A method and system for autism diagnosis based on interpretable brain map neural networks with multi-teacher distillation

Abstract

This invention discloses a method and system for autism diagnosis based on interpretable brain graph neural networks (GCNs) with multi-teacher distillation in the field of medical image processing technology. The method includes: preprocessing functional magnetic resonance imaging (fMRI) data of subjects in a resting state to obtain time-series data of blood oxygenation level-dependent signals in various brain regions of interest (ROIs); constructing a brain network BN matrix by calculating the Pearson correlation coefficient between any two ROIs; constructing a brain network graph based on the BN matrix; inputting the brain network graph into a trained autism diagnosis model for autism diagnosis to obtain a raw score for the predicted category; and converting the raw score for the predicted category into a diagnostic result. This invention uses multiple GCN teacher models to capture BN topological knowledge at different scales to guide the brain graph neural network, and combines an attention mechanism to adaptively fuse information from different subgraphs, enhancing the model's representation learning ability and effectively achieving autism diagnosis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to an autism diagnosis method and system based on an interpretable brain map neural network with multi-teacher distillation. Background Technology

[0002] Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by impaired social communication and interaction skills, leading to dysfunction in daily living. Early and accurate diagnosis is crucial for improving the prognosis and quality of life for individuals with ASD; however, due to the heterogeneity of the disorder, early and accurate diagnosis remains challenging. Traditional symptom-based diagnostic methods, relying solely on behavioral observation and subjective assessment, may lead to biased clinical reasoning and incorrect diagnostic decisions. Therefore, designing interpretable methods based on underlying biological mechanisms rather than symptoms is gradually becoming a consensus.

[0003] Resting-state functional magnetic resonance imaging (rs-fMRI), as an emerging neuroimaging technique, has been widely used in the automated diagnosis of brain diseases by quantifying blood oxygenation level dependent (BOLD) signals as a proxy for brain activity. Typically, rs-fMRI can be modeled as a brain network (BN), where each node represents a region of interest (ROI), and each edge represents functional connectivity, measured, for example, by the Pearson correlation between BOLD signals of paired ROIs. These connections can reveal synergistic activity between different brain regions and provide a framework for exploring the neurobiological basis through graph-based methods. Accumulated evidence from multiple independent studies indicates that many brain diseases are accompanied by abnormal patterns of functional connectivity in brain networks. Therefore, significant efforts have been devoted to studying brain networks to explore the relationship between brain diseases and abnormal interactions, which may serve as potential biomarkers for the diagnosis of brain diseases.

[0004] In recent years, Graph Convolutional Networks (GCNs) have attracted much attention due to their powerful ability to process graph-structured data, and have naturally been applied to the analysis of Batch Normalization (BN). For BN-based classification tasks, GCNs treat each BN network as a graph, where each row / column of the BN matrix represents the feature of the corresponding node, and the adjacency matrix is ​​constructed by applying hard thresholding to the BN matrix or the k-nearest neighbor (k-NN) algorithm.

[0005] Another limitation of existing Graph Convolutional Networks (GCNs) is that they use a fixed number of layers to learn the representation of each BN, failing to fully integrate topological information into the learning process. For example, the degree of a node in a brain network is quantified by the number of connections or edges connected to nodes in a specific brain region, reflecting the node's centrality and providing information about its surrounding topology, while node degree values ​​can vary significantly across different brain regions. In practice, it has been observed that GCN models with more layers perform better for nodes with higher degree during node representation learning, while models with fewer layers are better suited for nodes with lower degree. To enhance the expressive power of brain network analysis, knowledge distillation (KD) has emerged as a promising technique and has been successfully extended to GCNs. KD aims to transfer knowledge from a complex teacher model to a lightweight student model during the learning process. Although many GCN-based knowledge distillation methods have made significant progress in brain disease analysis, most methods rely on a single GCN teacher model for knowledge distillation, ignoring the diversity of topologies that may benefit student networks. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide an autism diagnosis method and system based on a multi-teacher distillation interpretable brain graph neural network. By capturing BN topological knowledge at different scales through multiple GCN teacher models, the brain graph neural network is guided. Combined with the attention mechanism, information from different subgraphs is adaptively fused to enhance the model's representation learning ability, which can effectively achieve the judgment of autism.

[0007] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution: In a first aspect, the present invention provides a method for diagnosing autism based on an interpretable brain map neural network with multi-teacher distillation, comprising: Preprocessing was performed on the functional magnetic resonance imaging data of the subjects at rest to obtain time-series data of blood oxygenation level dependent signals for each brain region of interest. Based on the time-series data of blood oxygenation level dependent signals in various brain regions of interest, a brain network BN matrix is ​​constructed by calculating the Pearson correlation coefficient between any two regions of interest, and a brain network map is constructed based on the brain network BN matrix. The brain network diagram is input into the trained autism diagnosis model to diagnose autism and obtain the raw score of the predicted category. Convert the raw scores of the predicted categories into diagnostic results; Inputting the brain network diagram into a trained autism diagnostic model for autism diagnosis includes: Based on the brain network diagram, feature extraction and average pooling are performed using a multi-teacher module to obtain the multi-teacher brain map representation features. Based on the brain network diagram, a subgraph generator is used to generate a multi-level subgraph for each node; Based on the multi-teacher mind map representation features and the multi-level subgraph of each node, the aggregated mind map representation is generated using the subgraph aggregation module. Based on the aggregated brain map representation, a multilayer perceptron is used to generate the original scores for the predicted categories.

