A brain network dynamic analysis method and system based on sub-network alignment

Abstract

The application belongs to the technical field of brain network analysis, and aims to solve the problem of inaccurate brain network analysis in the prior art, and provides a brain network dynamic analysis method and system based on sub-network alignment, which reorders each time window to arrange nodes belonging to the same functional sub-network in sequence, aligns the functional sub-network graph in each time window to capture the spatial interaction mode of the nodes by a spatial mixing expert, models the independent time sequence evolution of the sequence token of each functional sub-network in the whole time context by a time mixing expert, captures the cooperative or competitive dynamics across the functional sub-networks by a multi-head self-attention mechanism, fuses the global topology information of the original functional connection matrix for the updated time sequence token of each functional sub-network, introduces a learnable brain state prototype to obtain a subject state-level compact representation by aggregation, and finally obtains the brain network analysis result. The application improves the accuracy of brain disease classification and the interpretability of neuroscience.

Description

Technical Field

[0001] This invention belongs to the technical field of brain network analysis, and in particular relates to a method and system for dynamic analysis of brain networks based on subnetwork alignment. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Resting-state functional magnetic resonance imaging (fMRI) allows for non-invasive observation of spontaneous brain activity by measuring low-frequency fluctuations in blood oxygenation-dependent signals. Dynamic functional connectivity further characterizes the evolution of functional interactions between brain regions over time. Numerous studies have shown that the dynamic properties of dynamic functional connectivity are closely related to cognitive behavior and mental illness. Therefore, constructing sensitive and interpretable biomarkers based on dynamic functional connectivity is of significant value for the early diagnosis and understanding of pathological mechanisms of brain diseases.

[0004] However, existing dynamic functional connectivity modeling methods suffer from the following key drawbacks: First, the brain is not a single, synchronous system, but a heterogeneous system composed of multiple functional subnetworks. These subnetworks have different time scales and state transition rhythms: default mode networks change slowly, while attention networks can switch rapidly. Most existing methods use the whole-brain connectivity matrix within a sliding window as the basic unit, employing a global temporal model to uniformly process information from all subnetworks. This forces the heterogeneous dynamics of different subnetworks to be mixed in the same latent space, causing phase mixing problems. This makes it difficult for the model to simultaneously capture fast transients and slow, large-scale fluctuations. Second, although graph neural networks have been used for brain network analysis, they typically treat all brain region nodes equally. Highly dense connections within a subnetwork and sparse connections between subnetworks have fundamentally different functional meanings, but the shared isomorphic message passing mechanism makes it difficult to distinguish these patterns. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a method and system for dynamic analysis of brain networks based on subnetwork alignment. Through spatial propagation of subnetwork alignment, subnetwork-specific temporal modeling, and feedback mechanisms, it improves the accuracy of brain disease classification and neuroscientific interpretability.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for dynamic analysis of brain networks based on sub-network alignment, comprising: For the resting-state functional magnetic resonance imaging time series of the subjects, a dynamic functional connectivity sequence was constructed using a sliding window. Based on the pre-divided functional subnetworks, the functional connection matrix and node features of each time window are reordered so that nodes belonging to the same functional subnetwork are arranged continuously, resulting in a functional subnetwork alignment map. For the functional subnetwork alignment graph within each time window, neighborhood aggregation and nonlinear transformation are performed by a spatial hybrid expert to capture the spatial interaction patterns of nodes and obtain a node-level spatial representation. The spatial representations at the node level within each functional subnetwork are averaged and pooled to obtain temporal tokens at the functional subnetwork level. The contextual representations of the sequence tokens of each functional subnetwork over the entire time are then modeled by a temporal hybrid expert to obtain the temporal representations of the functional subnetworks. Within each time window, a multi-head self-attention mechanism is used to perform cross-sub-network interactive update processing on the time-series tokens of each functional sub-network, capturing the cooperative or competitive dynamics of cross-functional sub-networks. The timing tokens updated for each functional subnetwork are fused with the global topology information of the original functional connection matrix to obtain the enhanced timing tokens for the functional subnetworks. By introducing learnable brain state prototypes, calculating the probability that each functional subnetwork augmented temporal token belongs to each brain state, and obtaining compact representations of the subject's state level through aggregation, the brain network analysis results are obtained.

[0007] Secondly, the present invention provides a brain network dynamic analysis system based on subnetwork alignment, comprising: The data processing module is configured to: construct a dynamic functional connectivity sequence using a sliding window for the resting-state functional magnetic resonance imaging time series of the subjects; The subnetwork alignment module is configured to: reorder the functional connection matrix and node features of each time window according to the pre-divided functional subnetworks, so that the nodes belonging to the same functional subnetwork are arranged continuously, and thus obtain the functional subnetwork alignment map. The spatial hybrid expert module is configured to: perform neighborhood aggregation and nonlinear transformation on the functional subnetwork alignment graph within each time window, capture the spatial interaction patterns of nodes, and obtain a node-level spatial representation; The temporal hybrid expert module is configured to: perform mean pooling on the spatial representation at the node level within each functional subnetwork to obtain temporal tokens at the functional subnetwork level; and model the contextual representation of the sequence tokens of each functional subnetwork over the entire time period through the temporal hybrid expert to obtain the temporal representation of the functional subnetwork. The cross-subnetwork interaction module is configured to: within each time window, use a multi-head self-attention mechanism to perform cross-subnetwork interaction update processing on the time-series tokens of each functional subnetwork, and capture the collaborative or competitive dynamics of cross-functional subnetworks. The global connectivity fusion module is configured to: merge the timing tokens updated by each functional subnetwork with the global topology information of the original functional connectivity matrix to obtain enhanced timing tokens for the functional subnetworks; The analysis module is configured to: introduce learnable brain state prototypes, calculate the probability that each functional subnetwork augmented temporal token belongs to each brain state, obtain compact representations of the subject's state level through aggregation, and then obtain brain network analysis results.

[0008] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0009] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0010] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0011] The above one or more technical solutions have the following beneficial effects: In this invention, by reordering the functional connectivity matrix and node features of each time window, nodes belonging to the same functional subnetwork are arranged continuously. Then, spatial and temporal mixing experts are used to learn different interaction patterns within and between functional subnetworks, avoiding information ambiguity. By modeling each functional subnetwork independently, the different evolutionary rhythms of different functional subnetworks can be captured, overcoming the "phase mixing" problem of global temporal models. This invention improves the accuracy of brain disease classification and neuroscience interpretability through spatial propagation of subnetwork alignment and subnetwork-specific temporal modeling.

