Application of graph neural network based on arbitrary value adjacent matrix in ADHD classification

Abstract

The invention discloses an application of a graph neural network based on an arbitrary value adjacent matrix in ADHD (attention deficit hyperactivity disorder) classification, which comprises the following steps of: 1, acquiring brain region biological signals and brain region biological signals, taking the brain region biological signals as node features, and taking the brain region biological signals as node features; the brain interval biological signals serve as any numerical value adjacency matrix to represent the excitation or inhibition relation of the brain interval; 2, in order to reduce the complexity of the graph, the first K edges of the adjacent matrix are reserved, and sparse processing of the graph is realized; 3, the node features and the processed adjacent matrix serve as input of a graph convolution layer, a degree matrix is obtained by calculating the absolute value of the adjacent matrix, and normalization operation of the adjacent matrix of any numerical value is achieved; and 4, performing image convolution operation by using the normalized adjacent matrix. The method is used for the image convolution operation of which the adjacent matrix is any numerical value, and is used for modeling and processing of a complex brain network structure.

Description

Technical Field

[0001] The present invention relates to an application of a graph neural network based on an arbitrary value adjacency matrix in ADHD (attention deficit hyperactivity disorder) classification, belonging to the field of intelligent auxiliary classification diagnosis of mental illness. Background Art

[0002] With the rapid development of deep learning technology, graph neural networks have shown great potential in processing data with complex network structures. In the biomedical field, especially in neuroscience research, the connection relationship between brain regions can be naturally represented as a graph structure, where brain regions are nodes and biological signals between brain regions are edges. In graph theory, the adjacency matrix of a graph has a non-negative property, and this method performs well on problems with simple structures and limited categories. However, when processing complex data such as brain functional connectivity, non-negative adjacency matrices may not fully express the mutual excitation or inhibition relationship between brain regions. This limits their application in processing biological network data. ADHD is a common neurodevelopmental disorder with diverse clinical manifestations and unknown brain regions. In recent years, research based on brain region biological signals and brain region biological signals has provided new ideas for the classification of ADHD. However, how to effectively use these biological signal data, especially how to build and train graph neural network models that can process arbitrary numerical adjacency matrices, remains a challenging problem. Summary of the invention

[0003] In view of the problem that existing graph neural networks can only process non-negative adjacency matrices, the present invention provides an improved graph neural network technology, the key of which is to allow the adjacency matrix of the graph to be any real value, so as to more accurately express complex brain functional connection patterns. Specifically, we apply this technology to ADHD classification, where the nodes of the graph represent the amplitude of low-frequency fluctuations (ALFF) data of the intensity of regional brain activity, and the adjacency matrix of the graph directly uses the original value of functional connectivity (FC).

[0004] The brain topology network construction method of the present invention can efficiently utilize a variety of biological data calculated from magnetic resonance imaging (MRI) acquisition data, including characteristic data on brain regions, such as brain volume density (VBM), low-frequency amplitude wave (ALFF), homomorphic consistency (Reho), degree centrality (DC), characteristic data between brain regions, such as brain functional connectivity (FC), effector connectivity, etc., to describe and analyze the patient's brain function state from multiple angles. Due to the use of multiple data, through the invention, a graph network close to the real brain network can be constructed, making full use of multimodal information for subsequent disease classification and biomarker detection.

[0005] Traditional graph neural networks usually use non-negative adjacency matrices to describe the connection relationship between nodes, while the present invention introduces the concept of arbitrary-valued adjacency matrices. The present invention allows the elements of the adjacency matrix to be arbitrary real values, thereby more accurately expressing complex brain functional connection patterns. Among them, when the edge is positive, it can simulate the mutual stimulation and excitation between brain regions; when the edge is negative, it can simulate the mutual inhibition between brain regions. Compared with traditional methods, arbitrary-valued adjacency matrices can more carefully reflect the strength and diversity of connections between nodes. In order to reduce computational complexity, the present invention adopts an effective adjacency matrix sparse processing method. By retaining the top K edges with the highest weight, the structure of the graph is effectively simplified, while ensuring that key connection information is not lost. The present invention proposes a new adjacency matrix normalization method. By calculating the absolute value of the adjacency matrix and the corresponding degree matrix, effective normalization of the arbitrary-valued adjacency matrix is ​​achieved. This step not only retains the characteristics of the graph structure, but also effectively enhances the ability of the neural network to capture complex relationships between node features. This method not only broadens the application scope of graph neural networks, enabling them to process adjacency matrices containing arbitrary values, but also provides a new perspective for the structural analysis and feature extraction of complex brain networks.

