A function-structure brain network representation learning method for brain disease diagnosis

Abstract

This invention belongs to the field of image processing technology and relates to a functional-structural brain network representation learning method for brain disease diagnosis, including: 1. Multimodal brain network data acquisition and preprocessing; 2. Topology-aware hypergraph modeling; 3. Attention mask generation based on structural topological distance; 4. Joint representation learning of topology-aware hypergraph and attention; 5. Graph-level representation and brain disease diagnosis output. In clinical applications, this invention can be used for auxiliary screening and classification of patients with brain diseases such as ASD, and provides doctors with reference and basis through the output classification labels and key brain regions, supporting earlier risk assessment, follow-up monitoring, and individualized intervention decisions. This invention provides strong and effective technical support for the early diagnosis and precision treatment of brain diseases, and has broad clinical application prospects and important medical research value.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a functional-structural brain network representation learning method for brain disease diagnosis. Background Technology

[0002] Early identification and classification of brain diseases (such as autism spectrum disorder (ASD), Alzheimer's disease (AD), and mild cognitive impairment (MCI)) are crucial for clinical intervention and prognostic assessment. In recent years, neuroimaging-based brain network analysis has been widely used in the auxiliary diagnosis of brain diseases. By abstracting the interactions between brain regions into network connections, disease-related abnormal connection patterns can be identified, and potential biomarkers can be further explored. Structural and functional brain networks represent brain region interactions from different perspectives; joint analysis of both is considered helpful for a more comprehensive understanding of brain region interactions and the pathogenesis of brain diseases. Currently, there are brain network representation learning methods for brain disease diagnosis, mainly including the following categories: 1. Graph Models / Graph Neural Networks (GNNs): Modeling pairwise relationships between brain regions, focusing on short-range dependencies. 2. Hypergraph Modeling: Using hyperedges to express higher-order associations between multiple brain regions; hyperedges construct semantic similarity of multi-dependent functional connections. 3. Transformer Self-Attention: Capturing long-range dependencies between brain regions, but mostly functional-side global attention learning.

[0003] However, although existing brain network representation learning methods for diagnosing brain diseases have achieved some success on certain datasets, from the perspective of structure-function coupling, these methods generally struggle to uniformly and stably characterize higher-order correlations and short / long-range dependencies in functional networks under the constraints of structural network topology. This results in limitations in diagnostic accuracy, robustness, and interpretability. Specifically: methods primarily based on graph-based pairwise relationship modeling struggle to fully express higher-order correlations across multiple brain regions; hypergraph modeling methods often rely solely on the semantic relevance of functional brain networks, neglecting structural topological constraints, leading to potential deviations of higher-order relationships from structural pathways; GNN / HGNN methods rely on message-passing local aggregation mechanisms, resulting in insufficient capture of long-range dependencies; while Transformer self-attention methods excel at long-range dependencies, the lack of structural topological distance constraints can lead to global generalization of attention, resulting in ineffective associations and noise.

[0004] Therefore, an integrated representation learning method that combines "high-order correlation + short-range / long-range dependency" under structural topological constraints is needed to solve the above technical problems. Summary of the Invention

[0005] This invention provides the following technical solution: a functional-structural brain network representation learning method for brain disease diagnosis, comprising the following steps: Step 1: Acquisition and preprocessing of multimodal brain network data; Step 2: Topology-aware hypergraph modeling; Step 3: Attention mask generation based on structural topological distance; Step 4: Joint representation learning of topology-aware hypergraph and attention; Step 5: Graph-level representation and brain disease diagnosis output.

[0006] Preferably, step 1 specifically includes: acquiring and preprocessing the subject's fMRI and DTI data to obtain a functional connectivity matrix and a structural connectivity matrix; and binarizing the structural connectivity matrix to obtain a structural topological adjacency matrix. The functional connection matrix As the initial feature of the node; where N is the number of brain regions.

[0007] Step 2 specifically includes: under the topological constraints of structural connections, using node feature similarity as a transition probability to guide sampling upstream in the structural topology, thereby generating hyperedges that better conform to structure-function coupling to express higher-order associations between brain regions.

[0008] Step 3 specifically includes: defining the topological distance function between brain regions from the topological graph formed by structural connections. Take the shortest path hop count between two nodes (if unreachable, set it to a value much larger than the maximum hop count) and generate two types of topology masks: a hard mask M to filter associations beyond k hops, and a soft mask... Used to smoothly decay and recalibrate the attention within k hops with topological distance.

