loss Defined as: , in, Indicates the first The weighted rotation angles of pathological and genetic modalities in quaternion space for each sample. Indicates the number of samples.
[0036] The final joint loss is ,in The weights for the rotation loss term.
[0037] In summary, the multimodal survival analysis system of this invention can integrate pathological images and genomic data to accurately predict the survival risk of cancer patients. The system was tested on five publicly available TCGA datasets (including bladder cancer (BLCA), breast cancer (BRCA), endometrial cancer (UCEC), glioma (GBMLGG), and lung adenocarcinoma (LUAD)) and one private nasopharyngeal carcinoma (NPC) dataset. All data included paired whole-slice pathological images, genomic data, and corresponding survival information. Five-fold cross-validation was used to ensure the stability of the results. Regarding comparative methods, several representative baseline models were covered: unimodal methods included SNN and SNNTrans based on gene data, and AttnMIL, DeepAttnMISL, CLAM-MB, CLAM-SB, and TransMIL based on pathological images; multimodal methods included Porpoise, MCAT, MOTCat, SurvPath, and PIBD, thus achieving a comprehensive and fair performance comparison.
[0038] Table 1
[0039] As shown in Table 1, experimental results demonstrate that our proposed method achieves state-of-the-art performance across all datasets, with an average C-index of 0.710, representing an improvement of approximately 0.2%–4.7% compared to the current best multimodal method. Furthermore, it exhibits low variance across different datasets, demonstrating good stability. The key to this performance improvement lies in two aspects: firstly, the quaternion-based rotation alignment mechanism explicitly models the directional consistency between pathological and genetic modalities, effectively capturing shared information; secondly, the non-commutativity of quaternion multiplication enhances the model's ability to express complex cross-modal nonlinear relationships. Through the synergistic effect of these two factors, the model achieves more accurate and interpretable survival risk predictions while ensuring no loss of modality-specific information.
[0040] Based on the same inventive concept, such as Figure 5 As shown, this embodiment of the invention also provides a tumor survival prediction method based on quaternion-aligned multimodal data, including the following steps: S1, using the quaternion encoding module to map the multimodal tumor feature data into corresponding quaternions; S2, the quaternion rotation alignment module calculates the difference quaternion based on the quaternion of each modal feature, and extracts the rotation axis and rotation angle from the difference quaternion to construct the rotation quaternion. The rotation quaternion is used to rotate the quaternion of at least one modal feature to achieve the alignment of feature directions between modalities. S3 utilizes the multimodal feature fusion module to fuse the quaternions of aligned modal features to extract intermodal interaction relationships and generate a global representation; S4 utilizes the risk prediction output module to output tumor survival risk prediction results based on the global characterization.
[0041] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, including a memory and one or more processors, wherein the memory is used to store a computer program, and the processor is used to implement the above-described tumor survival prediction method based on quaternion-aligned multimodal data when executing the computer program.
[0042] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a computer, implements the aforementioned tumor survival prediction method based on quaternion-aligned multimodal data.
[0043] It should be noted that the tumor survival prediction method, electronic device, and computer-readable storage medium based on quaternion-aligned multimodal data provided in the above embodiments all belong to the same inventive concept as the tumor survival prediction system based on quaternion-aligned multimodal data. For details of their specific implementation process, please refer to the embodiments of the tumor survival prediction system based on quaternion-aligned multimodal data, which will not be repeated here.
[0044] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A tumor survival prediction system based on quaternion-aligned multimodal data, characterized in that, Including those connected sequentially: The quaternion encoding module is used to map multimodal tumor feature data to corresponding quaternions; The quaternion rotation alignment module is used to calculate the difference quaternion based on the quaternions of each modal feature, and extract the rotation axis and rotation angle from the difference quaternion to construct the rotation quaternion. The rotation quaternion is used to rotate the quaternion of at least one modal feature to achieve the alignment of feature directions between modalities. The multimodal feature fusion module is used to fuse the quaternions of aligned modal features to extract intermodal interaction relationships and generate a global representation. The risk prediction output module is used to output tumor survival risk prediction results based on the global characterization.
2. The tumor survival prediction system based on quaternion-aligned multimodal data according to claim 1, characterized in that, The multimodal tumor feature data includes at least pathological image features and genomic features. The pathological image features are obtained by feature extraction from the histopathological images of tumor patients, and the genomic features are obtained by dimensionality reduction and encoding of a set of tumor-related genes.
3. The tumor survival prediction system based on quaternion-aligned multimodal data according to claim 1, characterized in that, The quaternion encoding module uses a quaternion encoder to map each modal feature to the quaternion space and encode them into corresponding quaternions.
4. The tumor survival prediction system based on quaternion-aligned multimodal data according to claim 1, characterized in that, In the quaternion rotation alignment module, the calculation of difference quaternions based on quaternions of each modal feature, and the extraction of rotation axis and rotation angle from the difference quaternions to construct rotation quaternions, includes: Select the quaternion of any modal feature as the benchmark, and calculate the difference quaternion between the quaternions of other modal features and the quaternion of the benchmark modal feature; The rotation angle is extracted from the real part of the differential quaternion, the rotation axis is extracted from the imaginary part, and a learnable rotation cost matrix is introduced to correct the rotation angle. Construct a rotation quaternion based on the rotation axis and the corrected rotation angle.
5. The tumor survival prediction system based on quaternion-aligned multimodal data according to claim 1, characterized in that, The multimodal feature fusion module employs an attention-based feature fusion network to map the quaternions of aligned modal features into query matrices, key matrices, and value matrices, respectively. It captures the feature interaction relationships between modalities through attention calculation and uses a gated attention mechanism to weight the modalities based on their importance contributions. Finally, it fuses the modalities to generate a global representation.
6. The tumor survival prediction system based on quaternion-aligned multimodal data according to claim 1, characterized in that, The risk prediction output module adopts a fully connected layer, which includes at least two linear layers: the first linear layer is used to perform dimensionality reduction processing on the global representation of the input, and the second linear layer is used to receive the features after dimensionality reduction in the first layer and further calculate and output the patient's survival risk probability.
7. The tumor survival prediction system based on quaternion-aligned multimodal data according to claim 1, characterized in that, A joint loss function is used to train a tumor survival prediction model that includes a quaternion encoding module, a quaternion rotation alignment module, a multimodal feature fusion module, and a risk prediction output module. The joint loss function includes a survival analysis loss for supervising survival prediction and a rotation loss for constraining the consistency of quaternion rotation alignment.
8. A tumor survival prediction method based on quaternion-aligned multimodal data, implemented using the tumor survival prediction system based on quaternion-aligned multimodal data as described in any one of claims 1 to 7, characterized in that, Includes the following steps: The quaternion encoding module is used to map multimodal tumor feature data into corresponding quaternions; The quaternion rotation alignment module calculates the difference quaternion based on the quaternion of each modal feature, and extracts the rotation axis and rotation angle from the difference quaternion to construct the rotation quaternion. The rotation quaternion is then used to rotate the quaternion of at least one modal feature to achieve the alignment of feature directions between modalities. The quaternions of aligned modal features are fused using a multimodal feature fusion module to extract intermodal interaction relationships and generate a global representation. The risk prediction output module outputs tumor survival risk prediction results based on the global characterization.
9. An electronic device comprising a memory and one or more processors, the memory for storing a computer program, characterized in that, The processor is used to implement the tumor survival prediction method based on quaternion-aligned multimodal data as described in claim 8 when executing a computer program.
10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by a computer, it implements the tumor survival prediction method based on quaternion-aligned multimodal data as described in claim 8.