Explainable brain network classification method and system based on topology-aware graph convolution

Abstract

The application discloses an interpretable brain network classification method and system based on topology-aware graph convolution, relates to the technical field of medical image processing, acquires brain image data, divides the brain image data into a plurality of regions of interest based on a preset anatomical template of brain structure, inputs a brain network matrix into a pre-established brain region-aware graph convolution network model to perform brain region-aware topology modeling and brain region-aware representation learning, outputs brain region-level features in the brain network, inputs the brain network matrix into a pre-established brain community-aware graph convolution network model to perform community-aware topology modeling and brain community-aware representation learning, outputs community-level features in the brain network, fuses the brain region-level features in the brain network and the community-level features in the brain network to obtain fused features, classifies based on the fused features, and obtains a classification result.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to an interpretable brain network classification method and system based on topological perceptual graph convolution. Background Technology

[0002] Brain network (BN) analysis is a cutting-edge field in neuroscience, aiming to reveal the organizational patterns of the human brain and the pathological mechanisms of various neurological diseases. Given the inherent graph structure of BNs, methods based on graph convolutional networks (GCNs) have been widely applied to BN analysis. In these methods, each row or column of the BN matrix is ​​typically treated as a feature vector of a region of interest (ROI), and the graph's topology (represented by an adjacency matrix) is generally constructed by applying a threshold to the BN matrix or using a k-nearest neighbor (k-NN) method. Through message passing mechanisms, graph convolutional networks learn meaningful graph representations from BNs, thereby aiding in the identification of brain diseases.

[0003] Existing methods often rely on predefined topologies, which may fail to accurately capture the intrinsic relationships between brain regions of interest (ROIs), thus limiting the representational capabilities of graph convolutional networks (BNs). Furthermore, BNs are essentially composed of multiple functional communities, with ROIs belonging to the same community working collaboratively to perform similar neural activities. Current techniques typically assume that each ROI contributes only to its own single community, neglecting its contributions to other communities, thereby limiting the diagnostic performance of the model. Summary of the Invention

[0004] To address the shortcomings mentioned in the background section, the present invention aims to provide an interpretable brain network classification method and system based on topologically perceptual graph convolution.

[0005] Firstly, the objective of this invention can be achieved through the following technical solution: an interpretable brain network classification method based on topological sensing graph convolution, the method comprising the following steps: Brain imaging data is acquired, and the brain within the brain imaging data is divided based on a pre-defined anatomical template of brain structure. The time series of blood oxygenation level dependent signals for each region of interest is obtained, and the Pearson correlation coefficient between each pair of regions of interest is calculated as a brain network matrix. The brain network matrix is ​​input into a pre-established brain region perception map convolutional network model to perform brain region perception topology modeling and brain region perception representation learning, and outputs brain region level features in the brain network. The brain network matrix is ​​input into a pre-established brain community-aware graph convolutional network model to perform community-aware topology modeling and brain community-aware representation learning, and outputs community-level features within the brain network. The brain region-level features and community-level features within the brain network are fused to obtain fused features. Classification is then performed based on these fused features to obtain the classification results.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the brain region-aware topology modeling of the brain region-aware map convolutional network model is performed by using self-expression learning to perform topology modeling at the brain region level, as follows: set up For a subject's brain network matrix, where Indicates the first Characteristics of each brain region Self-expression is represented as an equation ,in For the self-expression coefficient matrix, set With diagonal elements of 0, the self-expressive representation is defined as the following optimization problem: In the formula, for The Frobenius square norm, where λ is the regularization parameter. for The diagonal elements; Brain region similarity matrix The calculation process is as follows: calculate The nth-order singular value decomposition, denoted as ,in d is the maximum dimension of the subspace, and K is the number of brain communities; calculate ,Will Normalization, then set The negative elements are all zero, resulting in a nonnegative matrix. ; Constructing a brain region similarity matrix ,in , reserve The adjacency matrix is ​​constructed using the median of the first preset percentage of edges. .