[0006] The technical solution of the present invention is as follows:

[0007] A method for applying a graph neural network based on an arbitrary numerical adjacency matrix in ADHD classification includes the following steps:

[0008] Step 1: Obtain brain region biological signals and inter-brain biological signals, use brain region biological signals as signals on brain topology network nodes, i.e., node features, and use inter-brain biological signals as signals on brain topology network edge connections, i.e., adjacency matrix, where the elements of the adjacency matrix are arbitrary real values;

[0009] Step 2: Sparse the adjacency matrix and retain the first K edges to ensure the connectivity of the graph;

[0010] Step 3: Use the absolute value of the adjacency matrix to calculate the degree matrix and obtain the normalized adjacency matrix A norm .

[0011] In the above step 2, the adjacency matrix is ​​sparsely processed, and the first k% of the absolute value of the adjacency matrix is ​​retained. The larger the value, the stronger the correlation between brain regions. This method achieves graph sparsification while ensuring graph connectivity.

[0012] In the above step 3, the adjacency matrix is ​​normalized, including the following processing steps:

[0013] Step 1: For a given adjacency matrix A, each element a of A ijRepresents the connection weight between node i and node j (i.e., the biological signal strength between brain regions). In order to consider the characteristic information of the node itself, self-loops are added to the diagonal of the adjacency matrix; the weight of the self-loop can be set to a fixed constant, which represents the connection strength between the node and itself. Next, the degree of each node is calculated, that is, the degree d of node i is the sum of the absolute values ​​of all elements in the i-th row of the adjacency matrix A; then, these degree values ​​are combined into a diagonal matrix D, where the elements on the diagonal are the degree values ​​of the corresponding nodes, and the remaining elements are 0;

[0014] Step 2: Use the calculated degree matrix D to normalize the adjacency matrix A to obtain the normalized adjacency matrix A norm ; The normalization operation is usually performed by adding each element a of the adjacency matrix A ij Divide by the degree d of node i and node j i and d j The square root of A is used to complete the norm =D -0.5 AD -0.5 ; This normalization method is called symmetric normalization, which helps reduce the impact of node degree on graph convolution operations, thereby improving the performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is an overall framework of the application of a graph neural network based on an arbitrary valued adjacency matrix in ADHD classification of the present invention;

[0016] Figure 2 It is the graph network model of the present invention; DETAILED DESCRIPTION

[0017] In conjunction with the accompanying drawings, the application of a graph neural network based on an arbitrary numerical adjacency matrix in ADHD classification of the present invention is further described below.

[0018] like Figure 1 , Figure 2 As shown, an application of a graph neural network based on an arbitrary numerical adjacency matrix in ADHD classification includes the following steps:

[0019] Step 1: Obtain brain region biological signals and inter-brain biological signals, use the brain region biological signal ALFF as the signal on the brain topology network node, that is, the node feature, and use the inter-brain biological signal FC as the signal on the brain topology network edge connection, that is, the adjacency matrix, where the elements of the adjacency matrix are arbitrary real values;

[0020] Step 2: Sparse the adjacency matrix, retain the first 10% of the absolute value of the adjacency matrix, and record the rest as 0; this sparse method retains important positive and negative connections and reduces the impact of noise connections on classification;

[0021] Step 3: Feature selection, such as Figure 1 As shown, the test data is labeled as healthy people (HC) or ADHD patients, and is put into the feature selection module together with the training data. The node features are input into the feature selector (the feature selector is an existing selector, such as support vector machine recursive feature elimination (SVM-RFE) feature selector, Lasso Regression feature selector, etc.), and the contribution of the node features to the target mental illness classification is obtained and output in the form of weights. The weight values ​​of all node features are sorted from large to small, and the M node features with the smallest weight values ​​are deleted, and the remaining N node features are retained;

[0022] Step 4: Design of graph convolutional layer based on arbitrary numerical adjacency matrix;

[0023] Step 4-1: Use the absolute value of the adjacency matrix to calculate the degree matrix and obtain the normalized adjacency matrix A norm ;

[0024] Step 4-2: For a given adjacency matrix A, each element a of A ij represents the connection weight between node i and node j (i.e., the biological signal strength between brain regions). Self-loops are added to the diagonal of the adjacency matrix; the weight of the self-loop is set to 1, indicating the connection strength between the node and itself. Subsequently, the degree of each node is calculated, i.e., the degree of node i, d i is the sum of the absolute values ​​of all elements in the i-th row of the adjacency matrix A; then, these degree values ​​are placed on the diagonal of a matrix to form a diagonal matrix D, where the elements on the diagonal are the degree values ​​of the corresponding nodes, and the remaining elements are 0;

[0025] Step 4-3: Use the calculated degree matrix D to normalize the adjacency matrix A to obtain the normalized adjacency matrix A norm ; The normalization operation is to multiply the adjacency matrix A on the left and right by the square root of the diagonal matrix D, denoted by A norm =D - 0.5 AD -0.5 ;

[0026] Step 4-4: Use the node features and the normalized adjacency matrix as the input of the graph convolution layer to perform graph convolution operation;

