An interpretable dynamic brain network analysis method based on functional and structural priors

Abstract

The application discloses an interpretable dynamic brain network analysis method based on function and structure priori, which comprises the following steps: preprocessing fMRI signals, dividing the fMRI signals to construct a dynamic functional brain network through non-overlapping sliding windows; embedding the function and structure priori into the dynamic functional brain network to obtain a hierarchical dynamic functional brain network representation based on brain region clustering, capturing the spatiotemporal topological relationship between functional subnets, then extracting spatiotemporal topological features from the hierarchical dynamic functional brain network through a tensor graph convolution network with intra-graph propagation and inter-graph propagation, and enhancing the feature representation consistency within the subnets and the difference between the feature representations of the subnets through the function subnet constraint, so that the topological prior characteristics of the brain network are better reflected. Finally, the global brain spatiotemporal embedding obtained through feature fusion combined with an attention mechanism is embedded into a multilayer perceptron to predict the category of each subject, and the influence of brain diseases on the spatiotemporal topological properties of the brain is analyzed.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, specifically relating to brain imaging image processing technology, and more particularly to an interpretable dynamic brain network analysis method based on functional and structural priors. Background Technology

[0002] In recent years, with the continuous development of science and technology, brain functional imaging technology has become one of the most popular research hotspots and cutting-edge directions in the field of brain science. Among them, functional magnetic resonance imaging (fMRI) combines functional, anatomical, and imaging information, which can not only show the location, size, and extent of brain functional activation areas, but also directly show the exact anatomical location of the activation areas. It has many superior characteristics such as being non-invasive, non-traumatic, radiation-free, repeatable, precisely localizable, and having high temporal and spatial resolution, and is therefore widely used in brain science research in various fields.

[0003] fMRI-based brain functional network connectivity analysis is an important aspect of neuroscience research using this technology. Because the brain is a highly complex system constantly in flux, quantitatively describing the dynamic changes in brain functional connectivity patterns over time better reflects the nature of brain neural activity and provides richer information for exploring the fundamental properties of brain networks. Therefore, refining and developing dynamic brain functional network connectivity analysis methods to explore the dynamic characteristics of the brain through dynamic functional connectivity patterns between fMRI brain networks is of great significance. This not only effectively uncovers the intrinsic information hidden in fMRI signals but also better reveals the patterns of brain neural activity, promoting the important role of fMRI technology. However, current dynamic brain functional network connectivity analysis methods still face several key challenges in brain network construction and dynamic brain functional pattern extraction, requiring further research.

[0004] For example, a dynamic brain network analysis method and system based on optimal transmission disclosed in the prior art provides a method for dynamic brain functional network connectivity analysis. It uses PageRank to calculate the hub degree of each brain region in each sub-network, and then uses the hub degree of brain regions in adjacent brain networks as the source domain and target domain, respectively. It uses the structural brain network from DTI as the transmission cost, and uses optimal transmission to construct a hub degree propagation model to learn the topological evolution between adjacent brain networks and other operations. Summary of the Invention

[0005] Purpose of the invention: The present invention aims to provide an interpretable dynamic brain network analysis method based on functional and structural priors. This method is intended to analyze the impact of brain diseases on the spatiotemporal topological properties of the brain and improve the interpretability of disease diagnosis.

[0006] Technical Solution: An interpretable dynamic brain network analysis method based on functional and structural priors, comprising the steps of constructing an interpretable spatiotemporal tensor graph convolutional network for dynamic functional brain network analysis:

[0007] S1. Use the DPARSF toolkit to preprocess fMRI signals;

[0008] S2. Use non-overlapping sliding windows to divide fMRI signals and construct dynamic functional brain networks;

[0009] S3. Prior knowledge of function and structure is embedded into the dynamic functional brain network to obtain a hierarchical dynamic functional brain network representation based on brain region clustering, which is used to capture the spatiotemporal topological relationships between functional subnetworks.

[0010] For each window, the hierarchical dynamic functional brain network is divided into different subnetworks based on functional and structural priors, including the default mode network, frontoparietal network, salience network, attention network, somatomotor network, visual network, and auditory network.

[0011] Construct C dynamic functional subnets Where A c Let F represent the adjacency matrix of the c-th dynamic functional subnet. c Let N represent the node characteristics of the c-th dynamic functional subnet, T represent the number of non-overlapping sliding windows, and N represent the node characteristics of the c-th dynamic functional subnet. c G represents the number of nodes in the c-th dynamic functional subnet. c Let R represent the c-th dynamic functional subnet;

[0012] To capture the evolutionary patterns of brain connectivity, each dynamic functional subnetwork is stored as a tensor.

