CN-116662850-B - Brain network generation system and method based on functional brain network classification task

CN116662850BCN 116662850 BCN116662850 BCN 116662850BCN-116662850-B

Abstract

The invention discloses a brain network generation system and method based on a functional brain network classification task, wherein the system comprises a time sequence encoder, a graph generator and a graph predictor based on a graph convolution neural network, which are sequentially connected, the time sequence encoder is used for encoding a Bold signal sequence, the graph generator is used for converting the encoded time sequence characteristics into a functional brain network graph, the graph predictor is used for carrying out classification prediction on the functional brain network graph, the time sequence encoder comprises a transform encoder layer and a multi-layer perceptron A, which are sequentially connected, the transform encoder layer is used for extracting the time sequence characteristics, and the multi-layer perceptron A is used for generating node characteristics of a plurality of ROIs after dimension reduction processing of the extracted time sequence characteristics. According to the invention, the correlation matrix of the brain network is generated by directly extracting the BOLD signal sequences from each brain region, so that the classification accuracy of the functional brain network is improved.

Inventors

XU JUNHAI
LI KEXIN
WEI JIANGUO

Assignees

天津大学

Dates

Publication Date: 20260512
Application Date: 20230427

Claims (10)

1. A brain network generation system based on a functional brain network classification task is characterized by comprising a time sequence encoder, a graph generator and a graph predictor based on a graph convolution neural network, which are sequentially connected, wherein the time sequence encoder is used for encoding a Bold signal sequence, the graph generator is used for converting the encoded time sequence characteristics into a functional brain network graph, the graph predictor is used for carrying out classification prediction on the functional brain network graph, the time sequence encoder comprises a encoder layer of a transducer and a multi-layer perceptron A which are sequentially connected, the encoder layer of the transducer is used for extracting the time sequence characteristics, the multi-layer perceptron A is used for generating node characteristics of a plurality of ROIs after dimension reduction processing of the extracted time sequence characteristics, and the graph generator is used for converting the dimension reduced time sequence characteristics into a communication matrix of a brain region.
2. The brain network generation system based on functional brain network classification tasks according to claim 1, wherein the graph generator generates a machine-learnable connected matrix Connected matrix The middle elements are the paired connectivity intensities between ROIs; wherein , For the time series features extracted by the time series encoder, Is a normalized time series feature.
3. The brain network generation system based on the functional brain network classification task according to claim 1, wherein the graph predictor comprises a graph convolution neural network, a regularization layer and a multi-layer perceptron B which are sequentially connected, wherein the graph convolution neural network is used for extracting node characteristics of the functional brain network graph, and the multi-layer perceptron B is used for carrying out classification prediction on the node characteristics.
4. A brain network generation method based on a functional brain network classification task is characterized by constructing a brain network generation system based on the functional brain network classification task, wherein the system comprises a time sequence encoder, a graph generator and a graph predictor based on a graph convolution neural network, which are sequentially connected, the time sequence encoder is used for encoding a Bold signal sequence, the graph generator is used for converting the encoded time sequence characteristics into a functional brain network graph, the graph predictor is used for carrying out classification prediction on the functional brain network graph, the time sequence encoder comprises an encoder layer of a transducer and a multi-layer perceptron A which are sequentially connected, the encoder layer of the transducer is used for extracting the time sequence characteristics, the multi-layer perceptron A is used for generating node characteristics of a plurality of ROIs after dimension reduction processing on the extracted time sequence characteristics, and the dimension reduced time sequence characteristics are converted into a communication matrix of a brain region by adopting the graph generator.
5. The brain network generation method based on the functional brain network classification task according to claim 4, characterized in that the method comprises the steps of: firstly, extracting time sequence features from a Bold signal sequence acquired by functional nuclear magnetic resonance by adopting an encoder layer of a transducer; Secondly, adopting a multi-layer perceptron A to perform dimension reduction treatment on the extracted time sequence features; thirdly, converting the time sequence characteristics after dimension reduction by adopting a graph generator to generate a communication matrix of the brain region; Assigning an initial value to each brain region node by a row vector of a brain region communication matrix calculated based on a Pearson correlation coefficient method to obtain initial characteristics of the brain region; step four, multiplying the initial characteristics of the brain region with a connected matrix generated by a graph generator to obtain a functional brain network graph, and inputting the functional brain network graph into a graph convolution neural network; And fifthly, extracting node characteristics of the functional brain network graph by using the graph convolution neural network, splicing all node characteristics to generate characteristics of the whole graph, inputting the characteristics into the multi-layer perceptron B, and outputting a classification result of the brain network by using the multi-layer perceptron B.
6. The brain network generation method based on the functional brain network classification task according to claim 5, wherein in the third step, the node is Node characteristics of (2) Initialization as node A vector of pearson correlation scores between all nodes in the functional brain network graph.
7. The brain network generation method based on functional brain network classification tasks according to claim 4, wherein the system is constructed based on an end-to-end framework, the system is trained and optimized by a downstream functional brain network classification task in the end-to-end framework, the training penalty includes inter-class penalty, matrix sparsity penalty and cross entropy penalty, the inter-class penalty is used for minimizing differences between connectivity matrices between the same class, the inter-class penalty is used for maximally improving differences between connection matrices between different classes while maintaining connection matrix similarity between the same class, and the matrix sparsity penalty is used for reducing the degree of bias caused by large values in the functional brain network graph.
8. The brain network generation method based on the functional brain network classification task according to claim 7, wherein the inter-class loss is set as Given a certain class Sum set The collection contains labels Is used to determine the index of all the samples of (a), The calculation formula is as follows: ; ; Wherein: c represents the set of all samples tested; c represents a test sample; q represents the index of the sample under test; A label representing a test specimen; k represents an index of the sample functional brain network map; a functional brain network graph generated by the graph generator; a functional brain network graph representing all samples of the same label; representing the mean of the different classes of samples; representing the variance of the same label sample.
9. The brain network generation method based on the functional brain network classification task according to claim 7, wherein the inter-class loss is set as , The calculation formula of (2) is as follows: ; Wherein: e represents the index of the first type of sample; f represents the index of the second class of samples; a represents a first type of sample; b represents a second type of sample; a functional brain network graph representing a first class of samples; a functional brain network graph representing a second type of sample; A variance of the functional brain network map representing a first class of samples; representing the variance of the functional brain network map of the second class of samples; A set of all functional brain network graphs representing a first class of samples; A set of all functional brain network graphs representing a second class of samples; A mean value of the functional brain network graph representing the first type of sample; a mean value of the functional brain network graph representing the second type of sample.
10. The brain network generation method based on the functional brain network classification task according to claim 4, wherein a position embedding layer added to the input data in the encoder layer of the transducer is removed, and the number of attention heads is set to 1.