[0012] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0013] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0014] Figure 1 This is an overall flowchart of the brain network dynamic analysis method based on sub-network alignment in Embodiment 1 of the present invention. Detailed Implementation

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0016] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0017] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0018] Example 1 like Figure 1 As shown, this embodiment discloses a method for dynamic analysis of brain networks based on sub-network alignment, including: For the resting-state functional magnetic resonance imaging time series of the subjects, a dynamic functional connectivity sequence was constructed using a sliding window. Based on the pre-divided functional subnetworks, the functional connection matrix and node features of each time window are reordered so that nodes belonging to the same functional subnetwork are arranged continuously, resulting in a functional subnetwork alignment map. For the functional subnetwork alignment graph within each time window, neighborhood aggregation and nonlinear transformation are performed by a spatial hybrid expert to capture the spatial interaction patterns of nodes and obtain a node-level spatial representation. The spatial representations at the node level within each functional subnetwork are averaged and pooled to obtain temporal tokens at the functional subnetwork level. The contextual representations of the sequence tokens of each functional subnetwork over the entire time are then modeled by a temporal hybrid expert to obtain the temporal representations of the functional subnetworks. Within each time window, a multi-head self-attention mechanism is used to perform cross-sub-network interactive update processing on the time-series tokens of each functional sub-network, capturing the cooperative or competitive dynamics of cross-functional sub-networks. The timing tokens updated for each functional subnetwork are fused with the global topology information of the original functional connection matrix to obtain the enhanced timing tokens for the functional subnetworks. By introducing learnable brain state prototypes, calculating the probability that each functional subnetwork augmented temporal token belongs to each brain state, and obtaining compact representations of the subject's state level through aggregation, the brain network analysis results are obtained.

[0019] In this embodiment, by reordering each time window, nodes belonging to the same functional subnetwork are arranged consecutively. Then, graph propagation experts are used to learn different interaction patterns within and between functional subnetworks, avoiding information ambiguity. Independent modeling of each functional subnetwork captures the distinct evolutionary rhythms of different functional subnetworks, overcoming the "phase mixing" problem of global temporal models. The temporal state at the functional subnetwork level is injected back into the spatial representation at the node level, making spatial propagation modulated by the temporal context, forming a more compact spatiotemporal representation. This invention improves the accuracy of brain disease classification and neuroscientific interpretability through subnetwork-aligned spatial propagation, subnetwork-specific temporal modeling, and feedback mechanisms.

[0020] The following is a detailed description of the brain network dynamic analysis method based on sub-network alignment proposed in this embodiment: S1. Obtain the resting-state functional magnetic resonance imaging time series of the subject, and construct a dynamic functional connectivity map sequence using a sliding window. Each diagram Includes a set of brain region nodes and functional connection matrix .

[0021] Given the time series of functional magnetic resonance imaging of the subject ,in, The number of brain regions, such as when using the CC200 atlas, ; Let be the time point number. A sliding window method is used, with a window width of . Step size is A total of The system operates in two windows. Within each window, the Pearson correlation coefficients of functional magnetic resonance imaging time series between different brain regions (ROIs) are calculated to obtain the functional connectivity matrix. Then the dynamic graph sequence is To eliminate noise, Fisher's Z-transform is usually also performed.

[0022] S2. Based on the pre-divided functional sub-networks, the functional connection matrix and node features of each time window are reordered so that nodes belonging to the same functional sub-network are arranged continuously, thus obtaining the functional sub-network alignment map.

[0023] Based on prior brain maps (such as Yeo 7 / 17 network, CC200 functional subnetwork partitioning), The brain regions are divided into Let there be three disjoint functional subnetworks, and let the functional subnetwork label mapping be denoted as . .

[0024] The functional subnetworks are divided into eight large-scale functional subnetworks: cerebellum and subcortical structures, visual network, somatic motor network, dorsal attention network, ventral attention network, limbic network, frontoparietal network, and default mode network.

[0025] All N Each brain region (node) is sorted according to its predefined functional subnetwork, and a permutation matrix is ​​constructed. This ensures that nodes belonging to the same subnetwork are arranged consecutively after sorting. In a specific implementation, such as using the CC200 graph and dividing it into 8 functional subnetworks like the default mode network and the attention network, the goal of reordering is to arrange these subnetworks in blocks within the matrix.

[0026] The functional adjacency matrix and node features for each time window are reordered:

[0027]

[0028] in, This is the initial node feature dimension, which can be 1, i.e., the BOLD signal value; For the number of brain regions, This represents the node characteristics of all nodes contained in all functional subnetworks within the time window t. This is the functional connection matrix.

[0029] Reordered functional adjacency matrix It exhibits a block-diagonal dominance structure, with each block corresponding to an internal connection within a subnetwork. To reduce computational complexity, the connection strength is retained for each time window. Set the edges with a percentage of 0 to zero, and set the other edges to 0 to obtain a sparse adjacency matrix. In the experiment Take 30%.

[0030] S3. For the functional subnetwork alignment graph within each time window, neighborhood aggregation and nonlinear transformation are performed by a spatial hybridization expert to capture the spatial interaction patterns of nodes and obtain the spatial representation at the node level.

[0031] The functional subnetwork alignment graph is a graph structure after node reordering, and its functional adjacency matrix is... Through permutation matrix P For the original functional connection matrix The nodes of the same functional subnetwork are reordered to form a block structure in the matrix.

[0032] In this embodiment, S3 specifically includes: S31, Let there be a total A space expert, in this embodiment of the experiment Each space expert receives the same input, but the difference lies in the fact that each space expert k corresponds to a parameter-independent lightweight multilayer perceptron. That is, each space expert uses different learnable parameters that are not shared with each other. Therefore, different space experts perform different nonlinear transformations on the same input, thereby capturing diverse spatial interaction patterns.

[0033] Space experts Lightweight multilayer perceptron Its input is the fusion of the current node's features and the normalized adjacency matrix, the th... The output of an expert for:

[0034] in, To hide dimensions, in the experiment ; This is the sparse adjacency matrix after normalization and sparsification of the functional subnetworks; The node features are reordered from the functional subnetwork; the subscript b represents the subject's index. This represents the number of brain regions.