[0013] S4. Construct a tensor graph convolutional network with intra-graph and inter-graph propagation to collaboratively extract spatiotemporal topological features from a hierarchical dynamic functional brain network.

[0014] Taking the l-th layer of tensor graph convolution as an example, intra-graph propagation is performed first, followed by inter-graph propagation:

[0015]

[0016] in d represents the hidden feature tensor of the l-th layer ST2GCN. l It is the dimension of the hidden feature tensor, F c (0) These are the initial node features. f represents the output feature tensor after performing intragraph propagation. intra and f interThese represent intra-graph propagation and inter-graph propagation operations, respectively.

[0017] Intragraph Propagation: Intragraph propagation aggregates the feature information of adjacent nodes in each brain region of a dynamic functional subnet, given an adjacency tensor. and feature tensor After propagation within the diagram, f intra The output features of the t-th graph in layer l are updated as follows:

[0018]

[0019] Where σ represents the ReLU activation function. It is A c The graph after symmetry normalization This represents the weight matrix during the feature extraction process of the t-th graph in the l-th layer;

[0020] Inter-graph propagation: To aggregate dynamic functional subnets G under different windows c The timing characteristics are achieved by connecting the dynamic function subnet G. c We construct a series of virtual graphs using the same nodes in the graph; since all nodes and edges in the virtual graph are equally important, we set the connection weight between node pairs in the virtual graph to 1.

[0021] Consider G c Contains N c There are N brain regions, therefore there are N corresponding to them. c A virtual graph is generated, forming a new graph adjacency tensor. Inter-graph information propagation on a virtual graph inter This can be achieved through the following formula:

[0022]

[0023] in This represents the output features of inter-graph information propagation, and also the input features of the (l+1)th layer ST2GCN. σ represents the ReLU activation function. Represents the learnable parameters during the inter-graph propagation process;

[0024] S5. Functional subnet constraints are introduced to enhance the consistency of feature representations within subnets and the differences in feature representations between subnets, thereby guiding the learned features to better reflect the topological prior features of the brain network.

[0025] This step involves designing constraints for brain functional subnetworks, embedding them into spatiotemporal tensor graph convolutional networks, encouraging the aggregation of feature representations of brain regions within subnetworks, and promoting discriminative feature representations of brain regions between different subnetworks. This includes using cosine similarity to measure the similarity of spatiotemporal feature representations of brain regions within subnetworks and encouraging this similarity to increase as training progresses.

[0026] S6. The spatiotemporal features of the extracted different dynamic functional subnetworks are fused through an attention mechanism to obtain a whole-brain level feature representation;

[0027] S7. Use a multilayer perceptron for classification, map the extracted spatiotemporal features to subject labels, feed the obtained global brain spatiotemporal embedding into the multilayer perceptron to predict the category of each subject, and analyze the impact of brain diseases on brain spatiotemporal topological properties to improve the interpretability of disease diagnosis.

[0028] Furthermore, the preprocessing described in step S1 includes deleting and removing data from the original DICOM data, denoising, head motion correction, removing scalp structures, denoising after image segmentation, and marking regions of interest.

[0029] Step S2 first uses a non-overlapping sliding window of length S to divide the original fMRI signal into T sub-segments. Then, the correlation between brain regions in each sub-segment is calculated using the Pearson correlation coefficient to obtain the dynamic functional brain network. Subsequently, the dynamic functional brain network is sparsified by setting the connection strength less than the threshold α to 0.

[0030] Furthermore, step S5 of the above scheme uses cosine similarity to measure the similarity of spatiotemporal feature representations of brain regions within the subnet, and promotes the increase of this similarity as training progresses. The mathematical expression is as follows:

[0031]

[0032] in and Let represent the feature embeddings of the i-th and j-th brain regions located under the t-th window of the c-th dynamic functional subnetwork, respectively; · represents the inner product operation; and |||| is the Frobenius norm.

[0033] To encourage the diversity of features among dynamic functional subnetworks, Euclidean distance is used to measure the difference in feature representations between different subnetworks, and this difference is guided to increase continuously as training progresses.