[0008] Optionally, the preprocessing of the acquired functional magnetic resonance imaging (fMRI) data of the subject at rest to obtain time-series data of blood oxygenation level-dependent signals for each region of interest in the brain includes: The obtained functional magnetic resonance brain imaging data of the subjects in the resting state were subjected to slice time correction, head movement correction, noise signal regression and time filtering to obtain the pre-processed resting state magnetic resonance imaging data. The resting-state magnetic resonance imaging data after preliminary processing was normalized to the MNI space using a nonlinear registration algorithm to obtain normalized resting-state magnetic resonance imaging data. Based on the anatomical automatic labeling template AAL, the brain in the normalized resting-state magnetic resonance imaging data is divided into multiple regions of interest, and the average time series of each region of interest is extracted to obtain the blood oxygen level dependent signal time series data of each region of interest in the brain.

[0009] Optionally, the calculation of the Pearson correlation coefficient between any two regions of interest is achieved using the following formula: , in, Indicates the first Areas of interest With the Areas of interest The Pearson correlation coefficient between them , This represents the total number of regions of interest. Indicates the first Areas of interest Blood oxygen levels depend on signal time-series data. Indicates the first Areas of interest Blood oxygen levels depend on signal time-series data. Expressing expectations, Indicates the first Areas of interest Blood oxygen levels depend on the average value of the signal time-series data. Indicates the first Areas of interest Blood oxygen levels depend on the average value of the signal time-series data. Indicates the first Areas of interest Blood oxygen levels depend on the standard deviation of the signal time series data. Indicates the first Areas of interest Blood oxygen levels depend on the standard deviation of the signal time series data. Indicates the total number of time steps. Indicates the first Areas of interest exist Blood oxygen levels at time steps depend on signals. Indicates the first Areas of interest exist Blood oxygen levels at time steps are signal-dependent.

[0010] Optionally, constructing a brain network graph based on the brain network BN matrix includes: By setting the elements in the brain network's BN matrix that correspond to the first p% of the values ​​to 1, and setting the remaining elements to 0, we obtain the adjacency matrix. ;in, This is the default value; Each region of interest is treated as a node, and the brain network's BN matrix is ​​used as the feature matrix for all nodes, combined with the adjacency matrix. Building a brain network map ;in, , Represents a set of nodes. , , This represents the total number of regions of interest, which is also the total number of nodes. Indicates the first 1 node Denotes the set of edges. , , Represents a node and nodes The edge between, Represents the feature matrix of all nodes. , Represents a node eigenvectors.

[0011] Optionally, the multi-teacher module includes Each teacher module consists of a GCN module, an average pooling module, and a multilayer perceptron. The process involves extracting features and performing average pooling based on the brain network map using a multi-teacher module to obtain multi-teacher brain map representation features, including: Input the brain network diagram into... Multi-teacher mind map representation features were extracted from each teacher module to obtain multi-teacher mind map representation features. ;in, ,in, Indicates the first A teacher's mind map represents a characteristic; Each teacher module performs the following steps during the training process: The GCN module is used to extract features from the brain network map, and the output features of the GCN module are obtained. The average pooling module is used to perform average pooling on the output features of the GCN module to obtain the teacher's mind map representation features. Based on the teacher's mind map representation features, a multilayer perceptron is used to generate the original score for the teacher's predicted category.

[0012] Optionally, based on the brain network diagram, a subgraph generator is used to generate multi-level subgraphs for each node, including: For any node Set the length of a single walk. It iteratively executes the subgraph generation step: using the nodo2vec graph embedding method, with nodes... Execution in the brain network diagram as the starting point A subindependent biased random walk, with a single walk length of... , obtain node of indivual Subgraph, settings ;in, As a preset value, Indicates the first One node; like Stop the subgraph generation step and obtain the nodes. Multi-order subgraph ;in, Indicates the preset order. , Represents a node The set of subgraph nodes, , This represents the union operator. Indicates the first The set of nodes visited by a biased random walk. Represents a node. Represents a node The set of edges of the subgraph Represents a node subgraph eigenmatrix Represents a node subgraph The adjacency matrix, Indicates an index. Represents a node The indivual Step diagram.

[0013] Optionally, the step of generating an aggregated mind map representation using a subgraph aggregation module based on the multi-teacher mind map representation features and the multi-level subgraphs of each node includes: Features of multi-teacher mind map representation By performing splicing and linear transformation, the fused mind map representation features are obtained. ;in, , Represents linear transformation operations. This indicates a concatenation operation. Indicates the first A teacher's mind map represents a characteristic; The multi-order subgraphs of all nodes are superimposed and input into the GCN. After average pooling, the multi-order feature representation matrix is ​​obtained. ;in, , Represents a node of Rank feature representation, Indicates the preset order. Indicates the total number of nodes; Multi-order feature representation matrix With fusion of brain map representation features After performing matrix multiplication, the matrix is ​​input into the softmax activation layer to obtain a multi-order weight matrix. ;in, , This represents the softmax activation function. This represents the matrix multiplication operator. Indicates transpose. Represents a node of Order weights; For multi-order weight matrices and multi-order feature representation matrix Element-wise matrix multiplication and node aggregation are performed to obtain a multi-order brain map representation. ;in, , This represents element-wise matrix multiplication. express Hierarchical brain map representation; The multi-level brain map representation By performing multi-level splicing, a spliced ​​brain map-level representation is obtained. ;in, ; Brain-level representation after splicing Perform linear projection to obtain the query matrix. Key matrix Sum matrix ;in, , , , This indicates a learnable query projection matrix. Represents the learnable key projection matrix. Represents the projection matrix of learnable values; Based on the query matrix Key matrix Sum matrix Perform scaling dot product attention operations and row-wise summation pooling operations to obtain the aggregated mind map representation. ;in, , This represents the scaled dot product attention matrix. This indicates a row-wise summation pooling operation. This represents the total number of rows in the scaled dot product attention matrix. , This represents the scaling factor.