[0035] This employs residual connections, allowing space experts the flexibility to choose to transmit only the original features or aggregated information.

[0036] S32. In order to assign a suitable spatial expert to each node, a gated network is introduced.

[0037] First, the node features of each node are summarized to obtain a node-level description vector. An effective summarization function is... A more refined approach is to use average pooling of all node features, or to compress high-dimensional features using a small MLP, and then generate the spatial weight matrix of each node with respect to each spatial expert through linear mapping and Softmax. :

[0038] in, C′ represents the result after the summary function. Each node in the output describes the dimension of the vector; K represents the number of spatial experts, and in the experiment, K=8; The node features of subject b after functional subnetwork reordering.

[0039] Weight matrix of gating network ∈ Map the C′-dimensional description vector of each node to a K-dimensional expert logic value, and then... The activation function yields the spatial expert weights for each node.

[0040] S33. To avoid increasing computational load and homogenizing spatial experts due to activating multiple spatial experts on all nodes, a hard Top-1 routing is adopted: on each node, only the output of the spatial expert with the highest spatial weight is retained, and the outputs of the other experts are set to 0. This forces each spatial expert to specialize in a certain type of node or interaction mode. For example, one spatial expert may be responsible for message passing within a subnetwork, while another spatial expert focuses on long-range connections across subnetworks.

[0041] S34. Weighted fusion yields the final node-level spatial representation. :

[0042] in, This indicates element-wise multiplication; The matrix represents the spatial weights of spatial expert k; the subscript b indicates the index of the subject; and K is the number of spatial experts. To hide dimensions; This represents the number of brain regions.

[0043] S4. Average pooling is performed on the spatial representations at the node level within each functional subnetwork to obtain time-series tokens at the functional subnetwork level.

[0044] To rise to the functional subnetwork level, for each functional subnetwork The node features within the network are averaged and pooled to obtain time-series tokens at the functional subnetwork level. :

[0045] in, For sub-networks The set of nodes, where i is the node index of the subnetwork.

[0046] For each subject Each functional subnetwork Each time window Through spatial representation of all nodes within the subnetwork Perform mean pooling to obtain a dimensional vector Stacking up all the vectors of subjects, subnetworks, and time windows constitutes a tensor. .in, It is a four-dimensional tensor with dimension . Number of subjects; The number of functional subnetworks, in the experiment ; Number of time windows; To hide dimensions, in the experiment .

[0047] S5. The token sequence of each functional subnetwork is independently evolved through time-mixed expert modeling to obtain the temporal representation of the functional subnetwork. Within each time window, the temporal tokens of each functional subnetwork are updated using a multi-head self-attention mechanism to capture the cooperative or competitive dynamics across functional subnetworks.

[0048] Since different subnetworks have different dynamic characteristics, a temporal hybrid expert is introduced. For the temporal token sequence of each functional subnetwork, its independent temporal evolution is modeled by the temporal hybrid expert module. Each functional subnetwork selects the most suitable temporal expert based on its overall temporal context through subnetwork-level gating. The expert output is weighted and fused to obtain the temporal representation of the subnetwork.

[0049] Specifically, for the subjects and each functional subnetwork Extract its time series Calculate the contextual representation of its entire time series: This context representation vector The overall behavioral tendencies of this functional subnetwork, such as overall activation level and fluctuation pattern, were captured.

[0050] set up A time expert, in the experiment Experts at every time It is a lightweight multilayer perceptron with input and output dimensions of 1. The entire time series is processed using a lightweight multilayer perceptron. Mapped to another sequence of the same length Each time expert has the same input, but each time expert has independent, learnable parameters.

[0051] The time expert may contain learnable recursive or convolutional structures (such as MLP), each time step is processed independently, and the sequence order depends on the time-ordered token sequence obtained by the window order.

[0052] Based on subjects context vector Time expert weights are generated through gating. :

[0053] in, These are trainable parameters; The number of time experts; .

[0054] context vector Linear mapping to The logistic values ​​of the dimension are then normalized to time-expert weights using a softmax activation function. Trainable parameters Learn end-to-end with the entire network.

[0055] A hard Top-1 routing method is used, retaining only the output of the time expert with the highest weight, and setting the outputs of other time experts to zero. This means that each sub-network is ultimately processed by only one time expert for its entire time series. This design allows different sub-networks to adaptively select the time expert best suited to handle that type of pattern based on their own temporal characteristics (slow changes or rapid fluctuations).

[0056] After passing through the time-mixing expert module, the subjects were obtained. Each functional subnetwork Complete temporal representation .

[0057] The advantage of this design is that different subnetworks can be assigned to different temporal experts, thus independently capturing their respective evolutionary rhythms. For example, the DMN might be assigned to an expert skilled at learning slow-wave changes, while the visual network might be assigned to an expert sensitive to transient stimuli.

[0058] The aforementioned temporal mixing experts process each subnetwork independently, but brain function depends on interactions between subsystems. To incorporate cross-subnetwork information, within each time window... Internally, the timing tokens of each functional sub-network will be stored. A standard Transformer encoder is input along the functional subnetwork dimension, enabling different functional subnetworks to exchange information and capture cooperative or competitive dynamics. This is achieved for a fixed time window. All functional subnetworks Time tokens Stack them along the sub-network dimensions to form a matrix. ; Indicates the subject subnetwork In the time window The timing tokens. The timing tokens for each functional sub-network. It is precisely because the sub-networks generated by S4 are in the same time window The temporal representation.

[0059] Transformer includes Each attention point, in the experiment of this embodiment Calculate the attention weights between each pair of subnetworks, so that each subnetwork token can aggregate information from other subnetworks, and output the updated time-series tokens of the functional subnetworks. .

[0060] S6. Inject the temporal state at the functional subnetwork level back into the spatial representation at the node level.

[0061] This embodiment uses a closed-loop feedback mechanism to add the temporal representation at the functional subnetwork level back to the spatial representation of each brain region node in the corresponding functional subnetwork in the form of residuals, thereby modulating the temporal state of the functional subnetwork on the spatial propagation at the node level and forming a two-way spatiotemporal coupling.

[0062] The closed-loop feedback mechanism is as follows: for those belonging to functional subnetworks nodes The update token of the corresponding functional sub-network in this time window is added to the node representation in the form of a residual:

[0063] in, It is a learnable scalar parameter, initialized to 0.5.

