Autism classification algorithm based on combination of dynamic and static brain networks

Abstract

The invention discloses an autism classification algorithm based on combination of dynamic and static brain networks. A static brain network encoder effectively extracts global brain network features and generates functional connections with richer semantic information; a dynamic brain network encoder independently encodes each subsequence in the dynamic connection and fuses all codes according to an attention mechanism; and the graph classifier recombines the results of the first two modules into a high-order function connection brain graph and performs classification based on a GCN algorithm. According to the method, deep features in a static function connection matrix are effectively learned by introducing the static brain network encoder, and the dynamic brain network encoder module is introduced, so that the method has the capabilities of learning time dynamics of a brain network and integrating dynamic function connection, and the brain fitting capability of the model is greatly improved; and finally, through representation learning of a graph classifier, the model can deeply learn classification features of the brain to diseases from two aspects of a dynamic brain network and a static brain network, so that the classification performance of the model can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of neuroimaging analysis, and in particular to an autism classification algorithm based on the combination of dynamic and static brain networks. Background Art

[0002] In recent years, with the rapid development of medical imaging technology, functional magnetic resonance imaging (fMRI) has become one of the new directions for researchers to diagnose autism. In some recent work, researchers have used fMRI data to construct a functional brain network (FBN) to represent each subject. In FBN, each node represents a corresponding brain region (Region of Interest, ROI), and the edge weight between nodes corresponds to the connection strength between the two ROIs.

[0003] In the existing FBN analysis framework, researchers usually build low-order brain networks for learning based on the Pearson Correlation Coefficient (PCC), but the Pearson correlation mainly measures linear relationships and is less sensitive to nonlinear relationships. This traditional functional brain network does not fit the brain well enough and cannot efficiently express the complex relationships between brain regions, thus affecting the accuracy of downstream classification tasks. In addition, existing methods usually use static functional connections or dynamic functional connections alone for training. Studies have shown that both dynamic and static functional connections make important contributions to the accurate classification of autism. However, due to the high dimensionality and complexity of dynamic functional connections, these two types of features cannot be combined well. Therefore, it is necessary to design a model that can successfully integrate the two types of functional connections, and this model must also be able to effectively extract deep relationships between brain regions. Summary of the invention

[0004] The purpose of the present invention is to make up for the fact that most algorithms in the existing technology usually directly use low-order functional brain networks for learning, which makes the model not fit the brain well enough. Most classification algorithms only use static functional connections or dynamic functional connections for training, which leads to the defect of incomplete model learning. This invention provides an autism classification algorithm based on the combination of dynamic and static brain networks.

[0005] The present invention is achieved through the following technical solutions: An autism classification algorithm based on the combination of dynamic and static brain networks includes a static brain network encoder, a dynamic brain network encoder and a graph classifier, and specifically includes the following steps: (1) The obtained static functional connection matrix is ​​input into the static brain network encoder, and the deep relationship between brain region nodes is learned from the static functional connection matrix through the self-attention mechanism to generate a high-order functional connection matrix; (2) The dynamic functional connectivity matrix obtained by the sliding window algorithm is input into the dynamic brain network encoder. After learning the dynamic features of the brain network using the self-attention mechanism, the channel attention mechanism is used to integrate the information to obtain the high-order dynamic feature embedding of each brain region. (3) Based on the results obtained by the static brain network encoder and the dynamic brain network encoder, the graph classifier uses the graph convolutional network algorithm and the topK pooling algorithm to learn and achieve the classification of autism.

[0006] In step (1), the Pearson correlation is used to obtain the static functional connectivity matrix of each subject. After obtaining the static functional connectivity matrix, a multi-head self-attention network is used as a functional connectivity encoder. For each pre-segmented brain region node, the corresponding row vector in the static functional connectivity matrix is ​​used as its feature vector, and it is input into the functional connectivity encoder to obtain the high-order embedding vector of each brain region and vertically concatenate it to obtain the brain region node feature representation of the subject. After obtaining the brain region node feature representation of the subject, a functional connectivity generator is used to generate a high-order functional connectivity matrix. The calculation method is to encode the learned brain region node features using the softmax function and multiply them with their transpose to obtain a new symmetric matrix, namely the high-order functional connectivity matrix.