[0014] Optionally, the multilayer perceptron includes two linear layers and a ReLU activation layer; The original score of the predicted category is obtained using the following formula: , in, Indicates the first linear layer. Indicates the second linear layer. Indicates the ReLU activation layer. The mind map represents the aggregated representation. This represents the original score of the predicted category; The diagnostic results are obtained using the following formula: , in, This represents the softmax activation function. This represents the index value function that retrieves the maximum value. This indicates the diagnostic result.

[0015] Optionally, the loss function of the autism diagnostic model is as follows: , in, Indicates the total loss. Indicates the loss weight. Represents classification loss, This indicates the loss from distillation by multiple teachers; The classification loss Calculated using the following formula: , in, Represents the cross-entropy loss function. Indicates the true label, This represents the original score of the predicted category; The multi-teacher distillation loss Calculated using the following formula: , in, This indicates the total number of teacher modules. Denotes the KL divergence function. This represents the softmax activation function. This represents the Logsoftmax activation function. Indicates the distillation temperature. Indicates the first Each teacher predicted the original score for the category.

[0016] Secondly, the present invention provides an autism diagnostic system based on an interpretable brain map neural network with multi-teacher distillation, comprising: The preprocessing module is used to: preprocess the acquired functional magnetic resonance brain imaging data of the subject in a resting state to obtain time-series data of blood oxygenation level dependent signals in various brain regions of interest; The brain network map construction module is used to: construct a brain network BN matrix by calculating the Pearson correlation coefficient between any two regions of interest based on the time-series data of blood oxygen level dependent signals in various brain regions of interest, and construct a brain network map based on the brain network BN matrix; The autism diagnosis module is used to: input brain network diagrams into a trained autism diagnosis model to diagnose autism and obtain the original score of the predicted category; The conversion module is used to convert the raw scores of the predicted categories into diagnostic results. Inputting the brain network diagram into a trained autism diagnostic model for autism diagnosis includes: Based on the brain network diagram, feature extraction and average pooling are performed using a multi-teacher module to obtain the multi-teacher brain map representation features. Based on the brain network diagram, a subgraph generator is used to generate a multi-level subgraph for each node; Based on the multi-teacher mind map representation features and the multi-level subgraph of each node, the aggregated mind map representation is generated using the subgraph aggregation module. Based on the aggregated brain map representation, a multilayer perceptron is used to generate the original scores for the predicted categories.

[0017] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: 1. The autism diagnosis method based on interpretable brain graph neural network with multi-teacher distillation provided by this invention guides the student model to learn by capturing BN topological knowledge at different scales through multiple GCN teacher models. It combines attention mechanism to adaptively fuse information from different subgraphs, enhances the model's representation learning ability, and can effectively realize the judgment of autism. It can provide a reference for doctors' clinical diagnosis. 2. The autism diagnosis system based on interpretable brain map neural network with multi-teacher distillation provided by this invention achieves autism diagnosis by setting up a preprocessing module, a brain network map construction module, an autism diagnosis module and a conversion module, which has practical significance and good application prospects. Attached Figure Description

[0018] Figure 1 A schematic diagram of an autism diagnosis method based on an interpretable brain map neural network with multi-teacher distillation, provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the autism diagnostic model structure provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the top 10 discriminative brain regions most relevant to autism identification identified by the autism diagnostic model provided in this embodiment of the invention in each subgraph sampling. Figure 4 The diagram shows the functional connections related to autism in each sub-diagram provided in the embodiments of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.

[0020] Example 1

[0021] This invention discloses an autism diagnosis method based on a multi-teacher distillation interpretable brain map neural network, with reference to... Figure 1 As shown, the specific steps include the following: S1, preprocess the acquired functional magnetic resonance brain imaging data of the subject in the resting state to obtain the time-series data of blood oxygenation level dependent signals of various brain regions of interest; S2, based on the time-series data of blood oxygenation level dependent signals in various brain regions of interest, construct the brain network BN matrix by calculating the Pearson correlation coefficient between any two regions of interest, and construct a brain network map based on the brain network BN matrix; S3: Input the brain network diagram into the trained autism diagnosis model to diagnose autism and obtain the original score of the predicted category. S4 converts the raw scores of the predicted categories into diagnostic results.

[0022] Specifically, in step S1, this embodiment preprocesses the acquired functional magnetic resonance imaging data of the subject in a resting state, divides the regions of interest according to the anatomical template of the brain structure, and obtains the time series of blood oxygen level dependent signals for each region of interest.

[0023] The preprocessing steps include: The obtained functional magnetic resonance brain imaging data of the subjects in the resting state were subjected to slice time correction, head movement correction, noise signal regression and time filtering to obtain the pre-processed resting state magnetic resonance imaging data. The resting-state magnetic resonance imaging data after preliminary processing was normalized to the MNI (Montreal Neurological Institute) space using the nonlinear registration algorithm ANTS, resulting in normalized resting-state magnetic resonance imaging data. Based on Anatomical Automatic Labeling (AAL), the brain in normalized resting-state magnetic resonance imaging data is divided into multiple regions of interest, and the average time series of each region of interest is extracted to obtain the blood oxygenation level dependent signal time series data of each region of interest in the brain.

