CN-121687549-B - ADHD brain function connection dynamic characterization system and method based on space-time diagram

CN121687549BCN 121687549 BCN121687549 BCN 121687549BCN-121687549-B

Abstract

ADHD brain function connection dynamic characterization system and method based on space-time diagram. Belongs to the technical field of medical image processing and artificial intelligence intersection. The method solves the defects of the existing deep learning method in constructing dynamic characterization of ADHD brain function connection. The system designs an end-to-end learnable self-adaptive graph construction module, automatically discovers an individuation topology through node embedding learning and dot product concentration scaling, combines Top-K sparsification and prior graph fusion based on RBF kernels to ensure physiological rationality, provides double-branch space-time graph convolution, uses expansion causal convolution for time branches, uses graph meaning force for space branches, realizes space-time joint modeling through self-adaptive gating fusion, introduces a multi-view comparison learning framework, designs three types of enhancement (time domain; frequency domain; graph structure) guided by knowledge in the field, and fully utilizes non-labeling data by adopting InfoNCE loss and two-stage training.

Inventors

HAN QIULEI
GU HE
LI ZIZHENG
YE HONGBIAO
SUN YAN
Song ze
ZHAO JIAN
Kuang Zhejun
SHI LIJUAN
TAN LU

Assignees

长春大学

Dates

Publication Date: 20260512
Application Date: 20260209

Claims (7)