[0034] Since different dynamic functional subnets contain different numbers of nodes, the extracted spatiotemporal feature dimensions also differ. Therefore, a fully connected layer is used to align the feature dimensions of multiple subnets: H c =FC(F c ), H c ∈R d This indicates the alignment feature embedding, where d is the embedding dimension.

[0035] In the above scheme, step S6 introduces a self-attention mechanism to weightedly combine the features of different functional subnetworks to form a whole-brain-level feature representation, specifically including:

[0036] First, the features of different subnets are packaged into a feature matrix. Then, matrix H is mapped using a linear mapping. C Perform different transformations:

[0037] Q = H C W q K = H C W k V=H C W v

[0038] Among them, W q W k W v The parameter matrices are used to generate queries, keys, and values, which are updated during model training via backpropagation of the network.

[0039] Then, the self-attention coefficients between functional sub-networks are captured by calculating the dot product correlation of Q and K, and mean pooling is used to obtain the integrated brain network embedding vector H:

[0040]

[0041] Softmax is used to normalize the attention weights.

[0042] Furthermore, based on functional and structural priors, an interpretable dynamic brain network analysis method can realize an interpretable spatiotemporal analysis model of dynamic brain networks:

[0043] Building blocks: hierarchical dynamic functional brain networks that integrate functional and structural priors;

[0044] Feature extraction module: Spatiotemporal tensor graph convolution integrates intra-graph propagation and inter-graph propagation into a unified convolutional framework for the collaborative extraction of spatiotemporal topological features in dynamic functional subnets;

[0045] Functional subnetwork constraint module: used to simulate the characteristics of strong coupling within subnetworks and loose coupling between subnetworks in the brain, enhance the consistency of feature representations within subnetworks, and emphasize the differences in feature representations between subnetworks;

[0046] Feature fusion module: Introduces a self-attention mechanism to form a whole-brain-level feature representation by weighted combination of features from different sub-networks;

[0047] The classification module maps the extracted whole-brain level feature representations to labels and trains the classifier using the cross-entropy loss function. This improves the model's brain disease diagnosis performance by narrowing the distance between the predicted and true labels.

[0048] Beneficial Effects: The present invention discloses an interpretable dynamic brain network analysis method based on functional and structural priors. By embedding prior knowledge of function and structure into a dynamic functional brain network, a hierarchical dynamic functional brain network representation based on brain region clustering is obtained. This effectively captures the spatiotemporal topological relationships between functional subnetworks. Furthermore, a tensor graph convolutional network with intra-graph and inter-graph propagation is designed to collaboratively extract spatiotemporal topological features from the hierarchical dynamic functional brain network. This also includes introducing functional subnetwork constraints to enhance the consistency of feature representations within subnetworks and the differences in feature representations between subnetworks. This guides the learned features to better reflect the topological prior features of the brain network. Accordingly, by narrowing the distance between predicted and true labels, the model's performance in diagnosing brain diseases is improved, enabling the analysis of the impact of brain diseases on the spatiotemporal topological properties of the brain and improving the interpretability of disease diagnosis. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the workflow of the method described in this invention;

[0050] Figure 2 This is an overall framework diagram of the method described in this invention;

[0051] Figure 3 This is a schematic diagram of the functional subnetwork constraints in this invention;

[0052] Figure 4 It is a comparison of the spatiotemporal variability of dynamic brain networks on different classification tasks. Detailed Implementation

[0053] Dynamic functional brain networks (DNFs) can capture topological changes in connectivity patterns during brain activity and have attracted increasing attention for detecting brain dysfunction. However, most current DNF analysis methods rely on data-driven modeling, neglecting key prior knowledge of brain structure and function, resulting in weak model interpretability. Furthermore, due to the complex spatiotemporal coupling of features in DNFs, effectively extracting dynamic topological features remains a challenging problem. To address these issues, this invention provides an interpretable DNF analysis method based on functional and structural priors, which can be referenced. Figure 1 As shown, the implementation steps of this method are as follows:

[0054] S1. Use the DPARSF toolkit to preprocess fMRI signals;

[0055] First, functional magnetic resonance imaging (fMRI) data is acquired and preprocessed to obtain fMRI signals.

[0056] fMRI data preprocessing: All fMRI data were preprocessed using SPM8, implemented in the DPARSF toolbox in MATLAB.

[0057] After obtaining the original DICOM data, since the images at the first few time points have a lot of noise, this invention first removes the first 10 time points.