[0007] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: a brain region perception representation learning process for the brain region perception map convolutional network model, as follows: Based on adjacency matrix Given a brain region perception map Using graph convolutional networks, representations of region-aware topology are learned from brain networks. The output node feature matrix after two layers of graph convolutional networks is... As a brain region-level feature within a brain network, it can be represented as: In the formula, , To increase the adjacency matrix of self-connections, for The degree matrix, and These are the trainable weight matrices for layers 1 and 2, respectively. This is the ReLU activation function.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: community-aware topology modeling of the brain community-aware graph convolutional network model is used to construct the connection relationships between brain communities, as follows: Using a multilayer perceptron equipped with a softmax activation function, the features of each node are... Mapped to the distribution of its soft community contributions : In the formula, Features of each node Mapped to K-dimensional vector Softmax transforms a K-dimensional vector Convert to a probability distribution, where K represents the number of clubs; Using brain region similarity matrix Optimize MLP to guide community detection to capture reliable brain community interactions, and define thresholds. and To retain only node pairs with extreme similarity values, where and Introducing subsets based on thresholds as follows: set up To be included in the community contribution matrix, only consider... The objective function is as follows: Discriminative node pairs are used to guide the learning of community contributions. In the formula, Brain regions with similar guidance functions show consistent contributions to community building. Guide brain regions to contribute to their respective communities. Denotes KL divergence, It is a one-hot encoded tag, representing a node. Which club / organization does it belong to? Constructing a community-level similarity matrix Preserve the similarity matrix of communities A second percentage of edges is pre-defined to construct a community-level adjacency matrix. The final constructed community feature matrix is ​​as follows: .

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: a brain community-aware representation learning process for the brain community-aware graph convolutional network model, as follows: Based on community-level adjacency matrix Given a brain community perception map Graph convolutional networks are used to learn representations of the brain community's perceptual topology from brain networks. The output node feature matrix after two layers of graph convolutional networks is... As a community-level feature within the brain network, it can be represented as: In the formula, , To increase the adjacency matrix of self-connections, for The degree matrix, and These are the trainable weight matrices for the first and second layers, respectively. This is the ReLU activation function.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the calculation process for the Pearson correlation coefficient between each pair of regions of interest, as follows: In the formula, This indicates the degree of linear correlation between brain region X and brain region Y, with a value between -1 and 1. and X represents the average value of the blood oxygenation level dependent signal in brain regions X and Y over time series, where X and Y represent the blood oxygenation level dependent signal values ​​in brain regions X and Y at various time points over time series.

[0011] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the classification based on the fused features is achieved by inputting the fused features into two fully connected layers and using softmax as the activation function to finally obtain the classification result.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of inputting the fused features into two fully connected layers, as follows: Node feature matrix Community Feature Matrix The final feature map is obtained by fusion: In the formula This indicates a splicing operation. Based on the community index Mapping to The same dimension, will The input is fed into the row-level summation Readout layer to obtain a graph-level representation, the expression of which is: Each participant's fusion feature is a The vector is then fed into the classifier to obtain the raw class scores: In the formula, Indicates the first linear layer. ReLU represents the second linear layer, and ReLU represents the ReLU activation layer. This represents the original score of the predicted category; The classification results are obtained using the following formula: , In the formula, This represents the softmax activation function. This represents the index-value function that takes the maximum value, and Y represents the diagnostic result; The loss function is as follows: In the formula, and They are the first The original category scores and true labels of each subject, where, , It is the number of subjects.

[0013] Secondly, in order to achieve the above objectives, this invention discloses an interpretable brain network classification system based on topologically perceptual graph convolution, comprising: The data processing module is used to acquire brain imaging data, divide the brain within the brain imaging data based on a preset anatomical template of brain structure, obtain the time series of blood oxygen level dependent signals for each region of interest, and calculate the Pearson correlation coefficient between each pair of regions of interest as a brain network matrix. The first feature extraction module is used to input the brain network matrix into a pre-established brain region perception map convolutional network model to perform brain region perception topology modeling and brain region perception representation learning, and output brain region-level features in the brain network. The second feature extraction module is used to input the brain network matrix into a pre-established brain community-aware graph convolutional network model to perform community-aware topology modeling and brain community-aware representation learning, and output community-level features in the brain network. The classification output module is used to fuse brain region-level features and community-level features within the brain network to obtain fused features. Based on the fused features, classification is performed to obtain the classification result.