1. A space-time diagram-based dynamic representation system for ADHD brain function connections, the system comprising: The data preprocessing and node characteristic encoding module is used for collecting EEG signals of a subject for preprocessing, encoding the EEG signals of each channel into characteristic vectors through node characteristic encoding to obtain a node characteristic matrix ; The self-adaptive graph construction module sequentially performs node embedded learning, self-adaptive adjacent matrix generation and priori knowledge fusion on the node characteristic matrix to obtain a final adjacent matrix ; Node embedding learning is specifically to learn a low-dimensional embedding vector for each node for capturing the inherent properties of the node and its role in the graph by Is carried out, wherein, Representing the final node's embedding, Representing the parameters as Is a multi-layer perceptron of (2); the self-adaptive adjacency matrix is generated by calculating the similarity between node pairs as an edge weight based on final node embedding to obtain a similarity matrix For the following Each node in the (B) is subjected to Top-K sparsification strategy to obtain a sparse adjacent matrix For sparse adjacent matrix Performing symmetry and normalization to obtain data-driven adaptive adjacency matrix ; The prior knowledge fusion is specifically implemented by constructing a prior adjacency matrix according to the Euclidean distance between electrodes by using a Gaussian radial basis function Obtaining final adjacent matrix by weighting combination : Wherein Is a learnable weight parameter; The stacked space-time diagram convolution module comprises a first layer space-time diagram convolution module and a second layer space-time diagram convolution module which are respectively formed by the two layers of space-time diagram convolution modules, wherein in each layer of space-time diagram convolution module, the time sequence dynamic characteristics in each node are captured through time convolution branches, the space neighborhood information is aggregated through graph volume integration branches, the output of the time convolution branches and the output of the graph convolution branches are fused through a gating mechanism to obtain fusion characteristics, and the second layer of space-time diagram convolution module also obtains residual fusion characteristics through residual connection ; The image pooling and global characterization learning module is used for gradually updating the adjacent matrix through a layering image pooling strategy And residual fusion feature Residual fusion feature through adaptive graph reading mechanism Different importance weights are distributed to different nodes in the graph to obtain the global graph representation And robust enhancement operation is carried out, and finally, the final global map representation is obtained ; The contrast learning module is used for enhancing the EEG signals through a multi-view data enhancement strategy, constructing InfoNCE loss functions for contrast learning, and improving the final global diagram representation by adopting a two-stage training strategy Is an accurate rate of (a).
2. The dynamic characterization system of ADHD brain function connection based on space-time diagram according to claim 1, wherein the node feature encoding is performed by a node feature encoding module that adopts a one-dimensional time convolutional neural network architecture, and comprises three convolutional blocks and a global adaptive average pooling layer, each convolutional block is composed of a convolutional layer, a batch normalization layer and a ReLU activation function, wherein the first convolutional layer captures the short-time oscillation mode of EEG, the second convolutional layer further extracts abstract timing features, the third convolutional layer enhances feature expression capability while maintaining timing resolution, and the feature sequences of different lengths are unified into feature vectors of fixed length through global adaptive average pooling to obtain a node feature matrix 。
3. The space-time diagram based dynamic characterization system for ADHD brain function connection according to claim 2, wherein, In the first layer space-time diagram convolution module, the operations in the time convolution branches are specifically as follows: Wherein, the method comprises the steps of, Representing the output characteristics of the first layer time convolved branches, Representing a time-sequential convolutional network; the operations in the graph convolution branches are specifically: Wherein, the method comprises the steps of, Representing the output characteristics of the convolved branches of the first layer graph, Representing a graph attention network; The fusion characteristics of the convolution branches of the first layer diagram are as follows: the method comprises the steps of obtaining, among others, Is an adaptive gating factor.
4. A space-time diagram based ADHD brain function connection dynamic characterization system according to claim 3, characterized in that in the second layer space-time diagram convolution module, the operations in the time convolution branches are specifically: Wherein, the method comprises the steps of, Representing the output characteristics of the second layer time convolution branches; the operations in the graph convolution branches are specifically: Wherein, the method comprises the steps of, Representing the output characteristics of the convolution branches of the second layer graph; The fusion characteristics of the convolution branches of the second layer diagram are as follows: obtaining; obtaining residual fusion characteristics through residual connection in second-layer space-time diagram convolution module The method comprises the following steps: , wherein, Representing a regularization operation.
5. The space-time diagram based dynamic characterization system for ADHD brain function connection according to claim 4, The hierarchical graph pooling strategy specifically comprises the steps of calculating a node selection matrix , wherein, A graph convolution operation is shown and is described, Representing the number of layers that the hierarchical pooling strategy performs, ; Initial as Subsequently by It is derived that the method comprises the steps of, Initial as Subsequently by Obtaining; Global graph characterization By passing through The method comprises the steps of obtaining, among others, For the number of channels of the EEG signal, Representation of Middle (f) The feature vectors of the individual nodes are used, Represent the first Attention weights of the individual nodes; final global graph characterization By passing through The method comprises the steps of obtaining, among others, Representing a matrix of projection layer weights, Representing the projection layer bias vector, In order to activate the function, Wherein In order to perform the averaging operation, To take the maximum value.
6. The space-time diagram based dynamic characterization system for ADHD brain function connection according to claim 5, wherein, Multi-view data enhancement strategies include time domain enhancement, frequency domain enhancement, and graph structure enhancement; The InfoNCE loss function is specifically: Wherein, the method comprises the steps of, And Is a representation of two enhanced views of the same sample, Is the first The characterization vector of the individual samples after passing the projection head, Is of the size of a batch of material, Is an indication function of the display, Is a temperature parameter, a similarity function Cosine similarity is used: ; the two-stage training strategy includes self-supervised pre-training and supervised fine tuning.
7. A dynamic characterization method for ADHD brain function connection based on a space-time diagram, characterized in that the method comprises: collecting EEG signals of a subject, preprocessing, and encoding the EEG signals of each channel into feature vectors by node feature encoding to obtain a node feature matrix ; Sequentially performing node embedded learning, adaptive adjacency matrix generation and priori knowledge fusion on the node characteristic matrix to obtain a final adjacency matrix ; Node embedding learning is specifically to learn a low-dimensional embedding vector for each node for capturing the inherent properties of the node and its role in the graph by Is carried out, wherein, Representing the final node's embedding, Representing the parameters as Is a multi-layer perceptron of (2); the self-adaptive adjacency matrix is generated by calculating the similarity between node pairs as an edge weight based on final node embedding to obtain a similarity matrix For the following Each node in the (B) is subjected to Top-K sparsification strategy to obtain a sparse adjacent matrix For sparse adjacent matrix Performing symmetry and normalization to obtain data-driven adaptive adjacency matrix ; The prior knowledge fusion is specifically implemented by constructing a prior adjacency matrix according to the Euclidean distance between electrodes by using a Gaussian radial basis function Obtaining final adjacent matrix by weighting combination : Wherein Is a learnable weight parameter; The method comprises the steps of carrying out space-time diagram convolution by using two layers of space-time diagram convolution modules, wherein the two layers of space-time diagram convolution modules comprise a first layer of space-time diagram convolution module and a second layer of space-time diagram convolution module, capturing time sequence dynamic characteristics in each node by using a time convolution branch in each layer of space-time diagram convolution module, aggregating space neighborhood information by using a graph integration branch, fusing outputs of the time convolution branch and the graph convolution branch by using a gating mechanism to obtain fusion characteristics, and obtaining residual fusion characteristics by residual connection in the second layer of space-time diagram convolution module ; Gradually updating the adjacency matrix through a hierarchical graph pooling strategy And residual fusion feature Residual fusion feature through adaptive graph reading mechanism Different importance weights are distributed to different nodes in the graph to obtain the global graph representation And robust enhancement operation is carried out, and finally, the final global map representation is obtained ; Enhancement of EEG signals by a multi-view data enhancement strategy, construction InfoNCE of a loss function for contrast learning, and enhancement of final global map representation by a two-stage training strategy Is an accurate rate of (a).