[0058] Next, the Realign operation is used to correct head movement and exclude samples with significant head movement.

[0059] Using the Bet operation in the toolbox to remove scalp structures can reduce the influence of irrelevant structures during registration, thereby improving registration accuracy.

[0060] Match the T1 image to the functional image.

[0061] Using the Segment and New segment+DARTEL operations in the toolbox, the image is segmented into gray matter, white matter, and cerebrospinal fluid.

[0062] Nuisance covariates regression is used to remove noise signals such as white matter and cerebrospinal fluid.

[0063] After preprocessing, the fMRI was divided into 90 regions of interest at 240 time points using an automated anatomical atlas.

[0064] S2. Use non-overlapping sliding windows to divide fMRI signals and construct dynamic functional brain networks;

[0065] Based on the preprocessed fMR data obtained in step S1, the original fMRI signal is first divided into T sub-segments using a non-overlapping sliding window of length S. Then, the correlation between brain regions in each sub-segment is calculated using the Pearson correlation coefficient to obtain the dynamic functional brain network. The dynamic functional brain network is then sparsified by setting connections with a strength less than a threshold α to 0.

[0066] S3. Prior knowledge of function and structure is embedded into the dynamic functional brain network to obtain a hierarchical dynamic functional brain network representation based on brain region clustering, which effectively captures the spatiotemporal topological relationship between functional subnetworks.

[0067] For each functional brain network within a window, we divide brain regions and connections with similar functions into different subnetworks based on functional and structural priors. These include the Default Mode Network (DMN), Frontoparietal Network (FPN), Salience Network (SAN), Attention Network (ATN), Sensorimotor Network (SMN), Visual Network (VN), Auditory Network (AN), and other networks (ON) composed of brain regions and connections not included in the above subnetworks. Therefore, we obtain C dynamic functional subnetworks. Where A cLet F represent the adjacency matrix of the c-th dynamic functional subnet, T represent the number of non-overlapping sliding windows, and F represent the adjacency matrix of the c-th dynamic functional subnet. c N represents the node characteristics of the c-th dynamic functional subnet. c This represents the number of nodes contained in the c-th dynamic functional subnetwork. To capture the evolutionary patterns of brain connectivity, each dynamic functional subnetwork is stored as a tensor.

[0068] S4. A tensor graph convolutional network with intra-graph and inter-graph propagation was designed to collaboratively extract spatiotemporal topological features from a hierarchical dynamic functional brain network. Specifically, intra-graph propagation is used to extract spatial structure information from the dynamic functional subnet, while inter-graph propagation extracts dynamic temporal information from the dynamic functional subnet by constructing a series of virtual graphs.

[0069] To capture the dynamic evolution of the spatial topology of brain networks, this invention designs a Spatio-Temporal Tensor Graph Convolutional Network (ST2GCN) with intra-graph and inter-graph propagation. This network can directly perform convolutional learning on dynamic tensors to collaboratively capture spatio-temporal correlations. Taking the l-th layer of the tensor graph convolution as an example, intra-graph propagation is performed first, followed by inter-graph propagation:

[0070]

[0071] in d represents the hidden feature tensor of the l-th layer ST2GCN. l It is the dimension of the hidden feature tensor, F c (0) These are the initial node features. f represents the output feature tensor after performing intragraph propagation. intra and f inter These represent intra-graph propagation and inter-graph propagation operations, respectively.

[0072] Intragraph propagation: This method aggregates feature information of neighboring nodes in each brain region of a dynamic functional subnet. Given an adjacency tensor... and feature tensor After propagation within the diagram, f intra Update the output features of the t-th graph in layer l as follows:

[0073]

[0074] Where σ represents the ReLU activation function. It is A c The graph after symmetry normalization Let represent the weight matrix in the feature extraction process of the t-th graph in the l-th layer.

[0075] Inter-graph propagation: To aggregate dynamic functional subnets G under different windows c Due to the timing characteristics, this invention connects the dynamic function subnet G c We construct a series of virtual graphs using identical nodes from the original graph. Since all nodes and edges in the virtual graph are equally important, the connection weight between pairs of nodes in the virtual graph is set to 1. c Contains N c From each brain region, we can obtain N. c A virtual graph is generated, forming a new graph adjacency tensor. Inter-graph information propagation on a virtual graph inter This can be achieved through the following formula:

[0076]

[0077] in It serves as both the output feature for inter-graph information propagation and the input feature for the l+1 layer ST2GCN. σ represents the ReLU activation function. This represents the learnable parameters during the inter-graph propagation process.