[0014] The beneficial effects of this invention are: This invention achieves flexible and accurate brain region-level topology construction by learning the intrinsic relationships of brain regions through self-expression and adaptive learning, thereby improving BN representation learning. It combines brain community interaction learning methods to identify the interactions between brain functional communities, providing community-level topology structures, enhancing the model's representation learning ability, and improving the accuracy of brain network analysis. Attached Figure Description

[0015] 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the brain network analysis model structure provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the first fifteen pairs of discriminative connections and community interaction structures identified by the brain network analysis model provided in the embodiment of the present invention in the diagnosis of Alzheimer's disease.

[0016] Figure 4 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: like Figure 1 As shown, an interpretable brain network classification method based on topologically sensitive graph convolution includes the following steps: S101: Acquire brain imaging data, divide the brain within the brain imaging data based on a preset anatomical template of brain structure, obtain the time series of blood oxygenation level dependent signals for each region of interest, calculate the Pearson correlation coefficient between each pair of regions of interest, and use it as a brain network matrix. Specifically, the brain imaging data refers to functional magnetic resonance imaging (fMRI) brain imaging data of the subjects in a resting state, and the preprocessing process is as follows: Slicing time and head motion correction, spatial normalization to MNI space, spatial smoothing, and global drift removal.

[0019] Regions of interest are divided from brain imaging data based on a pre-defined anatomical template of brain structure: The brain was divided into 116 regions of interest (ROIs) based on the Anatomical Automatic Labelling (AAL) template. The average time series of each ROI was extracted to form the overall time series data of the subject's brain region (data format: N×R×T, where N is the number of samples, R represents the number of regions of interest, and T is the time series length). The brain network BN matrix was constructed by calculating the Pearson correlation coefficient between paired regions of interest. The brain network matrix for each subject is obtained by calculating the Pearson correlation coefficient. The formula for calculating the Pearson correlation coefficient between brain region X and brain region Y is as follows: In the formula, This indicates the degree of linear correlation between brain region X and brain region Y, with a value between -1 and 1. A higher value indicates a stronger correlation. The correlation value between a brain region and itself is 1. and X represents the average value of the blood oxygenation level dependent signal in brain regions X and Y over time series, where X and Y represent the blood oxygenation level dependent signal values ​​in brain regions X and Y at various time points over time series.

[0020] Using the above method, the data will be transformed into an M×N×N tensor.

[0021] S102: Input the brain network matrix into the pre-established brain region perception map convolutional network model to perform brain region perception topology modeling and brain region perception representation learning, and output brain region level features in the brain network. The brain region perception topology model of the brain region perception convolutional network model utilizes self-expressive learning to perform topology modeling at the brain region level, as follows: set up Let be the BN matrix of a subject, where Indicates the first The characteristics of each brain region mean that self-expression can be simply represented by an equation. ,in This is the self-expressing coefficient matrix. To avoid the case of natural solutions, we set... The diagonal elements are 0. Mathematically, a self-expressive model can be defined as the following optimization problem: In the formula, for The Frobenius square norm, where λ is the regularization parameter. for The diagonal elements.

[0022] Subsequently, the brain region similarity matrix The calculation process is as follows: calculate The nth-order singular value decomposition, denoted as ,in d is the maximum dimension of the subspace, and K is the number of brain communities.

[0023] calculate ,Will Normalization, then set The negative elements are all zero, resulting in a nonnegative matrix. .

[0024] Constructing a brain region similarity matrix ,in .

[0025] Subsequently, retain The median value is the first preset percentage of the edge (set). The adjacency matrix is ​​constructed by setting the first percentage maximum value to 1 and the rest to 0. In this embodiment, the preset first percentage is 5%. The brain region perception representation learning process of the brain region perception map convolutional network model is as follows: Based on adjacency matrix Given a brain region perception map Using graph convolutional networks, representations of region-aware topology are learned from brain networks. The output node feature matrix after two layers of graph convolutional networks is... As a brain region-level feature within a brain network, it can be represented as: In the formula, , To increase the adjacency matrix of self-connections, for The degree matrix, and These are the trainable weight matrices for layers 1 and 2, respectively. This is the ReLU activation function.

[0026] S103: Input the brain network matrix into the pre-established brain community-aware graph convolutional network model to perform community-aware topology modeling and brain community-aware representation learning, and output community-level features in the brain network. Community-aware topology modeling of the brain community-aware graph convolutional network model is used to construct the connection relationships between brain communities. The process is as follows: First, a multilayer perceptron equipped with a softmax activation function is used to process the features of each node. Mapped to the distribution of its soft community contributions : In the formula, Features of each node Mapped to K-dimensional vector Softmax transforms a K-dimensional vector Convert to a probability distribution, where K represents the number of communities.