[0064] After processing by the Transformer encoder in S5, the updated functional subnet timing token is obtained. ,in, Represents a node Subnetwork In the time window The updated subnetwork temporal representation is then added back as a residual to the spatial representation of each node in the corresponding subnetwork within the same time window. This achieves closed-loop feedback.

[0065] This operation injects the temporal state at the functional subnetwork level into the representation of all nodes within the corresponding functional subnetwork on a time-window basis, enabling the spatial propagation of subsequent time windows to perceive the subnetwork dynamics of the current time window. Thus, the spatial graph propagation can depend on the subnetwork dynamics of the previous or current time, forming a tight spatiotemporal coupling.

[0066] S7. The timing tokens updated for each functional subnetwork are merged with the global topology information of the original functional connection matrix to obtain the enhanced timing tokens for the functional subnetworks.

[0067] S7 in this embodiment specifically includes: S71. Functional subnetwork pooling can lose fine-grained connection information between nodes, especially sparse long-range connections between different subnetworks, which may be smoothed out by averaging. Therefore, it is necessary to extract global topological features from the original functional connection matrix that has not been reordered. .

[0068] For each time window, take the upper triangular portion (excluding the diagonal) and quantize it:

[0069] in, Vectorization is represented. The upper triangle is represented by N, which represents the number of brain regions. The original functional adjacency matrix represents subject b and time window t.

[0070] S72. Since this dimension may be very large (200*199 / 2 = 19900), through a linear layer Compress it into a hidden dimension , to obtain global embedding :

[0071] S73, For each functional subnetwork , its time token With global embedding The data is then stitched together and then fused using gated residuals.

[0072]

[0073]

[0074] in, For the Sigmoid function, For element-wise multiplication, For learnable matrices, , .

[0075] The gate passes It adaptively determines how much information the current sub-network token needs to supplement from the global connection, while the residual preserves the original sub-network dynamics.

[0076] S8. Introduce learnable brain state prototypes, calculate the probability that each functional subnetwork temporal token belongs to each brain state, and obtain compact representations of the subject's state level through aggregation, thereby obtaining brain network analysis results.

[0077] In this embodiment, the enhanced functional subnetwork temporal tokens are flattened into subnetwork-time window pair token sequences, and a learnable brain state prototype is introduced. The probability of each token belonging to each brain state is calculated by a soft assignment function based on the square of cosine similarity and the prototype modulus normalization. Then, a state-level compact representation of the subject is obtained through weighted aggregation.

[0078] S8 specifically includes: S81. To obtain a compact and interpretable subject-level representation, a functional subnetwork temporal state clustering module is introduced.

[0079] Flatten all functional subnetworks and time window tokens for each subject as 1 token, obtain the token matrix :

[0080] Where E represents the number of functional subnetworks and T represents the number of time windows; This represents the timing token of the augmented functional subnetwork, which is the subject b, the functional subnetwork e, and the time window t; the superscript T indicates transpose.

[0081] S82, Introduction A learnable brain state prototype These brain state prototypes are jointly optimized with the model parameters during training.

[0082] Brain state prototype These are learnable parameters that are jointly optimized with other parts of the model (such as spatial experts, temporal experts, and classifiers) during training. Specifically: Number of prototypes. These are hyperparameters, used in experiments. Each prototype It is a dimension Trainable vectors, in experiments Brain state prototypes are randomly initialized and then learned from data through backpropagation. These prototypes eventually converge to cluster centers representing different "brain states" (such as default mode network activity state, attention network activity state, etc.).

[0083] S83, For token matrix The m-th row vector is the m-th token. For each token... Calculate its relationship with each prototype The allocation score encourages tokens to be associated with prototypes in the same direction, and prototypes with longer modulo lengths receive higher sensitivity.

[0084] in, As a soft allocation weight, it is used to allocate each token Aggregates to various brain state prototypes. Here, squared similarity is used and divided by... Instead of the usual The purpose is to emphasize the degree of consistency with the prototype direction, while also regularizing the prototype module length. The allocation matrix is... .

[0085] S84. Obtain the prototype-level aggregated vector through weighted summation. :

[0086] in, . The Line number The column element is . For the first The first subject The probability vector of each token pair for all prototypes, i.e. . It is an allocation matrix The The row represents the probability vector of the last token.

[0087] therefore, The The line is the first An aggregate vector of brain state prototypes, i.e., for all tokens According to their state The probabilities are weighted and summed.

[0088] Intuitively, The The row is a weighted sum of all tokens that may be associated with that state, representing the subject's state. The overall representation is then vectorized to obtain the final compact representation of the subject's state. Where vec represents vectorization, Indicates the number of brain state prototypes; To hide the dimension.

[0089] Compact representation of the subject's state level Input the classification model to obtain the classification prediction results of brain diseases. The entire classification model is trained end-to-end using binary cross-entropy loss. The classification model adopts a two-layer multilayer perceptron.

[0090] Compact representation of the subject's state level Input a two-layer multilayer perceptron and obtain the output. :

[0091] The probability of disease is .

[0092] Using binary cross-entropy loss :

[0093] in, Indicates the number of subjects; For the subjects Disease probability; These are learnable parameters; , For bias; This is the predicted value for subject b.

[0094] The entire model (including space / time experts, Transformer, prototype) The prototype (including gating parameters, etc.) is trained end-to-end using the Adam optimizer with an initial learning rate of 0.001, weight decay of 1e-4, and a selectable batch size of 32. No additional loss term needs to be added for clustering; the prototype automatically generates discriminative states through joint optimization with the classification task.

[0095] This embodiment uses the ABIDE I dataset as an example to illustrate the solution of this embodiment: Step 1: Data acquisition and preprocessing.

[0096] Resting-state fMRI data from 403 autistic patients and 468 healthy controls were downloaded from the ABIDE I database. Standard preprocessing was performed using the DPABI toolkit: removal of the first 10 time points to accommodate magnetic field saturation; slice timing correction; head movement correction (removal of subjects with head translation >2 mm or rotation >2 degrees); structure-function registration; normalization to MNI152 space; 4 mm Gaussian smoothing; delinear drift; bandpass filtering (0.01–0.1 Hz). The whole brain was then divided into 200 ROIs using the CC200 atlas, and the 200 ROIs were assigned to the 8 subnetworks according to the 7-network partitioning scheme of Yeo et al. (further refined to 8 subnetworks for consistency with this patent: cerebellum and subcortical, visual, somatic motor, dorsal attention, ventral attention, limbic, frontoparietal, and default mode).