[0009] Step 4 specifically includes: designing a hypergraph Transformer framework, where each layer consists of parallel hypergraph convolution branches and topological constraint attention branches, and obtaining a unified node representation through a fusion module.

[0010] Step 5 specifically includes: after multi-layer joint representation learning, the node representation simultaneously contains higher-order correlations and short / long-range dependencies under structural topological constraints; concatenating all node representations yields a graph-level representation. Representing the graph level The input classifier MLP outputs diagnostic class probabilities, and cross-entropy loss is used for supervised training.

[0011] More preferably, the transition probability in step 2 is:

[0012] in This indicates that in the SC structure, it is related to... The set of directly connected neighbors. The Euclidean distance is used to obtain the transition probability matrix. .

[0013] The set of hyperedges is represented as the incidence matrix of the hypergraph. The elements of the correlation matrix are defined as follows:

[0014] in, Representing brain region nodes, Indicates the super-edge.

[0015] More preferably, in step 3, the hard mask is:

[0016] Where k is the attention range threshold; attention exceeding the threshold is masked after softmax to filter out potential noise.

[0017] The soft mask is:

[0018] in By controlling the decay rate, brain regions that are closer to the topology are given greater weight, while those that are farther away are given less weight, thereby encouraging brain network representations to better conform to the structural topological space.

[0019] More preferably, in step 4, the hypergraph convolution branch is:

[0020] in, For the first Layer node features The initial function is the connectivity matrix. For hyperedge features, , For degree matrix, For learnable parameters, For non-linear activation, W is the hyperedge weight matrix (set as the identity matrix), and H is the hypergraph incidence matrix.

[0021] More preferably, in step 4, the topology-aware Transformer branch is: The output of the multi-head self-attention is:

[0022] In the formula, Concat represents concatenation. Here is the learnable parameter matrix; the initial attention score is calculated as follows:

[0023] In the formula, and For learnable parameter matrix, for The first dimension; Attention matrix after introducing topological hard mask and soft mask:

[0024] Calculate the weighted sum:

[0025] In the formula, This is the learnable parameter matrix.

[0026] More preferably, in step 4, the two branches are merged as follows: In the The layers respectively yield the hypergraph convolution branch output and the topological attention branch output:

[0027] The output of this layer is then obtained through the fusion module:

[0028] This fusion allows the model to simultaneously utilize high-order collaborative information and global dependency information of topological constraints in the same representation learning process.

[0029] More preferably, in step 5, supervised training is performed as follows:

[0030] Where C is the number of categories, For real labels, To predict probabilities.

[0031] The beneficial effects of this invention are: 1. In the diagnosis of brain diseases, this invention uses the structural connections obtained from DTI as topological constraints. Within this topology, hyperedges are constructed through semantically guided walk sampling. This ensures that high-order cooperative relationships in the functional connectivity network are modeled only on structurally reachable connections, thus avoiding spurious correlations and noise propagation caused by constructing edges solely based on functional similarity. It also overcomes the shortcomings of traditional graph models, which only model pairwise relationships and struggle to express multi-brain region cooperation. Simultaneously, topological hard and soft masks are introduced into the Transformer's self-attention mechanism. The hard mask blocks interactions exceeding k-hop, while the soft mask recalibrates weights based on Gaussian decay of the shortest path distance. This allows the model to capture long-range dependencies while focusing more on connections in closer and more relevant brain regions on the topology, achieving a mechanism-wise attention allocation that "filters potential noise and enhances effective neighbor information."

[0032] 2. The accuracy of this invention is improved by 7.2%, 4.2%, 5.1%, and 5.3% respectively on four diagnostic tasks: ASD vs NC, AD vs NC, AD vs MCI, and MCI vs NC. (For example, the accuracy of ASD vs NC is improved from 74.7% to 81.9%), thus the diagnostic accuracy of this invention is higher.

[0033] 3. This invention uses attention maps under topological constraints and discriminative hyperedge analysis to locate key subnetworks / key brain regions and discriminative connections, providing a basis for interpretable diagnosis and biomarker mining.