[0027] To capture reliable brain community interaction relationships, a ROI similarity matrix was used. The MLP is optimized to guide community detection, ensuring that the predicted community contribution vectors remain functionally consistent across ROI pairs. Furthermore, due to noise in the data, a significant portion of the computed pairwise ROI similarity may deviate from the true topological relationships. Therefore, a threshold is defined. and This is to retain only node pairs with extreme similarity values, because these values ​​inherently contain greater discriminative information, where and Based on these thresholds, a subset is then introduced. as follows: set up Become part of the community contribution matrix. Only consider... The objective function is as follows: Discriminative node pairs are used to guide the learning of community contributions. In the formula, Brain regions with similar guidance functions show consistent contributions to community building. Guide brain regions to contribute to their respective communities. Denotes KL divergence, It is a one-hot encoded tag, representing a node. Which club / organization does it belong to? Then, construct a community-level similarity matrix. Similarly, the similarity matrix of communities is preserved. A second percentage of edges is pre-defined to construct a community-level adjacency matrix. Therefore, the community feature matrix can be constructed as follows: In this embodiment, the second percentage is preset to 50%.

[0028] The brain community perception representation learning process of the convolutional network model is as follows: Given a brain community perception map We use GCN to learn the representation of brain community perceptual topology from BN. The output node feature matrix after two layers of GCN is shown. It can be represented as: In the formula, , To increase the adjacency matrix of self-connections, for The degree matrix, and These are the trainable weight matrices for the first and second layers, respectively. This is the ReLU activation function.

[0029] S104: The brain region-level features and community-level features within the brain network are fused to obtain fused features. Classification is then performed based on the fused features to obtain the classification results.

[0030] Classification based on fused features is achieved by inputting the fused features into two fully connected layers and using softmax as the activation function, ultimately obtaining the classification result.

[0031] The process of inputting the fused features into two fully connected layers is as follows: Node feature matrix Community Feature Matrix The final feature map is obtained by fusion: In the formula This indicates a splicing operation. Based on the community index Mapping to The same dimension, will The input is fed into the row-level summation Readout layer to obtain a graph-level representation, the expression of which is: Each participant's fusion feature is a The vector is then fed into the classifier to obtain the raw class scores: In the formula, Indicates the first linear layer. ReLU represents the second linear layer, and ReLU represents the ReLU activation layer. This represents the original score of the predicted category; The classification results are obtained using the following formula: , In the formula, This represents the softmax activation function. This represents the index-value function that takes the maximum value, and Y represents the diagnostic result; The loss function is as follows: In the formula, and They are the first The original category scores and true labels of each subject, where, , It is the number of subjects.

[0032] Specifically, the present invention will be further illustrated below through embodiments: In this embodiment, the ADNI dataset is used as the training sample to illustrate the proposed interpretable brain network modeling and analysis process based on topology-aware graph convolution, including the following steps: Step 1: Obtain the data.

[0033] Download and prepare the ADNI dataset.

[0034] All rs-fMRI samples in this study were obtained from the ADNI (Alzheimer's Disease Neuroimaging Initiative) dataset. The ADNI dataset contains 563 rs-fMRI datasets, including 99 Alzheimer's patients, 165 patients with early-stage mild cognitive impairment, 145 patients with late-stage mild cognitive impairment, and 154 healthy controls. Data were preprocessed using FSL FEAT. Preprocessing steps included slice time and head motion correction, spatial normalization to MNI space, spatial smoothing, and global drift removal. Subsequently, the Anatomical Automatic Labelling (AAL) template divided the brain into 116 Regions of Interest (ROIs), and the average time series data of each ROI was extracted to form the overall time series data of the subject's brain region. The Pearson correlation coefficient was calculated to obtain the brain network matrix for each subject, with a shape of 116×116. Demographic information for all participants is shown in Table 1. Table 1. Statistical information of the ADNI dataset sample (M represents male, F represents female) Step 2: Establish a classification network.