[0078] S5. Functional subnet constraints are introduced to enhance the consistency of feature representations within subnets and the differences in feature representations between subnets, thereby guiding the learned features to better reflect the topological prior features of the brain network.

[0079] The brain is a modular system composed of multiple subnetworks, each consisting of numerous brain regions and connections with specific functions. Brain regions within the same subnetwork are typically tightly connected, while those in different subnetworks are mostly sparsely connected. Based on this prior knowledge, a brain functional subnetwork constraint was designed and embedded into a spatiotemporal tensor graph convolutional network. This encourages the aggregation of feature representations of brain regions within a subnetwork while simultaneously promoting discriminative feature representations of brain regions across different subnetworks. Cosine similarity is used to measure the similarity of spatiotemporal feature representations of brain regions within a subnetwork, and this similarity is encouraged to increase with training.

[0080]

[0081] in and Let H represent the feature embeddings of the i-th and j-th brain regions located in the t-th window of the c-th dynamic functional subnetwork, respectively. · represents the inner product operation, and |||| is the Frobenius norm. To encourage feature diversity among dynamic functional subnetworks, Euclidean distance is used to measure the difference in feature representations between different subnetworks, and this difference is guided to increase continuously as training progresses. Since different dynamic functional subnetworks contain different numbers of nodes, the extracted spatiotemporal feature dimensions also differ. Therefore, a fully connected layer is first used to align the feature dimensions of multiple subnetworks: H c=FC(F c ), H c ∈R d This indicates the alignment feature embedding, where d is the embedding dimension.

[0082] S6. The spatiotemporal features of the extracted different dynamic functional subnets are fused through an attention mechanism;

[0083] Each functional subnetwork contributes differently to the brain network representation. Therefore, we introduce a self-attention mechanism to weightedly combine the features of different functional subnetworks to form a whole-brain-level feature representation. Specifically, we first package the features of different subnetworks into a feature matrix. Then we apply a linear mapping to matrix H. C Perform different transformations:

[0084] Q = H C W q K = H C W k V=H C W v

[0085] Among them, W q W k W v The parameter matrices are used to generate queries, keys, and values, which are updated via backpropagation during model training. Then, the self-attention coefficients between functional sub-networks are captured by calculating the dot product correlation of Q and K, and mean pooling is used to obtain the comprehensive brain network embedding vector H.

[0086]

[0087] Softmax is used to normalize the attention weights.

[0088] S7. The obtained global brain spatiotemporal embedding is fed into a multilayer perceptron to predict the category of each subject, and the impact of brain diseases on brain spatiotemporal topological properties is analyzed to improve the interpretability of disease diagnosis.

[0089] Classification is performed using a multilayer perceptron, mapping the extracted spatiotemporal features to subject labels, as shown in the following expression:

[0090] Output = MLP(H).

[0091] Further integration Figure 2-3As shown, the method described in this invention can realize the construction of a spatiotemporal analysis model of interpretable dynamic brain networks. The model includes a construction module for a hierarchical dynamic functional brain network that integrates functional and structural priors, a tensor graph convolution feature extraction module with intra-graph and inter-graph propagation, a functional sub-network constraint module, a feature fusion module, and a classification module.

[0092] This method achieves the overall framework structure as follows: Figure 2 As shown, it consists of a hierarchical dynamic functional brain network that integrates functional and structural priors, a spatiotemporal tensor graph convolutional network, functional subnetwork constraints, an attention mechanism, and a multilayer perceptron. The hierarchical dynamic brain network integrating functional and structural priors is used to capture the topological dependencies between subnetworks. The spatiotemporal tensor graph convolution integrates intra-graph and inter-graph propagation into a unified convolutional framework for the collaborative extraction of spatiotemporal topological features in the dynamic functional subnetworks. A detailed description of the functional subnetwork constraint module is as follows... Figure 3 As shown, this module is designed to simulate the strong coupling within subnetworks and the loose coupling between subnetworks in the brain, thereby enhancing the consistency of feature representations within subnetworks and emphasizing the differences in feature representations between subnetworks. Since each functional subnetwork contributes differently to the brain network representation, this invention introduces a self-attention mechanism to form a whole-brain-level feature representation by weighted combination of features from different subnetworks. The classifier's role is to map the extracted whole-brain-level feature representations to labels, and the classifier is trained using a cross-entropy loss function to improve the model's brain disease diagnostic performance by narrowing the distance between the predicted labels and the true labels.