[0007] In step (2), the dynamic functional connectivity matrix of each subject is obtained by using Pearson correlation. After obtaining the dynamic functional connectivity matrix, a multi-head self-attention network is used as a functional connectivity encoder. For different subsequence connection matrices in the dynamic functional connectivity matrix, a transformer network with shared weights is used to encode them. For each subsequence, the functional connectivity encoder is calculated in the same way as the functional connectivity encoder in the static connection encoder. After calculating multiple subsequences, a multi-channel brain region node feature representation of the subject is obtained, in which each channel corresponds to a subsequence. After obtaining the multi-channel brain region node feature representation, a channel attention network SENet is used to fuse its features. The calculation method is to use the SE block in SENet to generate a corresponding attention score for each channel, then multiply each channel by its score, and then add all the results to obtain the high-order dynamic feature embedding of each brain region.

[0008] In step (3), since the two results have the same number of brain region nodes, they are combined into a high-order functional connection map, in which the output matrix of the static brain network encoder is used as the adjacency matrix of the high-order brain map, and the input feature embedding of the dynamic brain network encoder is used as the node feature of the high-order brain map; after obtaining the high-order functional connection map, a two-layer graph convolutional network is used for learning. The calculation method is to first extract the features of each node and transmit its features to the neighboring nodes of the node, and then each node performs weighted fusion according to its own features and the features of the neighborhood and updates its own feature representation; after learning through the two-layer graph convolution layer, it is input into a topK pooling layer. The calculation method is to generate a corresponding importance score according to the feature vector of each node, retain the first K important nodes, discard the remaining nodes, and reduce the dimension of the feature; after the dimensionality reduction through the pooling layer, the feature representation of the entire image is obtained by using the readout layer and input into the multi-layer perceptron for classification. The calculation method is to multiply the feature vector by the weight vector learned by the perceptron, add the bias term, and then perform nonlinear activation through the activation function to finally obtain the classification result of the subject.

[0009] The advantages of the present invention are: by introducing a static brain network encoder, the present invention can effectively learn the deep features in the static functional connection matrix, and by introducing a dynamic brain network encoder module, it has the ability to learn the temporal dynamics of the brain network and integrate dynamic functional connections, which greatly improves the brain fitting ability of the model. Finally, after representation learning by a graph classifier, the model can deeply learn the brain's classification characteristics of diseases from both dynamic brain networks and static brain networks, thereby improving the classification performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 is a flow chart of the method of the present invention; Figure 2 is a diagram of a deep model architecture according to an embodiment of the algorithm proposed by the present invention; Figure 3 FIG. 4 is a diagram of a functional connection encoder architecture according to an embodiment of the algorithm proposed in the present invention. DETAILED DESCRIPTION

[0011] like Figure 1 As shown, an autism classification algorithm based on the combination of dynamic and static brain networks includes a static brain network encoder, a dynamic brain network encoder and a graph classifier, and specifically includes the following steps: (1) The obtained static functional connection matrix is ​​input into the static brain network encoder, and the deep relationship between brain region nodes is learned from the static functional connection matrix through the self-attention mechanism to generate a high-order functional connection matrix; In step (1), the Pearson correlation is used to obtain the static functional connectivity matrix of each subject. After obtaining the static functional connectivity matrix, in order to avoid the problem of insufficient brain fitting, a multi-head self-attention network is selected as the functional connectivity encoder; for each pre-segmented brain region node, the corresponding row vector in the static functional connectivity matrix is ​​used as its feature vector, and it is input into the functional connectivity encoder to obtain the high-order embedding vector of each brain region and vertically splice it to obtain the brain region node feature representation of the subject; after obtaining the brain region node feature representation of the subject, since the input of the subsequent classifier is required to be in the form of a symmetric matrix, a functional connectivity generator is required to generate a high-order functional connectivity matrix; the calculation method is to use the softmax function to encode the learned brain region node features, and multiply it with its transpose to obtain a new symmetric matrix, i.e., a high-order functional connectivity matrix.

[0012] (2) The dynamic functional connectivity matrix obtained by the sliding window algorithm is input into the dynamic brain network encoder. After learning the dynamic features of the brain network using the self-attention mechanism, the channel attention mechanism is used to integrate the information to obtain the high-order dynamic feature embedding of each brain region. In step (2), the dynamic functional connectivity matrix of each subject is obtained by using Pearson correlation. After obtaining the dynamic functional connectivity matrix, in order to improve the learning ability of the model, a multi-head self-attention network is selected as the functional connectivity encoder; for different subsequence connection matrices in the dynamic functional connectivity matrix, a transformer network with shared weights is used to encode them. For each subsequence, the calculation method of the functional connectivity encoder is the same as that of the static connection encoder. After calculating multiple subsequences, a multi-channel brain region node feature representation of the subject is obtained, in which each channel corresponds to a subsequence; after obtaining the multi-channel brain region node feature representation, in order to better apply it to downstream classification tasks, a channel attention network SENet is used to fuse its features; the calculation method is to use the SE block in SENet to generate a corresponding attention score for each channel, then multiply each channel by its score, and then add all the results to obtain the high-order dynamic feature embedding of each brain region.