[0027] Step 5: Feature extraction: The typical brain node features of the training data and the test data under different test data label assumptions are reduced in dimension through the feature extraction module to obtain the high-level features of the training data and the test data; Figure 2As shown, feature extraction is divided into two parts: encoding network, decoding network and classification network; both the encoding network and the decoding network here use the graph convolution layer designed by the method of the present invention; the encoding network is used to obtain high-level features, the decoding network requires that the high-level features can reconstruct the typical brain node features input by this module, and the classification network requires that the labels of training data and the hypothetical labels of test data can be identified from the high-level features; the purpose of the above operation is to enhance the feature clustering ability of the high-level features under the labels of training data and the hypothetical labels of test data;

[0028] Step 6: ADHD decision, calculate the variability scores of the high-level features of the training data and test data under different test data label assumptions (i.e., the ratio of intra-class divergence to inter-class divergence), compare the clustering performance, and select the hypothesis corresponding to the smaller score as the correct hypothesis; the network design form of step 5 and the decision method of step 6 can be referred to the detailed description of ADHD classification using auto-encoding neural network and binary hypothesis testing.

[0029] Embodiment 1:

[0030] 1) Experimental conditions

[0031] In this experiment, the resting fMRI blood oxygen concentration level (BOLD) data provided by the ADHD-200 public database was used and matched to the brain template (AAL-116 template) to generate BOLD signals based on brain regions. Subsequently, the BOLD signals between brain regions were subjected to time series filtering and Pearson correlation operations to obtain the low-frequency fluctuation amplitudes of 90 brain regions and the functional connectivity data between brain regions, and ADHD classification diagnosis was performed based on the graph network.

[0032] The graph convolution layers in the graph network model of the present invention all use the graph convolution process described in the present invention to extract more discriminative high-level features.

[0033] 2) Experimental content is based on Figure 2According to the specific experimental parameter settings in Implementation Example 1, classification experiments were carried out on data from multiple sites of the ADHD-200 database, and the classification accuracy was evaluated, as shown in Table 1. As can be seen from Table 1, the average site accuracy can reach 95.43%, indicating that the graph network method can effectively classify and diagnose ADHD in individual samples, and also provide a reliable guarantee for the identification of ADHD biomarkers. Due to the use of arbitrary value adjacency matrices, the graph neural network of the present invention performs well in processing complex brain functional connection data. Our technology not only improves the classification accuracy, but also provides a more flexible and sophisticated brain connection model. Therefore, the present invention provides an innovative graph neural network technology, which significantly improves the accuracy and applicability in the classification of neurological diseases such as ADHD by using arbitrary value adjacency matrices. This technology is expected to have a profound impact in the fields of clinical diagnosis and neuroscience research.

[0034] Table 1 Classification performance evaluation of the present invention on multiple site data of ADHD-200 database

[0035]

[0036] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An application of a graph neural network based on arbitrary numerical adjacency matrix in ADHD (attention deficit hyperactivity disorder) classification, characterized in that: The steps include: Step 1: Obtain brain region biological signals and inter-brain biological signals, use brain region biological signals as signals on brain topology network nodes, i.e., node features, and use inter-brain biological signals as signals on brain topology network edge connections, i.e., adjacency matrix, where the elements of the adjacency matrix are arbitrary real values; Step 2: Sparse the adjacency matrix and retain the first K edges to ensure the connectivity of the graph; Step 3: Use the absolute value of the adjacency matrix to calculate the degree matrix and obtain the normalized adjacency matrix A norm .

2. According to claim 1, the application of a graph neural network based on an arbitrary value adjacency matrix in ADHD classification, wherein in step 2, it is characterized in that The adjacency matrix is ​​sparsely processed, retaining the first k% of the absolute value of the adjacency matrix, a total of K edges, to form a sparse adjacency matrix.

3. According to claim 1, the application of a graph neural network based on an arbitrary value adjacency matrix in ADHD classification, wherein in step 3, it is characterized in that Adjacency matrix normalization includes the following processing steps: Step 1: For a sparse adjacency matrix A, where element a in A ij represents the connection weight between node i and node j (i.e., the biological signal strength between brain regions). Self-loops are added to the diagonal of the adjacency matrix; the weight of the self-loop can be set to a fixed constant, which represents the connection strength between the node and itself. Subsequently, the degree of each node is calculated, i.e., the degree of node i, d i is the sum of the absolute values ​​of all elements in the i-th row of the adjacency matrix A; then, these degree values ​​are placed on the diagonal of a matrix to form a diagonal matrix D, which is defined as the degree matrix, where the elements on the diagonal are the degree values ​​of the corresponding nodes, and the remaining elements are 0; Step 2: Use the calculated degree matrix D to normalize the adjacency matrix A to obtain the normalized adjacency matrix A norm ; The normalization operation is to multiply the adjacency matrix A on the left and right by the square root of the degree matrix D, denoted by A norm =D -0.5 AD -0.5 .

Images