[0024] In step S2, the Pearson correlation coefficient between any two regions of interest is obtained using the following formula: , in, Indicates the first Areas of interest With the Areas of interest The Pearson correlation coefficient between them , This represents the total number of regions of interest. Indicates the first Areas of interest Blood oxygen levels depend on signal time-series data. Indicates the first Areas of interest Blood oxygen levels depend on signal time-series data. Expressing expectations, Indicates the first Areas of interest Blood oxygen levels depend on the average value of the signal time-series data. Indicates the first Areas of interest Blood oxygen levels depend on the average value of the signal time-series data. Indicates the first Areas of interest Blood oxygen levels depend on the standard deviation of the signal time series data. Indicates the first Areas of interest Blood oxygen levels depend on the standard deviation of the signal time series data. Indicates the total number of time steps. Indicates the first Areas of interest exist Blood oxygen levels at time steps depend on signals. Indicates the first Areas of interest exist Blood oxygen levels at time steps are signal-dependent.

[0025] The construction of the brain network graph based on the brain network BN matrix includes: By setting the elements in the brain network's BN matrix that correspond to the first p% of the values ​​to 1, and setting the remaining elements to 0, we obtain the adjacency matrix. ;in, This is the default value; Each region of interest is treated as a node, and the brain network's BN matrix is ​​used as the feature matrix for all nodes, combined with the adjacency matrix. Building a brain network map ;in, , Represents a set of nodes. , , This represents the total number of regions of interest, which is also the total number of nodes. Indicates the first 1 node Denotes the set of edges. , , Represents a node and nodes The edge between, Represents the feature matrix of all nodes. , Represents a node eigenvectors.

[0026] In step S3, refer to Figure 2 As shown, the autism diagnostic model constructed in this embodiment includes a multi-teacher module, a subgraph generator, a subgraph aggregation module, and a multilayer perceptron. In this embodiment, the multi-teacher module is pre-trained, and knowledge distillation technology is used to transfer multi-scale brain network graph topology knowledge from multiple pre-trained teacher models to guide the training of the student model. In this embodiment, the student model refers to the GCN layer of the subgraph aggregation module.

[0027] The multi-teacher module includes Each teacher module includes a GCN module, an average pooling module, and a multilayer perceptron; pre-trained multi-teacher modules learn topological knowledge of the brain network graph at different scales by setting different numbers of GCN modules in each teacher module; in order to capture the topological knowledge of each node... Topological information, this embodiment adopts The first-order GCN module encodes this information; the second-order GCN module... In a message passing mechanism, a hidden layer can be represented as: , in, Brain network diagram as input Feature matrix , Represents the first GCN module The output of the hidden layer, Represents the first GCN module The weight matrix of the hidden layer, Represents the first GCN module The output of the hidden layer, Brain network diagram The adjacency matrix with added self-connections, , Represents the identity matrix. express The angle matrix, This represents the ReLU activation function.

[0028] Each teacher module performs the following steps during the training process: The GCN module is used to extract features from the brain network map, and the output features of the GCN module are obtained. The average pooling module is used to perform average pooling on the output features of the GCN module to obtain the teacher's mind map representation features. Based on the teacher's mind map representation features, a multilayer perceptron is used to generate the original score for the teacher's predicted category.

[0029] The expression for the average pooling module is as follows: , in, Indicates the first The characteristics of teacher mind map representation for each teacher module. Indicates the first The first feature of the GCN module output of the teacher module. Line, that is, the node-related features in the output features of the GCN module. eigenvectors.

[0030] Inputting brain network diagrams into a trained autism diagnostic model for autism diagnosis includes: S3.1 Based on the brain network diagram, feature extraction and average pooling are performed using a multi-teacher module to obtain the multi-teacher brain map representation features; S3.2, Based on the brain network diagram, use a subgraph generator to generate a multi-level subgraph for each node; S3.3, Based on the multi-teacher mind map representation features and the multi-level subgraphs of each node, the subgraph aggregation module is used to generate the aggregated mind map representation; S3.4, Based on the aggregated brain map representation, generate the original score for the predicted category using a multilayer perceptron.

[0031] In step S3.1, the brain network diagram is input into... Multi-teacher mind map representation features were extracted from each teacher module to obtain multi-teacher mind map representation features. ;in, ,in, Indicates the first A teacher's mind map represents a characteristic.

[0032] In step S3.2, the subgraph generator module aims to collect different subgraphs for each node by performing random walks on the brain network graph; based on the brain network graph, the subgraph generator generates multi-order subgraphs for each node, including: For any node Set the length of a single walk. It iteratively executes the subgraph generation step: using the nodo2vec graph embedding method, with nodes... Execution in the brain network diagram as the starting point A subindependent biased random walk, with a single walk length of... , obtain node of indivual Subgraph, settings ;in, As a preset value, Indicates the first One node; like Stop the subgraph generation step and obtain the nodes. Multi-order subgraph ;in, Indicates the preset order. , Represents a node The set of subgraph nodes, , This represents the union operator. Indicates the first The set of nodes visited by a biased random walk. Represents a node. Represents a node The set of edges of the subgraph Represents a node subgraph eigenmatrix Represents a node subgraph The adjacency matrix, Indicates an index. Represents a node The indivual Step diagram.

