An autism recognition method based on functional partition and graph convolutional neural network

Abstract

The application belongs to the field of autism recognition, and particularly relates to an autism recognition method based on function partition and graph convolutional neural network; comprising: acquiring autism brain imaging data, partitioning the brain by using DMN mode, and constructing a local function adjacency matrix and a global dynamic adjacency matrix according to the partition; processing the autism brain imaging data by using a space-time feature extractor to obtain space-time features; sending the space-time features into a frequency feature extractor to obtain frequency features; processing the frequency features and the local function adjacency matrix by using a local function feature extractor to obtain local function features; processing the frequency features and the global dynamic adjacency matrix by using a global dynamic feature extractor to obtain global dynamic features; fusing the local function features and the global dynamic features; inputting the fused features into a classifier to obtain an autism recognition result; and the application reduces data differences between different autism patients and improves recognition effect.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of autism recognition, and particularly relates to an autism recognition method based on functional partition and graph convolutional neural network. BACKGROUND

[0002] Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by difficulties in social interaction, language and nonverbal communication, as well as narrow interests, repetitive behaviors and behavioral stereotypy. However, because of the overlapping nature of symptoms and the lack of biological tests that can serve as a clear and quantitative gold standard, defining and diagnosing ASD is a very difficult process. So far, the diagnosis of ASD patients is basically based on the behavioral description of symptomatology, and the diagnosis result is obtained through the joint consultation of experienced doctors and the families of ASD patients. However, during the consultation of doctors, personal observation and subjective decision are prone to misdiagnosis or overdiagnosis of mild cases. Therefore, it is urgent to implement effective computer-aided diagnosis technology to help doctors diagnose ASD.

[0003] The prior art usually uses traditional machine learning methods such as K-Nearest Neighbors (k-NN) and Artificial Neural Networks (ANN) to recognize autism. However, traditional machine learning methods usually rely on manual selection and preprocessing of features, and the performance of ASD varies due to individual differences, task types, environmental influences and other factors. Therefore, the generalization ability of traditional machine learning methods under different sample sets or different task conditions may be poor. The interpretability of traditional methods is relatively poor.

[0004] In order to solve the above problems, researchers have introduced deep learning and graph network technology in recent years to improve the recognition ability of ASD autism. Deep learning models can automatically extract high-dimensional feature representations and capture complex temporal and spatial relationships in fmri signals. And the graph network is naturally suitable for the unstructured data of the brain. Therefore, it is natural to use graph neural networks for research. However, existing methods only focus on single graph structure when extracting features, ignoring the potential features of local function and global dynamic changes. SUMMARY

[0005] In view of the deficiencies of the prior art, the application provides an autism recognition method based on functional partition and graph convolutional neural network, which comprises: acquiring brain imaging data and inputting it into a trained autism recognition model to obtain an autism recognition result.

[0006] The training process of the autism recognition model comprises:

[0007] S1: autism brain imaging data is acquired, the brain is partitioned using DMN mode, and a local functional adjacency matrix and a global dynamic adjacency matrix are constructed according to the partition;

[0008] S2: the autism brain imaging data is processed using a space-time feature extractor to obtain space-time features;

[0009] S3: the space-time features are input into a frequency feature extractor to obtain frequency features;

[0010] S4: the frequency features and the local functional adjacency matrix are processed using a local functional feature extractor to obtain local functional features;

[0011] S5: the frequency features and the global dynamic adjacency matrix are processed using a global dynamic feature extractor to obtain global dynamic features;

[0012] S6: the local functional features and the global dynamic features are fused to obtain fused features;

[0013] S7: the fused features are input into a classifier to obtain an autism recognition result;

[0014] S8: the total loss of the model is calculated, and the model parameters are adjusted according to the total loss of the model, and when the pre-set training times are reached, the model training is completed.

[0015] Preferably, the space-time feature extractor includes a channel attention layer and a spatial attention layer;

[0016] The calculation formula of the channel attention layer is:

[0017] T (x) =σ(W1(W0(x ag ))+W1(W0(x max )))

[0018] The calculation formula of the spatial attention layer is:

[0019] S (x) =σ(f 7*7 ([x ag :x max ]))

[0020] Wherein, T (x) represents the output of the channel attention layer, S (x) represents the output of the spatial attention layer, sigma represents the sigmoid activation function, W1 and W0 represent the first and second training weight matrices, x ag and x max represent average pooling and maximum pooling respectively, and f 7*7 represents convolution with a convolution kernel size of 7*7.