[0093] This experiment uses two tensor-spatiotemporal graph convolutional layers. This invention utilizes a single-head self-attention mechanism to fuse features from different sub-networks to obtain a whole-brain feature representation. In the classifier, the two fully connected layers have 32 and 2 neurons respectively, and softmax is used as the activation function in the last layer.

[0094] The following description, in conjunction with specific examples, provides further details.

[0095] (I) Experimental Analysis

[0096] (1) Comparative Experiment: To verify the effectiveness of the proposed method in the diagnosis of brain diseases, extensive comparative analyses were conducted on collected epilepsy datasets and the publicly available ADNI dataset. Classification performance was measured using three evaluation metrics: accuracy (ACC), F1 score (F1), and receiver operating characteristic (AUC). The experimental results are shown in Tables 1 and 2, respectively. As can be seen from Tables 1 and 2, the method described in this invention achieved excellent diagnostic performance across multiple evaluation metrics, which fully demonstrates the advanced nature of this method in the diagnosis of brain diseases.

[0097] (2) Ablation Experiments: To verify the importance of the three main modules in the proposed method—spatiotemporal tensor graph convolution, intra-subnetwork constraints, and inter-subnetwork constraints—for the diagnosis of brain diseases, multiple ablation experiments were conducted. The experimental results are shown in Table 3. In Table 3, the baseline indicates that only multilayer perceptron was used for classification, and √ indicates that the corresponding module was used. The ablation experiment results show that all three modules can improve the performance of disease diagnosis compared to the baseline. In addition, the classification effect is better when two modules are used together than when only one module is used, which shows that the proposed modules are effective and that the modules promote each other. Spatiotemporal tensor graph convolution overcomes the shortcomings of traditional spatiotemporal models in handling dynamic graphs, and extracts spatiotemporal features collaboratively through intra-graph propagation and inter-graph propagation. Intra-subnetwork constraints and inter-subnetwork constraints integrate prior knowledge of function and structure, enabling the extracted spatiotemporal features to better capture the multi-level topological changes of dynamic functional brain networks, resulting in more discriminative features.

[0098] (II) Interpretability Analysis

[0099] (1) Importance of each functional subnetwork: The importance of each subnetwork to the diagnostic task was obtained by summing the attention mechanisms in step 6, and the experimental results are shown in Table 4. The table shows that different functional subnetworks have different diagnostic effects on brain diseases. Specifically, in the classification task between the normal control group and patients with frontal lobe epilepsy (NC vs. FLE), and in the classification task between the normal control group and patients with temporal lobe epilepsy (NC vs. TLE), the two most important functional subnetworks were the DMN and ATN. This is because frontal lobe epilepsy (FLE) originates from abnormal electrical neural activity in the frontal lobe, and temporal lobe epilepsy (TLE) originates from temporal lobe dysfunction, and these two areas are mainly located in the DMN and ATN. In the classification task between the normal control group and patients with subjective memory anxiety (NC vs. SMC), the two most important functional subnetworks were the ATN and FPN. The ATN and FPN are responsible for allocating attention and regulating cognitive resources to support complex cognitive tasks, respectively. Due to the occurrence of subjective memory anxiety (SMC), the brain may need to allocate more cognitive resources to compensate for the deficit in memory function. In addition, memory problems can cause worry and anxiety in patients, making it difficult for individuals to concentrate. Therefore, SMC may originate in part from abnormalities in the ATN and FPN. In the normal control group versus early mild cognitive impairment (EMCI) classification task, the two most important functional subnetworks are the ATN and VN. Early mild cognitive impairment (EMCI) can lead to cognitive decline, including memory, attention, and visual perception. The ATN controls memory and attention, while the VN determines an individual's visual and spatial perception abilities. Therefore, the proposed interpretable model can autonomously identify important subnetworks in different diagnostic tasks.