[0035] The network primarily consists of a brain region-aware GCN, a community-aware GCN, and fully connected layers. Specifically, the topologies of the brain region-aware and community-aware graphs are first modeled using region-aware GCN and community-aware GCN modules, respectively. The brain region-aware topology (represented by an adjacency matrix) is constructed through self-representation learning. The brain community-aware topology is modeled using a brain community interaction learning method, reflecting the interactions between functional communities. Subsequently, multiple GCN layers are used to learn Batch Normalization (BN) representations at the brain region and brain community levels. Finally, these representations are fused and fed into fully connected layers, where the Softmax activation function is used to transform the classification output values.

[0036] Step 4: Set network parameters.

[0037] The network model was implemented in the PyTorch framework, specifically PyTorch version 2.5.0 and CUDA version 12.4. The training batch size was 8, the learning rate was 1e-3, and the number of training epochs was 100. In brain region perceptual representation learning, the two weight matrices of GCN were 116 × 256 and 256 × 128, respectively. In brain community perceptual representation learning, the two weight matrices of GCN were 1128 × 64 and 64 × 32, respectively. The activation function after the graph convolutional layer was ReLU, and the dropout was 0.2. The weight matrices of the two fully connected layers were 160 × 64 and 64 × 2, respectively. The loss function included... , The thresholds were set to 0.1 and 0.9 respectively. Set it to 0.3, and train the model on this configuration to obtain a trained brain network classification network.

[0038] Step 5: Use the trained interpretable topology-aware graph convolutional network to classify and predict the samples.

[0039] Through the learning and training in the third and fourth steps, the optimal network model parameters are obtained by manually tuning the parameters, and the class prediction is performed on the test sample data in the dataset.

[0040] Five-fold cross-validation was used to evaluate the classification performance of the model on the dataset. Specifically, the dataset was first randomly divided into five subsets of roughly equal size. Then, one subset was used as the test set and the remaining subsets were used as the training set. This process was repeated five times to ensure that each subset was used as the test set once. This invention performed three classification tasks: AD vs. NC classification task, EMCI vs. LMCI classification task, and AD vs. LMCI classification task. Four metrics were used for performance evaluation: accuracy (ACC, the number of cases correctly classified), sensitivity (SEN, the proportion of actual patients correctly classified by the model), specificity (SPE, the proportion of actual normal people correctly classified by the model), and area under the receiver operating characteristic curve (AUC). The average of the five-fold cross-validation results for each metric was taken as the final result. Meanwhile, in order to verify the effectiveness of the model (TAGNN) proposed in this invention, it was compared with the baseline model BrainNetCNN using convolution, and also compared with three other types of models, namely (1) Transformer-based models: ConnectomeTF and BrainNetTF. (2) Graph neural network-based models: GCN, GIN, and FBNetGen. (4) Models based on predefined brain community structures: Com-BrainTF. The comparison results with the above 7 methods are given in Table 2; Table 2 Comparison results of 7 methods in 3 binary classification tasks on the ADNI dataset The following conclusions can be drawn from Table 2. First, graph neural network-based methods (i.e., GCN, GIN, FBNetGen, and TAGNN) outperform BrainNetCNN, indicating that BN is more suitable for processing graph-structured data, thus effectively extracting features for brain disease classification. Second, GCN methods based on learnable graph topologies (i.e., FBNetGen and TAGNN) outperform GCN methods relying on predefined graph topologies (i.e., GCN and GIN). For example, TAGNN achieves an accuracy of 88.91% in the AD vs. NC classification task, significantly higher than GCN's 75.89%. This shows that the learned brain graph topology can better capture brain connectivity patterns, thereby improving the performance of brain disease diagnosis tasks. Furthermore, compared to graph neural network-based methods (i.e., GCN, GIN, and FBNetGen), brain community-based methods (i.e., Com-BrainTF) achieve better results, indicating that communities provide valuable topological information that can further improve classification performance. Finally, TAGNN consistently outperforms other comparative methods in terms of accuracy. Specifically, TAGNN achieved an average accuracy of 85.24% across all tasks on the ADNI dataset, a 7.09% improvement over the second-best method, Com-BrainTF (78.15%). The results indicate that further incorporating community-level features helps improve the model's classification performance.