[0097] Step 2: Dynamic functional connections are built and aligned with subnetworks.

[0098] The sliding window width is set to 20 TRs (TR = 2s, i.e., 40 seconds), and the step size is set to 5 TRs (10 seconds). Each participant will have approximately 140-200 time windows; this example uses [a specific time window]. As a typical example, the Pearson correlation within each window is calculated to obtain the functional connectivity matrix. Fisher's Z-transform is used to approximate a normal distribution for the correlation coefficients. A permutation matrix is ​​constructed based on the sub-network labels. For the functional connection matrix The node features are reordered, and the top 30% of edges by strength are retained to obtain a sparse adjacency matrix. .

[0099] Step 3: Spatial Mixing Expert.

[0100] Set up space experts Each expert MLP structure: input feature dimension (BOLD signal value only), first pass through a linear layer to 128, batch normalization, ReLU, then another linear layer to 128, residual connection. Compressed into a 200-dimensional vector, through Mapped to Weights, softmax. Using hard Top-1 routing, we obtain... . This indicates average pooling based on the feature dimension.

[0101] Step 4: Subnetwork tokenization and time-based hybrid expert.

[0102] By average pooling across the 8 subnetworks, we obtain Time experts Each expert is a two-layer MLP (128→128), processing sequences independently step-by-step (weights are shared horizontally, but since gating is applied to the entire sequence of each sub-network, experts can learn transformations in different frequency domains). Context vector. Dimension 128, gated output with 4-dimensional weights. After hard Top-1, the result is... .

[0103] Step 5: Attention across subnetworks and closed-loop feedback.

[0104] For each time window, the tokens of the 8 subnetworks are input into the Transformer encoder (4 heads, feedforward layer dimension 256), and the updated temporal tokens of the functional subnetworks are output. Then, for each node, extract the corresponding [item / symbol] from the sub-network token. Multiply by the learnable coefficient (Initial 0.5), added to the final spatial representation Up, refresh the node representation.

[0105] Step 6: Global connection fusion.

[0106] Extract the original functional connection matrix of each window The upper triangular layer has 19,900 dimensions, which is compressed to 128 dimensions by linear layers. For each sub-network, it is spliced... Gating calculation Residual fusion yields enhanced functional subnetwork timing tokens .

[0107] Step 7: Sub-network temporal state clustering.

[0108] Flatten all tokens: Each has 128 dimensions. Set the number of states. .prototype Random initialization. Calculate soft-assigned weights. ,have to Vectorization yields 768 dimensions.

[0109] Step 8: Categorize.

[0110] A two-layer MLP (768→256→1) with sigmoid output was used. Cross-entropy loss, Adam optimizer, learning rate 0.001, 200 training epochs, and 5-fold cross-validation were employed. The final accuracy was 76.98%, and the F1 score was 77.47%, outperforming all compared methods.

[0111] To verify the effectiveness of the proposed method, a series of representative baseline methods were selected for comparison, covering both static and dynamic brain network modeling research directions.

[0112] (1) BrainGNN: A typical graph neural network method based on functional connectivity (FC), which models the brain connectivity graph through graph convolution or message passing mechanism and is used for brain network classification tasks.

[0113] (2) Brain Transformer: A Transformer-based model that captures global interaction relationships between different brain regions through self-attention mechanisms.

[0114] (3) Dynamic functional connection modeling methods: including BrainMass and HGST, which explicitly model the temporal evolution between dynamic functional connection windows through temporal aggregation or hierarchical spatiotemporal modeling.

[0115] (4) Knowledge or structure-aware graph model methods: including KMGCN (a graph convolutional network based on knowledge and multi-graph structure) and CAGT (a graph Transformer model based on community structure awareness).

[0116] To ensure the fairness of the experiment, all methods were trained and evaluated under the same data preprocessing procedure and 5-fold cross-validation settings.

[0117] The proposed method demonstrates classification performance on three rs-fMRI benchmark datasets (ABIDE I, REST-MDD, and ADHD-200). Overall, the proposed method achieves state-of-the-art or competitive performance on all datasets. Specifically, on the ABIDE I dataset, the method achieves 76.98% accuracy and 77.47% F1 score; on the REST-MDD dataset, the accuracy reaches 71.40% and the F1 score reaches 73.38%, both outperforming dynamic modeling and community structure awareness methods, including CAGT and CroMDBN. Furthermore, this performance improvement is reflected in multiple evaluation metrics such as ACC, AUC, and F1, indicating that the model's improvement is not limited to a specific metric but stems from a more robust feature representation capability. These results demonstrate that modeling dynamic functional connectivity through spatiotemporal interaction via subnetwork alignment can more effectively characterize the complex dynamic features of brain networks compared to methods relying on globally shared temporal modeling or static subnetwork priors.

[0118] On the more challenging ADHD-200 dataset, the proposed method achieves an F1 score of 70.37%, a 4.79 percentage point improvement over the current state-of-the-art HGST method, while maintaining competitive classification accuracy. Compared to existing dynamic graph models and hierarchical spatiotemporal modeling methods, this method demonstrates a significant advantage in modeling temporal variations arising from multi-site data. This result is particularly noteworthy considering the significant subject heterogeneity and acquisition protocol differences in the ADHD-200 dataset. Overall, the stable performance on three multicenter datasets and various neuropsychiatric disease tasks further validates the robustness and generalization ability of the proposed framework in modeling the heterogeneous and asynchronous subnetwork dynamics in rs-fMRI.

[0119] Five ablation variants were designed: (i) Spatial MoE: replacing the spatial mixture expert with a single GIN; (ii) Temporal MoE: replacing the temporal mixture expert with a global LSTM; (iii) Subnetwork aligned routing: not performing node reordering but maintaining other structures; (iv) Closed-loop feedback: removing the closed-loop feedback step; and (v) Subnetwork temporal state clustering: removing subnetwork temporal state clustering and replacing it with direct global average pooling. The system was retrained and evaluated using a 5-fold algorithm under each variant.