[0100] (2) Spatiotemporal variability: This invention calculates the variance of the total connectivity strength of the brain network across all sliding windows, using the variance as the spatiotemporal variability rate to explore the impact of brain diseases on the spatiotemporal characteristics of dynamic brain networks. The experimental results are as follows: Figure 1 As shown. From Figure 1 The study revealed that patients with spatiotemporal variability were significantly higher than those with neuronal non-contrast (NC), indicating that brain diseases can lead to instability in brain connectivity. Epilepsy is caused by excessive electrical activity in neurons due to a disruption of neuronal inhibitory mechanisms. This abnormality can spread to other areas of the brain network, leading to either interrupted or enhanced connections, making the changes in brain connectivity more drastic. Spinal muscular dystrophy (SMC) and endocrine disorders (EMCI) are cognitive impairments in which patients' cognitive abilities, such as memory and attention, gradually deteriorate, and the brain undergoes neurodegenerative changes. Furthermore, the brain compensates for these functional impairments by remodeling its networks, thus making the brain network even more unstable.

[0101] Table 1. Classification performance (mean / standard deviation) on the epilepsy dataset (%).

[0102]

[0103] Table 2. Classification performance (mean / standard deviation) on the ADNI dataset (%).

[0104]

[0105]

[0106] Table 3. Ablation experimental results (%) on the epilepsy and ANDI datasets

[0107]

[0108]

[0109] In Table 3: 1 represents the spatiotemporal tensor graph convolution module; 2 represents the intra-subnet constraint module; 3 represents the inter-subnet constraint module.

[0110] Table 4. Importance of Functional Subnetworks in Different Diagnostic Tasks

[0111]

[0112] In Table 4, we have bolded the two most important subnets for different tasks.

Claims

1. A method for interpretable dynamic brain network analysis based on functional and structural priors, the method comprising constructing an interpretable spatiotemporal tensor graph convolutional network for dynamic functional brain network analysis, characterized in that, Includes the following steps: S1. Use the DPARSF toolkit to preprocess fMRI signals; S2. Use non-overlapping sliding windows to divide fMRI signals and construct dynamic functional brain networks; S3. Prior knowledge of function and structure is embedded into the dynamic functional brain network to obtain a hierarchical dynamic functional brain network representation based on brain region clustering, which is used to capture the spatiotemporal topological relationships between functional subnetworks. For each window, the hierarchical dynamic functional brain network is divided into different subnetworks based on functional and structural priors, including the default mode network, frontoparietal network, salience network, attention network, somatomotor network, visual network, and auditory network. Construct C dynamic functional subnets Where A c Let F represent the adjacency matrix of the c-th dynamic functional subnet. c Let N represent the node characteristics of the c-th dynamic functional subnet, T represent the number of non-overlapping sliding windows, and N represent the node characteristics of the c-th dynamic functional subnet. c This represents the number of nodes contained in the c-th dynamic functional subnet; To capture the evolutionary patterns of brain connectivity, each dynamic functional subnetwork is stored as a tensor. S4. Construct a tensor graph convolutional network with intra-graph and inter-graph propagation to collaboratively extract spatiotemporal topological features from a hierarchical dynamic functional brain network. Taking the l-th layer of tensor graph convolution as an example, intra-graph propagation is performed first, followed by inter-graph propagation: in d represents the hidden feature tensor of the l-th layer ST2GCN. l It is the dimension of the hidden feature tensor, F c(0) These are the initial node features. f represents the output feature tensor after performing intragraph propagation. intra and f inter These represent intra-graph propagation and inter-graph propagation operations, respectively. Intragraph Propagation: Intragraph propagation aggregates the feature information of adjacent nodes in each brain region of a dynamic functional subnet, given an adjacency tensor. and feature tensor After propagation within the diagram, f intra The output features of the t-th graph in layer l are updated as follows: Where σ represents the ReLU activation function. It is A c The graph after symmetry normalization This represents the weight matrix during the feature extraction process of the t-th graph in the l-th layer; Inter-graph propagation: To aggregate dynamic functional subnets G under different windows c The timing characteristics are achieved by connecting the dynamic function subnet G. c We construct a series of virtual graphs using the same nodes in the graph; since all nodes and edges in the virtual graph are equally important, we set the connection weight between node pairs in the virtual graph to 1. Consider G c Contains N c There are N brain regions, therefore there are N corresponding to them. c A virtual graph is generated, forming a new graph adjacency tensor. Inter-graph information propagation on a virtual graph inter This can be achieved through the following formula: in This represents the output features of inter-graph information propagation, and also the input features of the (l+1)th layer ST2GCN. σ represents the ReLU activation function. Represents the learnable parameters during the inter-graph propagation process; S5. Functional subnet constraints are introduced to enhance the consistency of feature representations within subnets and the differences in feature representations between subnets, thereby guiding the learned features to better reflect the topological prior features of the brain network. This step involves designing constraints for brain functional subnetworks, embedding them into spatiotemporal tensor graph convolutional networks, encouraging the aggregation of feature representations of brain regions within subnetworks, and promoting discriminative feature representations of brain regions between different subnetworks. This includes using cosine similarity to measure the similarity of spatiotemporal feature representations of brain regions within subnetworks and encouraging this similarity to increase as training progresses. S6. The spatiotemporal features of the extracted different dynamic functional subnetworks are fused through an attention mechanism to obtain a whole-brain level feature representation; S7. Use a multilayer perceptron for classification, map the extracted spatiotemporal features to subject labels, feed the obtained global brain spatiotemporal embedding into the multilayer perceptron to predict the category of each subject, and analyze the impact of brain diseases on brain spatiotemporal topological properties to improve the interpretability of disease diagnosis.