[0033] In step S3.3, this embodiment introduces a subgraph aggregation module based on an attention mechanism to aggregate subgraphs of different orders to learn graph-level representations, specifically including: To find information-rich subgraphs guided by multi-scale representations of multi-teacher modules, we first analyze the features of multi-teacher mind map representations. The data is concatenated and then input into a linear layer to obtain the fused mind map representation features. ;in, , Represents linear transformation operations. This indicates a concatenation operation. Indicates the first A teacher's mind map represents a characteristic; The multi-order subgraphs of all nodes are superimposed and input into the GCN. After average pooling, the multi-order feature representation matrix is ​​obtained. ;in, , Represents a node of Rank feature representation, Indicates the preset order. Indicates the total number of nodes; Multi-order feature representation matrix With fusion of brain map representation features After performing matrix multiplication, the matrix is ​​input into the softmax activation layer to obtain a multi-order weight matrix. ;in, , This represents the softmax activation function. This represents the matrix multiplication operator. Indicates transpose. Represents a node of Order weights; For multi-order weight matrices and multi-order feature representation matrix Element-wise matrix multiplication and node aggregation are performed to obtain a multi-order brain map representation. ;in, , This represents element-wise matrix multiplication. express Hierarchical brain map representation; The multi-level brain map representation By performing multi-level splicing, a spliced ​​brain map-level representation is obtained. ;in, ; Brain-level representation after splicing Perform linear projection to obtain the query matrix. Key matrix Sum matrix ;in, , , , This indicates a learnable query projection matrix. Represents the learnable key projection matrix. Represents the projection matrix of learnable values; Based on the query matrix Key matrix Sum matrix Perform scaling dot product attention operations and row-wise summation pooling operations to obtain the aggregated mind map representation. ;in, , This represents the scaled dot product attention matrix. This indicates a row-wise summation pooling operation. This represents the total number of rows in the scaled dot product attention matrix. , This represents the scaling factor.

[0034] In step S3.4, the multilayer perceptron includes two linear layers and a ReLU activation layer; the original score for the predicted category is obtained using the following formula: , in, Indicates the first linear layer. Indicates the second linear layer. Indicates the ReLU activation layer. The mind map represents the aggregated representation. This represents the original score for the predicted category.

[0035] The autism diagnostic model BGNN proposed in this embodiment extracts effective information from the brain network graph BN by integrating discriminant subgraphs and utilizing multi-scale topological information from a pre-trained teacher model, thus reducing the total loss. Defined as classification loss And multiple teachers' distillation losses Combination, classification loss Multi-teacher distillation loss is used to reduce the distance between predicted and true labels. The difference between the two probability distributions is used for the student model and each teacher model.

[0036] The loss function of the autism diagnostic model is as follows: , in, Indicates the loss weight. Represents classification loss, This indicates the loss from distillation by multiple teachers; Classification loss Calculated using the following formula: , in, Represents the cross-entropy loss function. Indicates the true label, This represents the original score of the predicted category; Multi-teacher distillation loss Calculated using the following formula: , in, This indicates the total number of teacher modules. Denotes the KL divergence function. This represents the softmax activation function. This represents the Logsoftmax activation function. Indicates the distillation temperature. Indicates the first Each teacher predicted the original score for the category.

[0037] In step S4, the diagnostic result is obtained using the following formula: , in, This represents the softmax activation function. This represents the index value function that retrieves the maximum value. This indicates the diagnostic result.

[0038] In this embodiment, using the ABIDE dataset as training samples, the following experiments were conducted on an autism diagnosis method based on a multi-teacher distillation interpretable brain map neural network: In this embodiment, all resting-state functional magnetic resonance imaging (rs-fMRI) samples were obtained from the Autism Brain Imaging Data Exchange (ABIDE). The ABIDE database aggregates data from 17 different collection sites, including 531 autistic patients (ASD) and 571 normal controls (NC). This embodiment selected data from the three sites with the largest sample sizes: NYU, UM, and USM, each with more than 100 participants. Specifically, NYU contains 79 ASD patients and 105 NCs, UM contains 68 ASD patients and 77 NCs, and the USM site contains 58 ASD patients and 43 NCs.

[0039] The data used were preprocessed using C-PAC, an open-source software package developed by FCP-INDI. Preprocessing steps included slice time and head motion correction, noise signal regression, and temporal filtering. Then, the preprocessed rs-fMRI samples were normalized to the MNI space using the nonlinear registration algorithm ANTS. Subsequently, an automatic anatomical labeling template was used to divide 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 subjects is shown in Table 1, where Site represents the site, Category represents the category, Scan represents the number, Age represents age, M represents male, F represents female, and Gender (M / F) represents the male-to-female ratio.

[0040] Table 1. Demographic information of the ABIDE dataset sample

[0041]

[0042] The autism diagnostic model was implemented using the PyTorch framework, specifically PyTorch version 2.5.0 and CUDA version 12.4. The training batch size was 1, the learning rate was 1e-4, and the number of training epochs was 150. In the pre-trained multi-teacher module, three GCN modules of different depths were used to capture multi-scale Batch Normalization (BN) features. The first GCN module, GCN-1, contained a single hidden layer with a weight matrix of size 116×128. The second GCN module, GCN-2, contained two hidden layers with weight matrices of sizes 116×128 and 128×128, respectively. The third GCN module, GCN-3, contained three hidden layers with weight matrices of sizes 116×128, 128×128, and 128×128, respectively. Hidden layers of size 28 and 128×128 are used; a ReLU activation function is added after each hidden layer; in the subgraph aggregation module, a single-layer GCN with a weight matrix of size 116×128 is used as the student model, and a ReLU activation function is added after it; the number of repetitions of the random walk is set to 5, and the walk lengths are set to 2, 3, and 4 respectively; in the subgraph aggregation module, the weight matrix of the linear layer is 128×128; the linear layer used to fuse multi-scale BN representations has a dimension of (128×3, 128) and is equipped with a ReLU activation function; the multilayer perceptron (MLP) consists of two fully connected layers with a dimension of (128, 32, 2), where the first layer is equipped with a ReLU activation function. The loss function... and Set them to 4 and 0.2 respectively.