[0013] (3) Based on the results obtained by the static brain network encoder and the dynamic brain network encoder, the graph classifier uses the graph convolutional network algorithm and the topK pooling algorithm to learn and achieve the classification of autism.

[0014] In step (3), in order to better integrate the calculation results of step (1) and step (2), since the two results have the same number of brain region nodes, they are combined into a high-order functional connection brain map, in which the output matrix of the static brain network encoder is used as the adjacency matrix of the high-order brain map, and the input feature embedding of the dynamic brain network encoder is used as the node feature of the high-order brain map; after obtaining the high-order functional connection brain map, in order to better learn the information in the brain map, a two-layer graph convolutional network is used for learning. The calculation method is to first extract the features of each node and transmit its features to the neighboring nodes of the node, and then each node is based on its own features and the neighborhood of the node. The features are weighted and fused and their own feature representations are updated; after learning through the two-layer graph convolution layer, in order to improve the model computing performance, it will be input into a topK pooling layer, which is calculated by generating a corresponding importance score based on the feature vector of each node, retaining the top K important nodes, discarding the remaining nodes, and reducing the dimension of the features; after dimensionality reduction through the pooling layer, the readout layer is used to obtain the feature representation of the entire image and input it into the multi-layer perceptron for classification. The calculation method is to multiply the feature vector by the weight vector learned by the perceptron, add the bias term, and then perform nonlinear activation through the activation function to finally obtain the classification result of the subject.

[0015] Now, the exemplary embodiments of the present disclosure will be described in detail. In the method embodiment of the present invention, the deep model architecture according to an embodiment of the algorithm proposed by the present invention is as follows: Figure 2 As shown, the functional connection encoder architecture according to an embodiment of the algorithm proposed by the present invention is as follows Figure 3 shown.

[0016] This embodiment includes three parts, namely, a static brain network encoder (Static Brain Network Encoder), a dynamic brain network encoder (Dynamic Brain Network Encoder) and a graph classifier (GraphClassifier). Specifically, the static brain network encoder can effectively extract global brain network features and generate functional connections with richer semantic information; the dynamic brain network encoder can encode each subsequence in the dynamic connection separately and fuse all the encodings according to the attention mechanism; the graph classifier reorganizes the results of the first two modules into a high-order functional connection brain map and classifies it based on the GCN algorithm. This embodiment will be connected to a topK pooling layer after the last graph convolution layer, which has two main functions: 1) Improve the representation ability of the model by removing some noise nodes while retaining nodes that are more important for disease classification, and improve the interpretability of the model by analyzing these retained nodes. 2) After removing some nodes, the dimension of the embedding vector of the individual graph will be greatly reduced, which will help to improve the training speed of the model.

[0017] In order to evaluate the performance of this embodiment, we chose to compare it with multiple advanced autism classification methods on the Autism Brain Imaging Database (ABIDE). The AAL (Anatomical Automatic Labeling, AAL) brain atlas is used to divide the brain into 116 brain regions. Since ABIDE involves multiple acquisition sites, considering that all acquisition protocols of different acquisition sites are different, we uniformly use the first 100 time points in its BOLD sequence for experiments. First, the static functional connection matrix of each subject is obtained using the Pearson correlation, and then the BOLD sequence is segmented using a sliding window with a step size of 10 and a window size of 50, and the dynamic functional connection matrix of each subject is obtained using the Pearson correlation. In the model setting, we set 500 epochs and set the batch size to 16. We use the Adam optimizer for learning, the learning rate is set to 0.0001, the discard rate in the model is set to 0.1, and in the topK pooling layer, we set the pooling rate to 0.7.

[0018] In order to evaluate the performance of the compared models, we chose to use accuracy (ACC), sensitivity (SEN) and specificity (SPE) as evaluation indicators. The results of the comparative test are shown in Table 1.