[0034] 4. In clinical applications, this invention can be used to assist in the screening and classification of patients with brain diseases such as ASD, and provide doctors with reference and basis through the output classification labels and key brain regions, supporting earlier risk assessment, follow-up monitoring and individualized intervention decisions. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of a functional-structural brain network representation learning method for brain disease diagnosis according to the present invention. Figure 2 This is a schematic diagram of the topology-aware hypergraph modeling module of the present invention. Detailed Implementation

[0036] The relevant technologies of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0037] like Figures 1-2 As shown, the functional-structural brain network representation learning method for brain disease diagnosis in this embodiment includes the following steps: Step S1: Acquisition and preprocessing of multimodal brain network data: fMRI and DTI data of the subjects were acquired and preprocessed to obtain functional connectivity and structural connectivity matrices; the structural connectivity matrix was binarized to obtain the structural topological adjacency matrix. The functional connection matrix As the initial feature of the node (N is the number of brain regions).

[0038] Step S2: Topology-aware hypergraph modeling: Existing hypergraph construction methods often model higher-order associations based solely on semantic relevance in functional connectivity, which can easily lead to hyperedge combinations that are inconsistent with structural pathways and introduce noise. This invention, under the topological constraints of structural connectivity, utilizes node feature similarity as a transition probability to guide sampling up the structural topology, thereby generating hyperedges that better conform to structure-function coupling to express higher-order associations between brain regions.

[0039] 1) Semantic similarity transfer probability calculation: Semantic similarity is represented by the reciprocal of the Euclidean distance of node features, and structural neighborhood is normalized to obtain the result from... arrive Transition probability:

[0040] in This indicates that in the SC structure, it is related to... The set of directly connected neighbors. The Euclidean distance is used. From this, the transition probability matrix is ​​obtained. .

[0041] 2) Semantically guided deterministic and random walks on structural connection topologies: For each node as the starting point, several sub-hyperedges are repeatedly sampled under structural topological constraints and their union is calculated to form the hyperedge corresponding to that node. In the deterministic walk, the structural neighbor with the highest transition probability is selected at each step, while the random walk performs preference sampling according to P (preferring semantically similar nodes) to take into account both "accuracy and diversity" in higher-order association modeling.

[0042] 3) Hypergraph Incidence Matrix Representation: The constructed set of hyperedges is represented as the incidence matrix of the hypergraph. Its elements are defined as:

[0043] in Representing brain region nodes, This represents a hyperedge. The hyperedge provides communication channels for multiple brain regions, enabling higher-order associations of multi-region collaboration supported by natural representational structural connectivity.

[0044] Step S3: Attention mask generation based on structural topological distance: To avoid the Transformer creating an "approximately fully connected" focus on the brain network that is inconsistent with the anatomical structure, this invention defines a topological distance function between brain regions from the topological graph formed by structural connections. Take the shortest path hop count between two nodes (if unreachable, set it to a value much greater than the maximum hop count), and generate two types of topology masks based on this: a hard mask M to filter associations beyond k hops, and a soft mask... Used for smooth decay and recalibration of attention within k hops with topological distance. The hard mask is defined as:

[0045] Where k is the attention range threshold (k-hops), and attention exceeding the threshold is masked after softmax to filter out potential noise.

[0046] The soft mask uses Gaussian kernel decay:

[0047] in By controlling the decay rate, brain regions that are closer to the topology are given greater weight, while those that are farther away are given less weight, thereby encouraging brain network representations to better conform to the structural topological space.

[0048] Step S4: Topology-aware hypergraph-attention joint representation learning: To capture both high-order collaborations (hypergraphs) and short / long-range dependencies (attention) within the same framework, this invention designs a hypergraph Transformer framework: each layer consists of parallel hypergraph convolution branches and topological constraint attention branches, and a unified node representation is obtained through a fusion module.

[0049] 1) Hypergraph Convolution (HGConv) Branch: Higher-order feature extraction Hypergraph convolution can be written as a two-stage message passing: first, nodes aggregate to the hyperedge, and then the hyperedge propagates back to the nodes.

[0050] in For the first Layer node features ( (Initially a functional connection matrix) For hyperedge features, , Let be the node degree and hyperedge degree matrix. For learnable parameters, This is a non-linear activation. This branch aggregates multi-brain region collaborative information on the topologically constrained hyperedge constructed by S2.