Description

ADHD brain function connection dynamic characterization system and method based on space-time diagram Technical Field The invention belongs to the technical field of medical image processing and artificial intelligence intersection, and particularly relates to a dynamic ADHD brain function connection characterization system and method based on a space-time diagram. Background Attention Deficit Hyperactivity Disorder (ADHD) is a neurological disorder in children with a global prevalence of 5-7%. ADHD classification methods based on electroencephalogram (EEG) are paid attention to due to noninvasive property and high time resolution, but the existing deep learning methods face three key challenges when constructing dynamic characterization of ADHD brain function connection (1) rely on static predefined graph structures, ignore individual heterogeneity and time-varying dynamics of function connection, (2) perform sequential and spatial modeling serial processing, not capture inherent space-time coupling characteristics of EEG, and (3) perform purely supervised learning to be limited in generalization capability under a small sample scene. Disclosure of Invention In order to solve the defects of the existing deep learning method in constructing dynamic characterization of ADHD brain function connection, the application provides a dynamic characterization system and a dynamic characterization method of ADHD brain function connection based on a space-time diagram. The system comprises: The data preprocessing and node characteristic coding module is used for collecting EEG signals of a subject to preprocess, and coding the EEG signals of each channel into characteristic vectors through node characteristic codes to obtain a node characteristic matrix; the self-adaptive graph construction module sequentially performs node embedded learning, self-adaptive adjacent matrix generation and priori knowledge fusion on the node characteristic matrix to obtain a final adjacent matrix ; The stacked space-time diagram convolution module comprises a first layer space-time diagram convolution module and a second layer space-time diagram convolution module which are respectively formed by the two layers of space-time diagram convolution modules, wherein in each layer of space-time diagram convolution module, the time sequence dynamic characteristics in each node are captured through time convolution branches, the space neighborhood information is aggregated through graph volume integration branches, the output of the time convolution branches and the output of the graph convolution branches are fused through a gating mechanism to obtain fusion characteristics, and the second layer of space-time diagram convolution module also obtains residual fusion characteristics through residual connection; The image pooling and global characterization learning module is used for gradually updating the adjacent matrix through a layering image pooling strategyAnd residual fusion featureResidual fusion feature through adaptive graph reading mechanismDifferent importance weights are distributed to different nodes in the graph to obtain the global graph representationAnd robust enhancement operation is carried out, and finally, the final global map representation is obtained; The contrast learning module is used for enhancing the EEG signals through a multi-view data enhancement strategy, constructing InfoNCE loss functions for contrast learning, and improving the global map representation by adopting a two-stage training strategyIs an accurate rate of (a). Further, node feature coding is performed by a node feature coding module, the module adopts a one-dimensional time convolutional neural network architecture and comprises three convolutional blocks and a global self-adaptive average pooling layer, each convolutional block consists of a convolutional layer, a batch normalization layer and a ReLU activation function, wherein the first convolutional layer captures a short-time oscillation mode of EEG, the second convolutional layer further extracts abstract time sequence features, the third convolutional layer enhances feature expression capability while maintaining time sequence resolution, and feature sequences with different lengths are unified into feature vectors with fixed lengths through global self-adaptive average pooling to obtain a node feature matrix。 Further, node embedding learning is specifically to learn a low-dimensional embedding vector for each node for capturing the intrinsic properties of the node and the roles in the graph byIs carried out, wherein,Representing the final node's embedding,Representing the parameters asIs a multi-layer perceptron of (2); the self-adaptive adjacency matrix is generated by calculating the similarity between node pairs as an edge weight based on final node embedding to obtain a similarity matrix For the followingEach node in the (B) is subjected to Top-K sparsification strategy to obtain a sparse