Description

Brain network generation system and method based on functional brain network classification task Technical Field The invention relates to the technical field of deep learning, in particular to a brain network generation system and method based on functional brain network classification tasks. Background Analysis of brain network functional connections, through which the individual's human brain tissue can be understood, is becoming increasingly important in brain imaging studies. There are a number of findings in neuroscience research that indicate that the neural circuit is critical for understanding differences in brain function between populations. Functional magnetic resonance imaging (fMRI) is one of the most common imaging modalities for studying brain function and tissue. The brain network of the neuroimaging community is of great interest in classifying individuals. In the last few years, graph roll-up neural networks (GCNs) have been developed as a powerful method to process graph structures. The advantages of GCN result from combining local node level information with global domain level information. Similar to the standard two-dimensional image convolution, which pools the values of neighboring nodes, the purpose of the graph convolution is to aggregate neighborhood information of nodes in the graph using the graph laplace operator. Existing brain network analysis work typically consists of two steps, the first of which is to generate a functional brain network from fMRI data of an individual. This is typically accomplished by selecting a brain map with a set of regions of interest (ROIs) as nodes, and extracting fMRI Blood Oxygen Level Dependent (BOLD) signal sequences from each brain region. For edge generation, pairwise connectivity between node pairs is calculated using pearson correlation and partial correlation metrics. In a second stage, the brain connectivity metrics obtained between all node pairs are then used to classify or predict the individual. By using the graph convolution neural network, higher-order information in the functional brain network can be better extracted, so that the classification accuracy of the functional brain network is improved. Most of the existing brain network classification methods are based on correlation scores, which suffer from two drawbacks, namely, firstly, the correlation method is focused on capturing linear correlations and ignoring the time sequence, which means that disrupting the time step does not change the results. Second, the trend of applying a graph-roll neural network (GCN) to brain connection matrices for functional magnetic resonance imaging (fMRI) is becoming more and more evident, while the mechanism of most GCNs (i.e., messaging) is not compatible with existing functional brain networks that have both positive and negative weighted sides. The prior art is not compatible with most graph convolutional neural networks, but does not capture the non-linear, temporal relationships of brain network data. Therefore, these efforts have to design additional processes in the graph roll-up neural network to handle the negative weights. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides a brain network generation system and method based on a functional brain network classification task. In order to achieve the above purpose, the present invention adopts the following technical scheme: A brain network generation system based on a functional brain network classification task comprises a time sequence encoder, a graph generator and a graph predictor based on a graph convolution neural network, wherein the time sequence encoder, the graph generator and the graph predictor are sequentially connected, the time sequence encoder is used for encoding a Bold signal sequence, the graph generator is used for converting the encoded time sequence characteristics into a functional brain network graph, the graph predictor is used for carrying out classification prediction on the functional brain network graph, the time sequence encoder comprises a transducer encoder layer and a multi-layer perceptron A, which are sequentially connected, the transducer encoder layer is used for extracting the time sequence characteristics, and the multi-layer perceptron A is used for generating node characteristics of a plurality of ROIs after dimension reduction processing on the extracted time sequence characteristics. Further, the graph generator generates a machine-leachable connected matrix A, wherein elements in the connected matrix A are paired connectivity intensities between the ROIs; Where h A＝softmax(he),he is the time series feature extracted by the time series encoder and h A is the normalized time series feature. The graph predictor further comprises a graph convolution neural network, a regularization layer and a multi-layer perceptron B which are connected in sequence, wherein the graph convolution neural network is used