[0051] 2) Topology-aware Transformer branch: The output of the multi-head self-attention is:

[0052] Initial attention score calculation:

[0053] Attention matrix after introducing topological hard mask and soft mask:

[0054] And calculate the weighted sum:

[0055] M is responsible for blocking attention from anything other than the k-jump. It is responsible for recalibrating the k-hops within the topological distance decay, making long-range dependent learning across brain regions more consistent with structural pathways.

[0056] 3) Two-branch fusion: Obtaining a unified joint representation of topological constraints. In the The layers respectively yield the hypergraph convolution branch output and the topological attention branch output:

[0057] The output of this layer is then obtained through the fusion module:

[0058] This fusion allows the model to simultaneously utilize high-order collaborative information and global dependency information of topological constraints in the same representation learning process.

[0059] Step S5: Graph-level representation and brain disease diagnosis output: After multi-layer joint representation learning, the node representations simultaneously contain higher-order correlations and short / long-range dependencies under structural topological constraints; concatenating all node representations yields a graph-level representation. The input classifier (MLP) outputs diagnostic class probabilities, and is trained under supervised supervision using cross-entropy loss.

[0060] Where C is the number of categories, For real labels, This is for predicting probabilities. This end-to-end framework is used for brain disease diagnostic tasks.

[0061] Example The implementation of this embodiment for brain disease diagnosis is as follows: Using ABIDE and ADNI as data sources, 95 subjects (53 with ASD and 42 with NC) with both fMRI and DTI were selected from ABIDE, and 502 subjects (94 with AD, 195 with MCI, and 213 with NC) were selected from ADNI. The original fMRI was preprocessed using DPARSF, and the original DTI was preprocessed using PANDA. The preprocessing procedure for ADNI was consistent with that of ABIDE. After preprocessing, the entire brain was divided into 116 brain regions using an AAL template. BOLD time series were extracted from each brain region, and the Pearson correlation coefficient between any two brain regions was calculated to construct a functional connectivity matrix. This functional connectivity matrix was then used as the initial node feature matrix. Simultaneously, the structural connectivity matrix is ​​obtained based on DTI fiber tracing and binarized to obtain the adjacency matrix. , is used to characterize the topological connections of a structure.

[0062] To construct a hypergraph under topological constraints to improve the diagnostic representation of brain diseases, this embodiment uses the reciprocal of the Euclidean distance between node features as the semantic similarity within the neighborhood defined by the structural topology A, and normalizes it to obtain the transition probability matrix P. Then, starting from each brain region, a walk is performed with a fixed step size t=5: a deterministic walk is performed only once, selecting the neighbor with the highest transition probability within the structural neighborhood at each step and moving forward, continuously walking t=5 steps to obtain a deterministic sub-hyperedge; a random walk samples neighbors according to the P distribution and continuously walks t=5 steps to obtain a random sub-hyperedge, and the random sub-hyperedge is generated 3 times (i.e., q=3). The union of the nodes appearing in the deterministic sub-hyperedge corresponding to each starting node and the 3 random sub-hyperedges forms the final hyperedge. All hyperedges are represented by an association matrix H for subsequent hypergraph convolution and attention calculations.

[0063] Then, the shortest path hop count in the structure graph is used. Generating topological hard and soft masks: Hard masks are used to mask more than [a certain number of items]. -hop range attention interactions, soft mask in Within the -hop range, the attention score is recalibrated using a Gaussian kernel that decays with topological distance. In multi-head self-attention, the attention matrix is ​​applied using a method of "first superimposing a hard mask, then element-wise modulating it with a soft mask." Internally, ReLU is used as the activation function. The hypergraph Transformer structure has L=2 layers, S=3 attention heads, and a topological threshold of k=6. The training phase is implemented in PyTorch using the Adam optimizer, with an initial learning rate set to... And apply weight decay regularization and take the weight decay coefficient as... Five-fold cross-validation was used to complete training and testing, thereby outputting diagnostic categories for brain diseases. The effectiveness was validated using four commonly used quantitative indicators in medical diagnosis: accuracy, sensitivity, specificity, and AUC.