[0043] After model training, five-fold cross-validation is used to evaluate the model's classification performance on the dataset. Specifically, the dataset is first randomly divided into five roughly equal subsets; then, one subset is used as the test set, and the remaining subsets are used as the training set; this process is repeated five times to ensure that each subset is used as the test set once.

[0044] For performance evaluation, six metrics were used: accuracy (ACC), sensitivity (SEN), specificity (SPE), F1 score, precision (PRE), and area under the receiver operating characteristic curve (AUC). ACC represents the proportion of all samples correctly classified, SEN represents the proportion of actual patients correctly classified by the model, and SPE represents the proportion of actual healthy individuals correctly classified by the model. The average of the five-fold cross-validation results for each metric was taken as the final result.

[0045] To verify the effectiveness of the BGNN model proposed in this embodiment, it was compared with the baseline model BrainNetCNN, which uses convolution, and also with three other types of models. The first type is graph neural network-based models GCN (Graph Convolutional Network), GAT (Graph Attention Network), and GIN (Graph Isomorphism Network); the second type is Transformer-based models BrainNetTF (Brain Network Transformer) and ConnectomeTF (Connectome Transformer); and the third type is knowledge distillation-based models KD (Knowledge Distillation), LSP (Local Structure Preserving), G-CRD (Graph Contrastive Representation Distillation), and DGKD (Decoupled Graph Knowledge Distillation). The comparison results of the BGNN model with the above 10 models are shown in Table 2.

[0046] Table 2. Comparison results of the BGNN model with 10 other models on the ABIDE dataset.

[0047] The following conclusions can be drawn from Table 2: First, in terms of accuracy, the graph neural network-based methods GCN, GIN, and GAT outperform BrainNetCNN. For example, GCN achieved an accuracy of 65.21% on the NYU site, significantly higher than BrainNetCNN's 61.43%. This may be due to the inherent graph structure of BN, which allows graph neural networks to better extract topological information relevant to brain disease classification. Second, the Transformer-based methods BrainNetTF and ConnectomeTF consistently outperform graph neural network-based methods. Transformers can capture long-range dependencies between ROIs, which is often crucial for brain disease classification, while GNNs are usually limited to local neighborhood aggregation. Third, traditional KD-based methods KD, LSP, G-CRD, and DGKD sometimes underperform Transformer-based methods. For example, on the UM site, G-CRD achieved an accuracy of 72.24%, lower than BrainNetTF. The accuracy of BGNN is 73.29%; this may be because limited medical data makes it difficult to pre-train a robust teacher model to transfer knowledge to the student model for brain disease classification. Finally, it can be observed that BGNN achieves state-of-the-art results on all datasets. Specifically, BGNN's mean accuracy and area under the curve (AUC) across all sites are 73.47% and 75.58%, respectively, representing performance improvements of 2.86% and 2.85% compared to the second-best accuracy and AUC of 70.61% and 72.73%, respectively. These results demonstrate that BGNN effectively utilizes multi-teacher knowledge distillation to preserve multi-scale topological information in BN, thereby guiding the extraction of discriminant subgraphs to improve brain disease classification.

[0048] This embodiment Figure 3 This image shows the top 10 discriminative brain regions most relevant to ASD recognition identified by BGNN in sub-map sampling at various levels. Specifically, PAL.R represents the globus pallidus (right), THA.R represents the thalamus (right), CAU.R represents the caudate nucleus (right), CAU.L represents the caudate nucleus (left), THA.L represents the thalamus (left), HIP.R represents the hippocampus (right), AMYG.R represents the amygdala (right), HIP.L represents the hippocampus (left), and PHG.R represents the parahippocampal gyrus (right). CAL.L represents the calcarine fissure (left), Vermis7 represents the cerebellar vermis VII, HIP.L represents the hippocampus (left), CRBL6.R represents the cerebellar lobule VI (right), and HIP.R represents the hippocampus (right). As can be seen from the figure, brain regions such as the amygdala (AMYG), hippocampus (HIP), caudate nucleus (CAU), and thalamus (THA) are frequently identified in the second, third, and fourth sub-graphs. These brain regions have been confirmed in previous studies to be closely related to the diagnosis of ASD.

[0049] Figure 4 Functional connectivity associated with ASD is illustrated in subgraphs at various levels to explain its potential pathogenesis. In the figures, PAL.L represents the globus pallidus (left), PoCG.L represents the postcentral gyrus (left), PUT.R represents the putamen (right), STG.R represents the superior temporal gyrus (right), CRBL6.R represents the cerebellar lobule VI (right), PUT.L represents the putamen (left), ROL.L represents the central tegmentum (left), ROL.R represents the central tegmentum (right), PHG.L represents the parahippocampal gyrus (left), TPOsup.R represents the superior temporal pole (right), TPOsup.L represents the superior temporal pole (left), AMYG.L represents the amygdala (left), and ANG.L represents the angular gyrus (left). It can be seen that, unlike NC, the abnormal connectivity patterns in ASD patients are mainly distributed around the AMYG, which is consistent with previous findings. In addition, NC shows strong connectivity in regions such as THA and the superior temporal gyrus STG, while such connectivity patterns were not observed in ASD patients, a finding consistent with several previous studies.