[0041] also, Figure 3 (Left) Shows the top 15 ROI-level connectivity patterns and corresponding brain regions identified by TAGNN in AD diagnosis. As can be seen from the figure, the cingulate gyrus (PCG), thalamus (THA), and lingual gyrus (LING) are identified as key brain regions in AD diagnosis, and these brain regions have been confirmed to be closely related to AD diagnosis in previous studies. Figure 3 (Right) This shows the brain community interaction patterns identified by TAGNN. It can be seen that the default mode network (DMN) is densely connected with the sensorimotor network (SMN) and the visual network (VN), indicating that AD patients have abnormal cross-network coupling and reduced network separation, which is consistent with previous research results.

[0042] Example 2: To achieve the above objective, such as Figure 4 As shown, based on Embodiment 1, this invention discloses an interpretable brain network classification system based on topologically aware graph convolution, comprising: Data processing module 11 is used to acquire brain imaging data, divide the brain within the brain imaging data based on a preset anatomical template of brain structure, obtain the time series of blood oxygen level dependent signals for each region of interest, and calculate the Pearson correlation coefficient between each pair of regions of interest as a brain network matrix. The first feature extraction module 12 is used to input the brain network matrix into the pre-established brain region perception map convolutional network model to perform brain region perception topology modeling and brain region perception representation learning, and output brain region level features in the brain network. The second feature extraction module 13 is used to input the brain network matrix into the pre-established brain community perception graph convolutional network model to perform community perception topology modeling and brain community perception representation learning, and output community-level features in the brain network. The classification output module 14 is used to fuse brain region-level features and community-level features in the brain network to obtain fused features, and to perform classification based on the fused features to obtain classification results.

[0043] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0044] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0045] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0046] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. An interpretable brain network classification method based on topologically sensitive graph convolution, characterized in that, The method includes the following steps: Brain imaging data is acquired, and the brain within the brain imaging data is divided based on a pre-defined anatomical template of brain structure. The time series of blood oxygenation level dependent signals for each region of interest is obtained, and the Pearson correlation coefficient between each pair of regions of interest is calculated as a brain network matrix. The brain network matrix is ​​input into a pre-established brain region perception map convolutional network model to perform brain region perception topology modeling and brain region perception representation learning, and outputs brain region level features in the brain network. The brain network matrix is ​​input into a pre-established brain community-aware graph convolutional network model to perform community-aware topology modeling and brain community-aware representation learning, and outputs community-level features within the brain network. The brain region-level features and community-level features within the brain network are fused to obtain fused features. Classification is then performed based on these fused features to obtain the classification results.

2. The interpretable brain network classification method based on topologically aware graph convolution as described in claim 1, characterized in that, The brain region-aware topology model of the brain region-aware convolutional network model utilizes self-expressive learning to perform topology modeling at the brain region level, as follows: set up For a subject's brain network matrix, where Indicates the first Characteristics of each brain region Self-expression is represented as an equation ,in For the self-expression coefficient matrix, set With diagonal elements of 0, the self-expressive representation is defined as the following optimization problem: In the formula, for The Frobenius square norm, where λ is the regularization parameter. for The diagonal elements; Brain region similarity matrix The calculation process is as follows: calculate The nth-order singular value decomposition, denoted as ,in d is the maximum dimension of the subspace, and K is the number of brain communities; calculate ,Will Normalization, then set The negative elements are all zero, resulting in a nonnegative matrix. ; Constructing a brain region similarity matrix ,in , reserve The adjacency matrix is ​​constructed using the median of the first preset percentage of edges. .

3. The interpretable brain network classification method based on topologically aware graph convolution according to claim 2, characterized in that, The brain region perception representation learning process of the brain region perception map convolutional network model is as follows: Based on adjacency matrix Given a brain region perception map Graph convolutional networks are used to learn representations of brain region perceptual topology from brain networks. The output node feature matrix after two layers of graph convolutional networks is... As a brain region-level feature within a brain network, it can be represented as: In the formula, , To increase the adjacency matrix of self-connections, for The degree matrix, and These are the trainable weight matrices for layers 1 and 2, respectively. This is the ReLU activation function.