[0120] The results are shown in Table 4: the performance of all variants decreased significantly, with the removal of subnetwork temporal state clustering having the greatest impact (accuracy decreased by more than 9%). Specifically: (1) Removing spatial MoE reduced ABIDE I accuracy to 74.84% and F1 to 72.84%; (2) Removing temporal MoE reduced REST-MDD F1 from 73.38% to 67.71%; (3) Removing subnetwork alignment reduced F1 on ABIDE I from 77.47% to 73.89%; (4) Removing closed-loop feedback reduced accuracy to 73.46%; and (5) Removing state clustering reduced accuracy by more than 9 percentage points. This demonstrates the necessity of each module.

[0121] Tables 1-3 show the comparison of experimental results between the method of this embodiment and the baseline model.

[0122] Table 1: Comparison of experimental results between the proposed method and the baseline model on the ABIDE I dataset.

[0123] Table 2: Comparison of experimental results between the proposed method and the baseline model on the REST-MDD dataset.

[0124] Table 3: Comparison of experimental results between the proposed method and the baseline model on the ADHD-200 dataset.

[0125] Table 4: Comparison of ablation experimental results of the present invention on different datasets

[0126] This embodiment demonstrates the dynamic subnetwork differential connectivity patterns extracted from the ABIDE I, REST-MDD, and ADHD-200 datasets, verifying the neuroscience interpretability of the model.

[0127] Step 1: Calculate the difference in average group levels.

[0128] For each dataset, the mean functional connectivity matrix for the patient group and the healthy control group is calculated separately within each time window. and Then calculate the difference matrix. .

[0129] Step 2: Sparsification and Visualization.

[0130] To highlight the most discriminative outliers, retain the data within each time window. Edges with values ​​in the top 30% (i.e., the upper 0.7 quantile) are set to zero, and the remaining edges are set to zero. Nodes are then colored according to the sub-network division, and the left and right hemispheres are drawn separately. Time windows are arranged from left to right and from top to bottom to observe the dynamic evolution of biomarkers.

[0131] Step 3: ABIDE I (Autism Spectrum Disorder) Results Analysis.

[0132] In the early time window, the differences were mainly concentrated in the connections between the dorsal attentional network (DAN), ventral attentional network (VAN), and frontoparietal network (FPN), as well as their interactions with sensory networks (visual VN, somatomotor SMN), suggesting possible abnormalities in external attention-oriented processing in the early stages of the disease. As the time window progressed, the differential connections gradually involved the coupling of the default mode network (DMN) with the frontoparietal and attentional networks, indicating a disharmony between self-reference processes and the control system in the later stages of the disease. Furthermore, hemispheric analysis showed that the differential connections in the right hemisphere were more concentrated in later time windows, reflecting a typical characteristic of right-sided abnormalities in autism spectrum disorders.

[0133] Step 4: REST-MDD (Depression) outcome analysis.

[0134] Aberrant connectivity was primarily concentrated in interactions between the DMN, FPN, and limbic network (LN), as well as coupling between the DMN and attentional networks. Specifically, DMN-FPN and DMN-LN aberrations spanned multiple time windows, while the involvement of sensory networks (VN, SMN) was more transient. The left hemisphere exhibited more persistent aberrations in later time windows, consistent with neuroimaging evidence of left prefrontal cortex dysfunction in depression.

[0135] Step 5: ADHD-200 (Attention Deficit Hyperactivity Disorder) Results Analysis.

[0136] Differential connectivity was primarily observed within higher-order cognitive networks such as the FPN, DAN, and VAN, and in their interactions with the DMN, while abnormalities in the sensorimotor networks were relatively less frequent. Notably, the control and attention networks in the right hemisphere showed more significant differences in later time windows, suggesting a possible persistent dysfunction of the right frontoparietal network in ADHD. This finding is consistent with the "right hemisphere delayed development" hypothesis of ADHD.

[0137] Step 6: Clinical significance.

[0138] These dynamic differential connectivity patterns not only validate that the subnetwork-specific asynchronous temporal dynamics captured in this invention are consistent with known neuropathological findings, but also provide potential quantitative biomarkers for disease subtyping and selection of individualized treatment time windows. For example, in early intervention for autism, changes in DAN-VN connectivity can be monitored; in the treatment of depression, DMN-FPN coupling strength can serve as an indicator for evaluating treatment efficacy.

[0139] This embodiment further utilizes the soft allocation extracted by the subnetwork temporal state clustering module. and state aggregation representation The study analyzed the brain state composition and temporal trajectory of the subjects.

[0140] Step 1: State definition.

[0141] Select the number of states This corresponds to six interpretable potential brain states (e.g., "high DMN-low attention", "high attention-low DMN", "sensory dominance", "control integration", "default hybridity", and "noisy state"). After training, each prototype... The weight vectors can be projected onto the subnetwork token space and assigned semantic labels through correlation analysis.

[0142] Step 2: Time series of individual subjects' states.

[0143] For a given subject, state assignment is performed on the tokens of each sub-network within each time window to obtain... The allocation matrix (each position represents the highest probability state) can be used to draw a "subnetwork-time window" state diagram, allowing for a visual observation of the asynchronous nature of state transitions between different subnetworks over time. For example, a healthy control may show the DMN frequently entering a "high DMN-low attention" state, while the attention network transitions to a "high attention-low DMN" state during task-related phases; whereas a patient with depression may be in a state of DMN overactivation for a long period and have difficulty switching to an attentional state.

[0144] Step 3: Statistics on differences in group-level status.

[0145] For both the patient and control groups, the average time spent in each state, the distribution of state dwell time, and the state transition probability matrix were statistically analyzed. Findings included: In ABIDE I, the patient group showed a significant increase in time spent in the "hybrid state" and "default hybrid state," while the transition probability from "high attention-low DMN" to "high DMN-low attention" decreased; in REST-MDD, the patient group showed a decrease in time spent in the "DMN-FPN coupled state," while the time spent in the "marginal dominant state" increased. These differences could serve as novel diagnostic features.

[0146] Step 4: Visualization and Explanation.

[0147] The prototype was obtained through t-SNE. Projecting the data onto a two-dimensional plane reveals a positive correlation between the distance between different prototypes and the connection strength between known functional systems, further demonstrating that the states learned in this invention have neuroscientific significance.

[0148] This embodiment details the selection of key hyperparameters and their impact, providing configuration guidance for practical applications.

[0149] (1) Sliding window parameters: Window width affects time-frequency resolution.

[0150] In this study, the window width was set to 20 TRs (TR = 2s, total 40 seconds), the step size was 5 TRs (10 seconds), and the overlap rate was 75%. Smaller window widths (e.g., 10 TRs) can improve temporal resolution but increase noise, while larger window widths (e.g., 40 TRs) allow for smooth and rapid changes. Through grid search, 20 TRs achieved the best balance between classification performance and trajectory smoothness.