[0050] In summary, the autism diagnosis method based on interpretable brain graph neural networks with multi-teacher distillation proposed in this embodiment guides the student model to learn by capturing BN topological knowledge at different scales through multiple GCN teacher models. It also enhances the model's representation learning ability by adaptively fusing information from different subgraphs using an attention mechanism. This method can effectively diagnose autism and provide a reference for doctors' clinical diagnosis.

[0051] Example 2

[0052] Based on the same inventive concept as Embodiment 1, this embodiment of the invention discloses an autism diagnostic system based on an interpretable brain map neural network with multi-teacher distillation, comprising: The preprocessing module is used to: preprocess the acquired functional magnetic resonance brain imaging data of the subject in a resting state to obtain time-series data of blood oxygenation level dependent signals in various brain regions of interest; The brain network map construction module is used to: construct a brain network BN matrix by calculating the Pearson correlation coefficient between any two regions of interest based on the time-series data of blood oxygen level dependent signals in various brain regions of interest, and construct a brain network map based on the brain network BN matrix; The autism diagnosis module is used to: input brain network diagrams into a trained autism diagnosis model to diagnose autism and obtain the original score of the predicted category; The conversion module is used to convert the raw scores of the predicted categories into diagnostic results. Inputting the brain network diagram into a trained autism diagnostic model for autism diagnosis includes: Based on the brain network diagram, feature extraction and average pooling are performed using a multi-teacher module to obtain the multi-teacher brain map representation features. Based on the brain network diagram, a subgraph generator is used to generate a multi-level subgraph for each node; Based on the multi-teacher mind map representation features and the multi-level subgraph of each node, the aggregated mind map representation is generated using the subgraph aggregation module. Based on the aggregated brain map representation, a multilayer perceptron is used to generate the original scores for the predicted categories.

[0053] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0054] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0055] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0056] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0057] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0058] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. An autism diagnostic system based on an interpretable brain map neural network with multi-teacher distillation, characterized in that, include: The preprocessing module is used to: preprocess the acquired functional magnetic resonance brain imaging data of the subject in a resting state to obtain time-series data of blood oxygenation level dependent signals in various brain regions of interest; The brain network map construction module is used to: construct a brain network BN matrix by calculating the Pearson correlation coefficient between any two regions of interest based on the time-series data of blood oxygen level dependent signals in various brain regions of interest, and construct a brain network map based on the brain network BN matrix; The autism diagnosis module is used to: input brain network diagrams into a trained autism diagnosis model to diagnose autism and obtain the original score of the predicted category; The conversion module is used to convert the raw scores of the predicted categories into diagnostic results. Inputting the brain network diagram into a trained autism diagnostic model for autism diagnosis includes: Based on the brain network diagram, feature extraction and average pooling are performed using a multi-teacher module to obtain the multi-teacher brain map representation features. Based on the brain network diagram, a subgraph generator is used to generate a multi-level subgraph for each node; Based on the multi-teacher mind map representation features and the multi-level subgraph of each node, the aggregated mind map representation is generated using the subgraph aggregation module. Based on the aggregated brain map representation, a multilayer perceptron is used to generate the original scores for the predicted categories. The multi-teacher module includes Each teacher module consists of a GCN module, an average pooling module, and a multilayer perceptron. The process involves extracting features and performing average pooling based on the brain network map using a multi-teacher module to obtain multi-teacher brain map representation features, including: Input the brain network diagram into... Multi-teacher mind map representation features were extracted from each teacher module to obtain multi-teacher mind map representation features. ;in, ,in, Indicates the first A teacher's mind map represents a characteristic; Each teacher module performs the following steps during the training process: The GCN module is used to extract features from the brain network map, and the output features of the GCN module are obtained. The average pooling module is used to perform average pooling on the output features of the GCN module to obtain the teacher's mind map representation features. Based on the teacher's mind map representation features, a multilayer perceptron is used to generate the original score of the teacher's predicted category. Based on the brain network diagram, a subgraph generator is used to generate multi-order subgraphs for each node, including: For any node Set the length of a single walk. It iteratively executes the subgraph generation step: using the nodo2vec graph embedding method, with nodes... Execution in the brain network diagram as the starting point A subindependent biased random walk, with a single walk length of... , obtain node of indivual Subgraph, settings ;in, As a preset value, Indicates the first One node; like Stop the subgraph generation step and obtain the nodes. Multi-order subgraph ;in, Indicates the preset order. , Represents a node The set of subgraph nodes, , This represents the union operator. Indicates the first The set of nodes visited by a biased random walk. Represents a node. Represents a node The set of edges of the subgraph Represents a node subgraph eigenmatrix Represents a node subgraph The adjacency matrix, Indicates an index. Represents a node The indivual Step diagram; The step of generating an aggregated mind map representation based on the multi-teacher mind map representation features and the multi-level subgraphs of each node using a subgraph aggregation module includes: Features of multi-teacher mind map representation By performing splicing and linear transformation, the fused mind map representation features are obtained. ;in, , Represents linear transformation operations. This indicates a concatenation operation. Indicates the first A teacher's mind map represents a characteristic; The multi-order subgraphs of all nodes are superimposed and input into the GCN. After average pooling, the multi-order feature representation matrix is ​​obtained. ;in, , Represents a node of Rank feature representation, Indicates the preset order. Indicates the total number of nodes; Multi-order feature representation matrix With fusion of brain map representation features After performing matrix multiplication, the matrix is ​​input into the softmax activation layer to obtain a multi-order weight matrix. ;in, , This represents the softmax activation function. This represents the matrix multiplication operator. Indicates transpose. Represents a node of Order weights; For multi-order weight matrices and multi-order feature representation matrix Element-wise matrix multiplication and node aggregation are performed to obtain a multi-order brain map representation. ;in, , This represents element-wise matrix multiplication. express Hierarchical brain map representation; The multi-level brain map representation By performing multi-level splicing, a spliced ​​brain map-level representation is obtained. ;in, ; Brain-level representation after splicing Perform linear projection to obtain the query matrix. Key matrix Sum matrix ;in, , , , This indicates a learnable query projection matrix. Represents the learnable key projection matrix. Represents the projection matrix of learnable values; Based on the query matrix Key matrix Sum matrix Perform scaling dot product attention operations and row-wise summation pooling operations to obtain the aggregated mind map representation. ;in, , This represents the scaled dot product attention matrix. This indicates a row-wise summation pooling operation. This represents the total number of rows in the scaled dot product attention matrix. , This represents the scaling factor.