[0064] The core idea of ​​this invention is to model and learn high-order associations and long-range dependencies in functional brain networks by guiding the topology of structural brain networks in the context of brain disease diagnosis, thereby forming a functional-structural coupled representation for classification. Around this core idea, there are several alternative implementation methods: for example, attention constraints are not limited to the form of a mask matrix; additive bias or multiplicative gating can also be used to restrict and recalibrate attention scores; the fusion of high-order collaborative information and long-range dependency information is not limited to a specific fusion structure, but can also employ different methods such as serial, parallel, residual addition, weighted summation, and gated fusion, and the fusion location can be completed within a layer, across layers, or at the end. As long as the alternative scheme still achieves high-order collaborative modeling and long-range interactive learning under structural topological constraints and outputs a graph-level discriminative representation for brain disease diagnosis, it is an equivalent substitute for the technical solution of this invention.

[0065] The key technical point of this invention is: 1. A High-Order Association Modeling Mechanism Based on Structural Topological Constraints: In the joint modeling of functional-structural brain networks for brain disease diagnosis, the functional network not only includes pairwise associations between brain regions but also high-order associations involving multiple brain regions. However, existing hypergraph methods often rely solely on semantic similarity on the functional side (such as k-nearest neighbors, sparse representation, etc.) when constructing high-order relationships, easily neglecting the support and constraints of structural brain network topological connections on high-order associations, thus introducing high-order combination relationships inconsistent with anatomical structures. This invention explicitly introduces the topological connections of the structural brain network into the hyperedge generation process: first, structural topological adjacency relationships are obtained from SC; then, semantically guided walks (including deterministic and random walks) are performed within the structural topological neighborhood based on the transition probabilities obtained from functional semantic correlation; and the subsets obtained from multiple walks are combined to generate hyperedges, thereby ensuring that hyperedge construction occurs within the space reachable by the structural topology, reducing noise and improving the stability and consistency of the constructed high-order associations between brain regions.

[0066] 2. Topologically Aware Attention Learning: In brain networks, the influence and information transmission of one brain region on other brain regions are constrained by the topological pathways of the structural brain network. Brain regions that are structurally closer and more accessible are more likely to form stable and interpretable functional interactions. If global self-attention is directly applied to functional features, the model may assign high weights to brain region pairs that are topologically distant or even structurally weakly supported, thus treating noise correlations as valid dependencies, introducing unstable attentional connections and affecting diagnostic judgment. Therefore, this invention explicitly introduces the topological distance (shortest path hop count) obtained from the structural brain network into attention calculation and constructs a topological mask: a hard mask is used to block brain region pairs exceeding k hops to filter structurally unreasonable long-range interactions, and a soft mask is used to attenuate and recalibrate the attention score within the k-hop range according to the topological distance, making attention more focused on the connection range that is structurally reasonable and has stronger pathway support. This results in a long-range dependency representation that is more consistent with the structure-function coupling mechanism, reduces the propagation of invalid associations, and improves the robustness and accuracy of brain disease diagnosis.

[0067] 3. A Topology-Aware Hypergraph-Attention Joint Representation Learning Framework: Abnormalities related to brain diseases often manifest in two types of network patterns: one is collaborative activity involving multiple brain regions (higher-order associations), and the other is long-distance communication abnormalities across brain regions and network modules (long-range dependence). Relying solely on pairwise connections or emphasizing only global attention can easily overlook one of these key patterns, thus affecting the sufficiency and stability of diagnostic judgments. This invention starts from the brain network mechanism and simultaneously acquires two complementary types of information in the same representation learning process: on the one hand, it aggregates multi-brain region collaborative information under topological constraints at the higher-order relation level to form higher-order features that can represent "multi-brain region co-activation under structural connectivity constraints"; on the other hand, it learns the long-range interaction strength between brain regions under structural topological constraints, making the modeling of cross-regional communication relationships more consistent with structural pathway support. The two types of information are fused to form a unified function-structure coupling representation and output graph-level discriminative features for classification, thereby more comprehensively covering abnormal brain network patterns and improving the ability to distinguish and diagnose brain diseases.

[0068] In summary, this invention achieves effective utilization of structural brain network topological connections in brain disease diagnosis through a high-order association modeling mechanism based on structural topological constraints, topology-aware attention learning, and a topology-aware hypergraph-attention joint representation learning framework. This avoids the problems of unreasonable high-order combination relationships and unstable attention connections that are easily introduced in existing methods. By integrating structural information into high-order association modeling, attention computation, and overall representation learning, it more comprehensively and accurately captures high-order associations and long-range dependencies in brain networks. The function-structure coupling representations generated by this method for classification are more stable and reliable, and can more fully cover abnormal patterns in brain networks, thereby significantly improving the ability to distinguish and diagnose brain diseases. In practical applications, it provides strong and effective technical support for the early diagnosis and precision treatment of brain diseases, has broad clinical application prospects and important medical research value, and is expected to promote the development of the field of brain disease diagnosis towards greater precision and efficiency.