[0151] (2) Functional subnetwork division: 200 ROIs were assigned to 8 functional subnetworks using the CC200 map. The division was based on the 7-network template of Yeo et al., with the addition of cerebellar and subcortical structures.

[0152] Other graphs (such as AAL and Harvard-Oxford) can also be used, but experiments show that the 8-sub-network partitioning achieves the best classification performance.

[0153] (3) Number of space experts Test from 2 to 16. Optimal performance at that time (accuracy 76.98%). Too little ( It is difficult to distinguish between different interaction modes, and there are too many ( This leads to some experts being idle and training becoming unstable. Therefore... It has the same number of subnets, and also reserves dedicated experts for each subnet.

[0154] (4) Number of time experts Test from 1 to 8. F1 is the highest. Equivalent to a global time series model, but with a significant performance degradation; It is prone to overfitting, especially for multicenter data.

[0155] (5) Hidden Dimensions The optimal performance levels are 64, 128, and 256. 128 achieves the best balance between performance and efficiency; 256 has a larger model capacity but doubles the training time and is prone to overfitting on small datasets.

[0156] (6) Number of brain states From 3 to 10. Time-based classification yields the best results, with each state having a clear semantic interpretation; too many states can lead to state fragmentation, making them difficult to interpret.

[0157] (7) Hard Top-1 routing vs. soft weighting: Hard routing is used for both spatial and temporal experts (only the highest-weighted expert is retained). Although soft weighting may slightly improve performance (about 0.3 percentage points), hard routing requires less computation and allows for greater specialization of each expert (because each node / subnetwork can only select one expert, forcing experts to learn different modes). Therefore, this invention uses hard Top-1 routing and utilizes load balancing loss (optional) to prevent expert collapse.

[0158] (8) Training parameters: Optimizer Adam, learning rate 0.001, weight decay 1e-4, batch size 32, maximum epochs 200, early stopping epochs 20, 5-fold cross-validation. All datasets used the same parameter settings to ensure fair comparison. Training a full model (5-fold average) on ABIDE I took approximately 2 hours (NVIDIA RTX 4090).

[0159] (9) Standardization and data augmentation: The BOLD signal of each subject was z-standardized within each ROI; when constructing the sliding window, the overlapping window itself was used to perform data augmentation without adding additional noise.

[0160] This embodiment acquires the resting-state functional magnetic resonance imaging (fMRI) time series of subjects and constructs a dynamic functional connectivity graph sequence using a sliding window. Based on prior knowledge of brain functional networks, brain regions are divided into multiple functional sub-networks, and nodes are reordered to achieve sub-network alignment. A spatial hybrid expert module performs multi-modal graph propagation on brain region nodes within each time window, with different experts focusing on heterogeneous spatial patterns such as intra-sub-network integration or cross-sub-network interaction. Sub-network tokenization and a temporal hybrid expert module independently model the temporal trajectory of each functional sub-network and capture inter-system dependencies using cross-sub-network self-attention. A closed-loop feedback mechanism injects the temporal state at the sub-network level back into the node-level spatial representation, achieving spatiotemporal bidirectional coupling. Global topological information from the original functional connectivity matrix is ​​fused to compensate for sub-network pooling loss. Interpretable latent brain states are extracted through sub-network temporal state clustering and aggregated into subject-level representations. Finally, the data is input into a classifier to output brain disease prediction results. This invention achieves state-of-the-art classification performance on three multicenter datasets for autism, depression, and ADHD. Ablation experiments verify the irreplaceability of each module, and dynamic differential connectivity analysis reveals subnetwork-specific time biomarkers with neuroscience significance.

[0161] Example 2 The purpose of this embodiment is to provide a brain network dynamic analysis system based on subnetwork alignment, including: The data processing module is configured to: construct a dynamic functional connectivity sequence using a sliding window for the resting-state functional magnetic resonance imaging time series of the subjects; The subnetwork alignment module is configured to: reorder the functional connection matrix and node features of each time window according to the pre-divided functional subnetworks, so that the nodes belonging to the same functional subnetwork are arranged continuously, and thus obtain the functional subnetwork alignment map. The spatial hybrid expert module is configured to: perform neighborhood aggregation and nonlinear transformation on the functional subnetwork alignment graph within each time window, capture the spatial interaction patterns of nodes, and obtain a node-level spatial representation; The temporal hybrid expert module is configured to: perform mean pooling on the spatial representation at the node level within each functional subnetwork to obtain temporal tokens at the functional subnetwork level; and model the contextual representation of the sequence tokens of each functional subnetwork over the entire time period through the temporal hybrid expert to obtain the temporal representation of the functional subnetwork. The cross-subnetwork interaction module is configured to: within each time window, use a multi-head self-attention mechanism to perform cross-subnetwork interaction update processing on the time-series tokens of each functional subnetwork, and capture the collaborative or competitive dynamics of cross-functional subnetworks. The global connectivity fusion module is configured to: merge the timing tokens updated by each functional subnetwork with the global topology information of the original functional connectivity matrix to obtain enhanced timing tokens for the functional subnetworks; The analysis module is configured to: introduce learnable brain state prototypes, calculate the probability that each functional subnetwork augmented temporal token belongs to each brain state, obtain compact representations of the subject's state level through aggregation, and then obtain brain network analysis results.