2. The autism diagnostic system based on interpretable brain map neural networks using multi-teacher distillation as described in claim 1, characterized in that, The preprocessing of the acquired functional magnetic resonance imaging (fMRI) data of the subject at rest yields time-series data of blood oxygenation level-dependent signals for various brain regions of interest, including: The obtained functional magnetic resonance brain imaging data of the subjects in the resting state were subjected to slice time correction, head movement correction, noise signal regression and time filtering to obtain the pre-processed resting state magnetic resonance imaging data. The resting-state magnetic resonance imaging data after preliminary processing was normalized to the MNI space using a nonlinear registration algorithm to obtain normalized resting-state magnetic resonance imaging data. Based on the anatomical automatic labeling template AAL, the brain in the normalized resting-state magnetic resonance imaging data is divided into multiple regions of interest, and the average time series of each region of interest is extracted to obtain the blood oxygen level dependent signal time series data of each region of interest in the brain.

3. The autism diagnostic system based on interpretable brain map neural networks using multi-teacher distillation according to claim 1, characterized in that, The Pearson correlation coefficient between any two regions of interest is calculated using the following formula: ， in, Indicates the first Areas of interest With the Areas of interest The Pearson correlation coefficient between them , This represents the total number of regions of interest. Indicates the first Areas of interest Blood oxygen levels depend on signal time-series data. Indicates the first Areas of interest Blood oxygen levels depend on signal time-series data. Expressing expectations, Indicates the first Areas of interest Blood oxygen levels depend on the average value of the signal time-series data. Indicates the first Areas of interest Blood oxygen levels depend on the average value of the signal time-series data. Indicates the first Areas of interest Blood oxygen levels depend on the standard deviation of the signal time series data. Indicates the first Areas of interest Blood oxygen levels depend on the standard deviation of the signal time series data. Indicates the total number of time steps. Indicates the first Areas of interest exist Blood oxygen levels at time steps depend on signals. Indicates the first Areas of interest exist Blood oxygen levels at time steps are signal-dependent.

4. The autism diagnostic system based on interpretable brain map neural networks using multi-teacher distillation according to claim 1, characterized in that, The construction of the brain network graph based on the brain network BN matrix includes: By setting the elements in the brain network's BN matrix that correspond to the first p% of the values ​​to 1, and setting the remaining elements to 0, we obtain the adjacency matrix. ;in, This is the default value; Each region of interest is treated as a node, and the brain network's BN matrix is ​​used as the feature matrix for all nodes, combined with the adjacency matrix. Building a brain network map ;in, , Represents a set of nodes. , , This represents the total number of regions of interest, which is also the total number of nodes. Indicates the first 1 node Denotes the set of edges. , , Represents a node and nodes The edge between, Represents the feature matrix of all nodes. , Represents a node eigenvectors.

5. The autism diagnostic system based on interpretable brain map neural networks using multi-teacher distillation according to claim 1, characterized in that, The multilayer perceptron comprises two linear layers and a ReLU activation layer; the original score of the predicted category is obtained by the following formula: ， in, Indicates the first linear layer. Indicates the second linear layer. Indicates the ReLU activation layer. The mind map represents the convergence of elements. This represents the original score of the predicted category; The diagnostic results are obtained using the following formula: ， in, This represents the softmax activation function. This represents the index value function that retrieves the maximum value. This indicates the diagnostic result.

6. The autism diagnostic system based on interpretable brain map neural networks using multi-teacher distillation according to claim 1, characterized in that, The loss function of the autism diagnostic model is as follows: ， in, Indicates the total loss. Indicates the loss weight. Represents classification loss. This indicates the loss from distillation by multiple teachers; The classification loss Calculated using the following formula: ， in, Represents the cross-entropy loss function. Indicates the true label, This represents the original score of the predicted category; The multi-teacher distillation loss Calculated using the following formula: ， in, This indicates the total number of teacher modules. Denotes the KL divergence function. This represents the softmax activation function. This represents the Logsoftmax activation function. Indicates the distillation temperature. Indicates the first Each teacher predicted the original score for the category.

Images