2. The interpretable dynamic brain network analysis method based on functional and structural priors according to claim 1, characterized in that, The preprocessing described in step S1 includes deletion and removal of the original DICOM data, denoising, head motion correction, removal of scalp structures, denoising after image segmentation, and marking of regions of interest.

3. The interpretable dynamic brain network analysis method based on functional and structural priors according to claim 1, characterized in that, Step S2 first uses a non-overlapping sliding window of length S to divide the original fMRI signal into T sub-segments. Then, the correlation between brain regions in each sub-segment is calculated using the Pearson correlation coefficient to obtain the dynamic functional brain network. Subsequently, the dynamic functional brain network is sparsified by setting the connection strength less than the threshold α to 0.

4. The interpretable dynamic brain network analysis method based on functional and structural priors according to claim 1, characterized in that, Step S5 uses cosine similarity to measure the similarity of spatiotemporal feature representations of brain regions within a subnet, and encourages this similarity to increase as training progresses. The mathematical expression is as follows: in and Let represent the feature embeddings of the i-th and j-th brain regions located under the t-th window of the c-th dynamic functional subnetwork, respectively; · represents the inner product operation; and ∥∥ is the Frobenius norm. To encourage the diversity of features among dynamic functional subnetworks, Euclidean distance is used to measure the difference in feature representations between different subnetworks, and this difference is guided to increase continuously as training progresses. Since different dynamic functional subnets contain different numbers of nodes, the extracted spatiotemporal feature dimensions also differ. Therefore, a fully connected layer is used to align the feature dimensions of multiple subnets: H c =FC(F c ), H c ∈R d This indicates the alignment feature embedding, where d is the embedding dimension.

5. The interpretable dynamic brain network analysis method based on functional and structural priors according to claim 1, characterized in that, Step S6 introduces a self-attention mechanism to weightedly combine features from different functional subnetworks, forming a whole-brain-level feature representation, specifically including: First, the features of different subnets are packaged into a feature matrix. Then, matrix H is mapped using a linear mapping. C Perform different transformations: Q=H C W q ,K=H C W k ,V=H C W v Among them, W q W k W v The parameter matrices are used to generate queries, keys, and values, which are updated during model training via backpropagation of the network. Then, the self-attention coefficients between functional sub-networks are captured by calculating the dot product correlation of Q and K, and mean pooling is used to obtain the integrated brain network embedding vector H: Softmax is used to normalize the attention weights.

6. The interpretable dynamic brain network analysis method based on functional and structural priors according to claim 1, characterized in that, This method enables the realization of interpretable spatiotemporal analysis models of dynamic brain networks: Building blocks: hierarchical dynamic functional brain networks that integrate functional and structural priors; Feature extraction module: Spatiotemporal tensor graph convolution integrates intra-graph propagation and inter-graph propagation into a unified convolutional framework for the collaborative extraction of spatiotemporal topological features in dynamic functional subnets; Functional subnetwork constraint module: used to simulate the characteristics of strong coupling within subnetworks and loose coupling between subnetworks in the brain, enhance the consistency of feature representations within subnetworks, and emphasize the differences in feature representations between subnetworks; Feature fusion module: Introduces a self-attention mechanism to form a whole-brain-level feature representation by weighted combination of features from different sub-networks; The classification module maps the extracted whole-brain level feature representations to labels and trains the classifier using the cross-entropy loss function. This improves the model's brain disease diagnosis performance by narrowing the distance between the predicted and true labels.

Images