[0162] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0163] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0164] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0165] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0166] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0167] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0168] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0169] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0170] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0171] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0172] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for dynamic analysis of brain networks based on sub-network alignment, characterized in that, include: For the resting-state functional magnetic resonance imaging time series of the subjects, a dynamic functional connectivity sequence was constructed using a sliding window. Based on the pre-divided functional subnetworks, the functional connection matrix and node features of each time window are reordered so that nodes belonging to the same functional subnetwork are arranged continuously, resulting in a functional subnetwork alignment map. For the functional subnetwork alignment graph within each time window, neighborhood aggregation and nonlinear transformation are performed by a spatial hybridity expert to capture the spatial interaction patterns of nodes, thus obtaining a node-level spatial representation; specifically: Within each time period, the node's own node features are summarized to obtain a node-level description vector; For the node-level description vector, the spatial weight matrix of each node to each spatial expert is obtained by using linear mapping and activation function. A Top-1 routing mechanism is adopted, in which only the spatial expert output with the highest spatial weight is retained at each node, and the outputs of other spatial experts are set to zero, thus obtaining a node-level spatial representation. The spatial representations at the node level within each functional subnetwork are averaged and pooled to obtain temporal tokens at the functional subnetwork level. The contextual representations of the sequence tokens of each functional subnetwork over the entire time are then modeled by a temporal hybrid expert to obtain the temporal representations of the functional subnetworks. Specifically: The spatial representations at the node level within each functional subnetwork are averaged and pooled to obtain time-series tokens at the functional subnetwork level. For each functional subnetwork, calculate the context vector of the corresponding time token throughout the entire time series; For each functional subnetwork, the corresponding functional subnetwork-level time expert weights are obtained through multiple time experts for the entire time series context vector. A hard Top-1 routing mechanism is adopted, which retains only the time expert output with the highest time weight for the functional sub-network, and sets the outputs of other time experts to zero, so as to obtain the time sequence representation of the functional sub-network. Within each time window, a multi-head self-attention mechanism is used to perform cross-sub-network interactive update processing on the time-series tokens of each functional sub-network, capturing the cooperative or competitive dynamics of cross-functional sub-networks. The timing tokens updated for each functional subnetwork are fused with the global topology information of the original functional connection matrix to obtain the enhanced timing tokens for the functional subnetworks. By introducing learnable brain state prototypes, calculating the probability that each functional subnetwork augmented temporal token belongs to each brain state, and obtaining compact representations of the subject's state level through aggregation, the brain network analysis results are obtained.

2. The method of claim 1, wherein, Based on the pre-divided functional subnetworks, each time window is reordered so that nodes belonging to the same functional subnetwork are arranged consecutively, resulting in a functional subnetwork alignment diagram, specifically: Based on prior brain maps, The brain region nodes are divided into A non-overlapping functional subnetwork; The functional connectivity matrix and node features of each time window are reordered using a permutation matrix to obtain the reordered functional connectivity matrix and node features, and then the functional subnetwork alignment graph is obtained.

3. The method of claim 1, wherein, Within each time window, the temporal tokens of each functional subnetwork are processed using a multi-head self-attention mechanism to capture the cooperative or competitive dynamics across functional subnetworks. Specifically, within each time window, the temporal tokens of each functional subnetwork are stacked along the subnetwork dimension and then input into the Transformer encoder to calculate the interrelationships between the temporal tokens of all functional subnetworks, thus obtaining the updated temporal tokens of the functional subnetworks.

4. The method of claim 1, wherein, The timing tokens updated for each functional subnetwork are fused with the global topology information of the original functional connection matrix to obtain the enhanced timing tokens for the functional subnetworks, specifically: Extract the upper triangular part from the original functional connectivity matrix and quantize it, then project it through a linear layer to obtain the global embedding; The time-series token updated by each functional subnetwork is concatenated with the global embedding, and the concatenation feature and the time-series token updated by the functional subnetwork are fused to obtain the functional subnetwork enhanced time-series token.

5. The method of claim 1, wherein, By introducing learnable brain state prototypes, the probability of each functional subnetwork temporal token belonging to each brain state is calculated. Through aggregation, a compact representation of the subject's state level is obtained, leading to the brain network analysis results, specifically: The soft allocation weight is obtained by calculating the allocation score of each token and each prototype through the allocation function; the allocation function encourages tokens to be associated with prototypes that are in the same direction. Based on the soft-assignment weight of each token, a prototype-level aggregate vector is obtained by weighted summation; Vectorize the prototype-level aggregated vectors to obtain the final subject state-level compact representation; The vectorized final subject state-level compact representation is input into the classification model to obtain brain network analysis results. 6.A brain network dynamic analysis system based on sub-network alignment, characterized in that, include: The data processing module is configured to: construct a dynamic functional connectivity sequence using a sliding window for the resting-state functional magnetic resonance imaging time series of the subjects; The subnetwork alignment module is configured to: reorder the functional connection matrix and node features of each time window according to the pre-divided functional subnetworks, so that the nodes belonging to the same functional subnetwork are arranged continuously, and thus obtain the functional subnetwork alignment map. The spatial hybrid expert module is configured to: for the functional subnetwork alignment graph within each time window, perform neighborhood aggregation and nonlinear transformation through the spatial hybrid expert to capture the spatial interaction patterns of nodes and obtain node-level spatial representations; specifically: Within each time period, the node's own node features are summarized to obtain a node-level description vector; For the node-level description vector, the spatial weight matrix of each node to each spatial expert is obtained by using linear mapping and activation function. A Top-1 routing mechanism is adopted, in which only the spatial expert output with the highest spatial weight is retained at each node, and the outputs of other spatial experts are set to zero, thus obtaining a node-level spatial representation. The temporal hybrid expert module is configured to: perform mean pooling on the spatial representation at the node level within each functional subnetwork to obtain temporal tokens at the functional subnetwork level; and model the contextual representation of the sequence tokens of each functional subnetwork over the entire time period through the temporal hybrid expert to obtain the temporal representation of the functional subnetwork. Specifically: The spatial representations at the node level within each functional subnetwork are averaged and pooled to obtain time-series tokens at the functional subnetwork level. For each functional subnetwork, calculate the context vector of the corresponding time token throughout the entire time series; For each functional subnetwork, the corresponding functional subnetwork-level time expert weights are obtained through multiple time experts for the entire time series context vector. A hard Top-1 routing mechanism is adopted, which retains only the time expert output with the highest time weight for the functional sub-network, and sets the outputs of other time experts to zero, so as to obtain the time sequence representation of the functional sub-network. The cross-subnetwork interaction module is configured to: within each time window, use a multi-head self-attention mechanism to perform cross-subnetwork interaction update processing on the time-series tokens of each functional subnetwork, and capture the collaborative or competitive dynamics of cross-functional subnetworks. The global connectivity fusion module is configured to: merge the timing tokens updated by each functional subnetwork with the global topology information of the original functional connectivity matrix to obtain enhanced timing tokens for the functional subnetworks; The analysis module is configured to: introduce learnable brain state prototypes, calculate the probability that each functional subnetwork augmented temporal token belongs to each brain state, obtain compact representations of the subject's state level through aggregation, and then obtain brain network analysis results.

7. An electronic device, comprising: It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-5.

Images