[0069] It should be emphasized that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A functional-structural brain network representation learning method for brain disease diagnosis, characterized in that, Includes the following steps: Step 1: Acquisition and preprocessing of multimodal brain network data; Step 2: Topology-aware hypergraph modeling; Step 3: Attention mask generation based on structural topological distance; Step 4: Joint representation learning of topology-aware hypergraph and attention; Step 5: Graph-level representation and brain disease diagnosis output.

2. The functional-structural brain network representation learning method for brain disease diagnosis according to claim 1, characterized in that, Step 1 specifically includes: acquiring and preprocessing the subject's fMRI and DTI data to obtain the functional connectivity matrix and the structural connectivity matrix; and binarizing the structural connectivity matrix to obtain the structural topological adjacency matrix. The functional connection matrix As the initial feature of the node; where N is the number of brain regions; Step 2 specifically includes: under the topological constraints of structural connections, using node feature similarity as a transition probability to guide sampling up the structural topology, thereby generating hyperedges that better conform to structure-function coupling to express higher-order associations between brain regions; Step 3 specifically includes: defining a topological distance function between brain regions from the topological graph formed by structural connections. The shortest path hop count between two nodes is taken, and two types of topology masks are generated: a hard mask M is used to filter associations outside of k hops, and a soft mask... Used to smoothly decay and recalibrate the attention within k hops with topological distance; Step 4 specifically includes: designing a hypergraph Transformer framework, each layer consisting of parallel hypergraph convolution branches and topological constraint attention branches, and obtaining a unified node representation through a fusion module; Step 5 specifically includes: after multi-layer joint representation learning, the node representation simultaneously contains higher-order correlations and short / long-range dependencies under structural topological constraints; concatenating all node representations to obtain a graph-level representation. Representing the graph level The input classifier MLP outputs diagnostic class probabilities, and cross-entropy loss is used for supervised training.

3. The functional-structural brain network representation learning method for brain disease diagnosis according to claim 2, characterized in that, The transition probability in step 2 is: in, This indicates that in the SC structure, it is related to... The set of directly connected neighbors. The Euclidean distance is used to obtain the transition probability matrix. ; The set of hyperedges is represented as the incidence matrix of the hypergraph. The elements of the correlation matrix are defined as follows: in, Representing brain region nodes, Indicates the super-edge.

4. The functional-structural brain network representation learning method for brain disease diagnosis according to claim 2, characterized in that, In step 3, the hard mask is: Where k is the attention range threshold; attention exceeding the threshold is masked after softmax to filter out potential noise; The soft mask is: in Control the decay rate.

5. The functional-structural brain network representation learning method for brain disease diagnosis according to claim 2, characterized in that, In step 4, the hypergraph convolution branch is: in, For the first Layer node features The initial function is the connectivity matrix. It is a hyperedge feature. , Let the node degree and hyperedge degree matrices be the matrix. For learnable parameters, For non-linear activation, W is the hyperedge weight matrix, and H is the hypergraph correlation matrix.

6. The functional-structural brain network representation learning method for brain disease diagnosis according to claim 5, characterized in that, In step 4, the topology-aware Transformer branch is as follows: The output of the multi-head self-attention is: In the formula, Concat represents concatenation. The learnable parameter matrix; The initial attention score is calculated as follows: In the formula, and For learnable parameter matrix, for The first dimension; Attention matrix after introducing topological hard mask and soft mask: Calculate the weighted sum: In the formula, This is the learnable parameter matrix.

7. The functional-structural brain network representation learning method for brain disease diagnosis according to claim 6, characterized in that, In step 4, the two branches are merged as follows: In the The layers respectively yield the hypergraph convolution branch output and the topological attention branch output: The output of this layer is then obtained through the fusion module: This fusion allows the model to simultaneously utilize high-order collaborative information and global dependency information of topological constraints in the same representation learning process.

8. The functional-structural brain network representation learning method for brain disease diagnosis according to claim 2, characterized in that, In step 5, supervised training is as follows: Where C is the number of categories. For real labels, To predict probabilities.

Images