4. The interpretable brain network classification method based on topologically aware graph convolution according to claim 1, characterized in that, The community-aware topology model of the brain community-aware graph convolutional network model is used to construct the connection relationships between brain communities, and the process is as follows: Using a multilayer perceptron equipped with a softmax activation function, the features of each node are... Mapped to its soft community contribution distribution : In the formula, Features of each node Mapped to K dimensional vector softmax will K dimensional vector Convert to a probability distribution K Indicates the number of clubs; Using brain region similarity matrix Optimize MLP to guide community detection to capture reliable brain community interactions, and define thresholds. and To retain only node pairs with extreme similarity values, where and Introducing subsets based on thresholds as follows: set up To be included in the community contribution matrix, only consider... The objective function is as follows: Discriminative node pairs are used to guide the learning of community contributions. In the formula, Brain regions with similar guidance functions show consistent contributions to community building. Guide brain regions to contribute to their respective communities. Denotes KL divergence, It is a one-hot encoded tag, representing a node. Which club / organization does it belong to? Constructing a community-level similarity matrix Preserve the similarity matrix of communities A second percentage of edges is pre-defined to construct a community-level adjacency matrix. The final constructed community feature matrix is ​​as follows: .

5. The interpretable brain network classification method based on topologically aware graph convolution according to claim 4, characterized in that, The brain community perception representation learning process of the convolutional network model is as follows: Based on community-level adjacency matrix Given a brain community perception map The graph convolutional network is used to learn the representation of the brain community's perceptual topology from the brain network. The output node feature matrix after two layers of graph convolutional network is used. As a community-level feature within the brain network, it can be represented as: In the formula, , To increase the adjacency matrix of self-connections, for The degree matrix, and These are the trainable weight matrices for the first and second layers, respectively. This is the ReLU activation function.

6. The interpretable brain network classification method based on topologically aware graph convolution according to claim 1, characterized in that, The calculation process for the Pearson correlation coefficient between each pair of regions of interest is as follows: In the formula, This indicates the degree of linear correlation between brain region X and brain region Y, with a value between -1 and 1. and X represents the average value of the blood oxygenation level dependent signal in brain regions X and Y over time series, where X and Y represent the blood oxygenation level dependent signal values ​​in brain regions X and Y at various time points over time series.

7. The interpretable brain network classification method based on topologically aware graph convolution according to claim 1, characterized in that, The classification based on the fused features is achieved by inputting the fused features into two fully connected layers and using softmax as the activation function to obtain the final classification result.

8. The interpretable brain network classification method based on topologically aware graph convolution according to claim 7, characterized in that, The process of inputting the fused features into two fully connected layers is as follows: Node feature matrix Community Feature Matrix The final feature map is obtained by fusion: In the formula This indicates a splicing operation. Based on the community index Mapping to The same dimension, will The input is fed into the row-level summation Readout layer to obtain a graph-level representation, the expression of which is: Each participant's fusion feature is a The vector is then fed into the classifier to obtain the raw class scores: , In the formula, Indicates the first linear layer. ReLU represents the second linear layer, and ReLU represents the ReLU activation layer. This represents the original score of the predicted category; The classification results are obtained using the following formula: In the formula, This represents the softmax activation function. This represents the index-value function that takes the maximum value, and Y represents the diagnostic result; The loss function is as follows: In the formula, and They are the first The original category scores and true labels of each subject, where, , It is the number of subjects.

9. An interpretable brain network classification system based on topologically sensitive graph convolution, employing the interpretable brain network classification method based on topologically sensitive graph convolution as described in any one of claims 1 to 8, characterized in that, include: The data processing module is used to acquire brain imaging data, divide the brain within the brain imaging data based on a preset anatomical template of brain structure, obtain the time series of blood oxygen level dependent signals for each region of interest, and calculate the Pearson correlation coefficient between each pair of regions of interest as a brain network matrix. The first feature extraction module is used to input the brain network matrix into a pre-established brain region perception map convolutional network model to perform brain region perception topology modeling and brain region perception representation learning, and output brain region-level features in the brain network. The second feature extraction module is used to input the brain network matrix into a pre-established brain community-aware graph convolutional network model to perform community-aware topology modeling and brain community-aware representation learning, and output community-level features in the brain network. The classification output module is used to fuse brain region-level features and community-level features within the brain network to obtain fused features. Based on the fused features, classification is performed to obtain the classification result.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on a processor. When the processor loads and executes the computer program, it employs the interpretable brain network classification method based on topological sensing graph convolution as described in any one of claims 1 to 8.

Images