[0019] Table 1 Comparative test results on AAL brain atlas (percentage)

[0020] From the experimental results, it can be seen that this embodiment achieved the best classification accuracy and the best classification sensitivity in the comparative test, and the specificity was 0.6% lower than the first place, which shows the effectiveness of the embodiment made according to the present invention in the autism classification model.

[0021] By extracting high-order fMRI features and integrating the research methods of dynamic brain networks and static brain networks, the present invention solves the common problems in the benchmark model, and the experimental results fully demonstrate the effectiveness of this brain network modeling method. In this experiment, the average accuracy of this embodiment on representative models such as SVM, DiagNet, and PLSNet was improved by 16.6%, 7.4%, and 2.9%, respectively, which fully demonstrates the superiority of the present invention in autism classification research.

Claims

1. An autism classification algorithm based on the combination of dynamic and static brain networks, characterized by: It includes a static brain network encoder, a dynamic brain network encoder and a graph classifier, and specifically includes the following steps: (1) The obtained static functional connection matrix is ​​input into the static brain network encoder, and the deep relationship between brain region nodes is learned from the static functional connection matrix through the self-attention mechanism to generate a high-order functional connection matrix; (2) The dynamic functional connectivity matrix obtained by the sliding window algorithm is input into the dynamic brain network encoder. After learning the dynamic features of the brain network using the self-attention mechanism, the channel attention mechanism is used to integrate the information to obtain the high-order dynamic feature embedding of each brain region. (3) Based on the results obtained by the static brain network encoder and the dynamic brain network encoder, the graph classifier uses the graph convolutional network algorithm and the topK pooling algorithm to learn and achieve the classification of autism.

2. The autism classification algorithm based on the combination of dynamic and static brain networks according to claim 1 is characterized by: In step (1), the Pearson correlation is used to obtain the static functional connectivity matrix of each subject. After obtaining the static functional connectivity matrix, a multi-head self-attention network is used as a functional connectivity encoder. For each pre-segmented brain region node, the corresponding row vector in the static functional connectivity matrix is ​​used as its feature vector, and it is input into the functional connectivity encoder to obtain the high-order embedding vector of each brain region and vertically concatenate it to obtain the brain region node feature representation of the subject. After obtaining the brain region node feature representation of the subject, a functional connectivity generator is used to generate a high-order functional connectivity matrix. The calculation method is to encode the learned brain region node features using the softmax function and multiply them with their transpose to obtain a new symmetric matrix, namely the high-order functional connectivity matrix.

3. The autism classification algorithm based on the combination of dynamic and static brain networks according to claim 1 is characterized by: In step (2), the dynamic functional connectivity matrix of each subject is obtained by using Pearson correlation. After the dynamic functional connectivity matrix is ​​obtained, a multi-head self-attention network is used as a functional connectivity encoder. For different subsequence connection matrices in the dynamic functional connectivity matrix, a transformer network with shared weights is used to encode them. For each subsequence, the calculation method of the functional connectivity encoder is the same as that of the static connection encoder. After calculating multiple subsequences, a multi-channel brain region node feature representation of the subject is obtained, in which each channel corresponds to a subsequence. After obtaining the multi-channel brain region node feature representation, a channel attention network SENet is used to fuse its features. The calculation method is to use the SE block in SENet to generate a corresponding attention score for each channel, then multiply each channel by its score, and then add all the results to obtain the high-order dynamic feature embedding of each brain region.

4. The autism classification algorithm based on the combination of dynamic and static brain networks according to claim 1 is characterized by: In step (3), since the two results have the same number of brain region nodes, they are combined into a high-order functional connection map, in which the output matrix of the static brain network encoder is used as the adjacency matrix of the high-order brain map, and the input feature embedding of the dynamic brain network encoder is used as the node feature of the high-order brain map; after obtaining the high-order functional connection map, a two-layer graph convolutional network is used for learning. The calculation method is to first extract the features of each node and transmit its features to the neighboring nodes of the node, and then each node performs weighted fusion according to its own features and the features of the neighborhood and updates its own feature representation; after learning through the two-layer graph convolution layer, it is input into a topK pooling layer. The calculation method is to generate a corresponding importance score according to the feature vector of each node, retain the first K important nodes, discard the remaining nodes, and reduce the dimension of the feature; after the dimensionality reduction through the pooling layer, the feature representation of the entire image is obtained by using the readout layer and input into the multi-layer perceptron for classification. The calculation method is to multiply the feature vector by the weight vector learned by the perceptron, add the bias term, and then perform nonlinear activation through the activation function to finally obtain the classification result of the subject.

Images