[0021] Preferably, the frequency feature extractor is a wavelet convolution layer, which utilizes a convolution recursive operation to simulate a wavelet function to extract frequency features from spatiotemporal features to obtain frequency features.

[0022] Preferably, the local function feature extractor includes a graph convolution layer, a RELU activation layer and a Dropout layer.

[0023] Preferably, the global dynamic feature extractor includes a graph dynamic convolution layer, a RELU activation layer and a Dropout layer.

[0024] Preferably, the process of fusing the local functional features and the global dynamic features includes concatenating the local functional features and the global dynamic features to obtain fused features.

[0025] Preferably, the process of classifier fusion features includes:

[0026] The fused features are input into the first fully connected layer for dimensionality reduction to obtain the dimensionality reduction result; the dimensionality reduction result is input into the RELU activation layer to obtain the intermediate features; the intermediate features are input into the second fully connected layer for dimensionality reduction to obtain autism-related features; the autism-related features are input into the Sofmax layer for normalization to obtain the autism recognition results.

[0027] Preferably, the formula for calculating the total loss of the model is:

[0028]

[0029] Among them, L cls Represents the total loss of the model, N t Indicates the number of training samples, C indicates the total number of classification categories, represents the true label of the i-th sample, Represents the predicted value of the i-th sample.

[0030] The present invention has the following beneficial effects: It uses the DMN model to partition brain function, deriving a general distribution pattern of human brain function, ensuring the independence of individual characteristics. By combining frequency features with graph neural networks, it identifies the different brain activity states between autistic patients and healthy individuals, improving the model's generalization and recognition accuracy, enabling doctors to make more accurate and scientific diagnoses of autistic patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of the structure of the autism recognition model in the present invention. DETAILED DESCRIPTION

[0032] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of the present application.

[0033] The present application provides an autism recognition method based on functional partition and graph convolutional neural network, which comprises the following contents:

[0034] The brain imaging data is acquired and input into the trained autism recognition model to obtain the autism recognition result.

[0035] As shown in Figure 1 The training process of the autism recognition model comprises:

[0036] S1: Acquire autism brain imaging data, partition the brain using the DMN mode, and construct a local functional adjacency matrix and a global dynamic adjacency matrix according to the partition.

[0037] The autism brain imaging data is acquired, and preferably, the ABIDE 1 public data set can be used as training data. The ABIDE 1 data set contains data from 17 research institutions, about 1112 participants, including 539 autism patients (ASD) and 573 healthy controls (TD). All subjects are divided according to the leave-one-out method, and all data are cross-validated to obtain a training set, a validation set and a test set.

[0038] The data set is collected in the cc200 mode, and there are 200 regions. The corresponding regions on the data set are regarded as a functional area by the DMN partition mode. A functional area only has internal connection relationship, thereby constructing a local functional adjacency matrix. At the same time, the global dynamic graph structure represents the connection between different functional areas. The internal connection of the graph structure only exists between different functional areas, thereby constructing a global dynamic adjacency matrix.

[0039] S2: Process the autism brain imaging data using a space-time feature extractor to obtain space-time features.

[0040] The autism brain imaging data is a time series data and has not been processed. Therefore, the original needs to be processed accordingly. In addition, the relevant time feature information and preliminary spatial information in the original signal need to be obtained. Therefore, the present application uses a space-time feature extractor to input the original signal into a space-time attention layer to achieve the above two purposes, so as to ensure that more accurate feature information can be obtained in the subsequent frequency feature extraction.

[0041] The spatio-temporal feature extractor comprises a channel attention layer and a spatial attention layer; a calculation formula of the channel attention layer is:

[0042] T (x) = σ(W1(W0(x ag ))+W1(W0(x max )))

[0043] A calculation formula of the spatial attention layer is:

[0044] S (x) = σ(f 7*7 ([x ag :x max ]))

[0045] Wherein, T (x) represents an output of the channel attention layer, S (x) represents an output of the spatial attention layer, sigma represents a sigmoid activation function, W1 and W0 represent first and second training weight matrices, x ag and x max respectively represent average pooling and maximum pooling, f 7*7 represents a convolution with a kernel size of 7*7.

[0046] S3: The spatio-temporal feature is sent to a frequency feature extractor to obtain frequency features.

[0047] The frequency feature information is more suitable for reflecting the motor state of the brain. Therefore, the spatio-temporal feature data obtained by the spatio-temporal attention layer is sent to the frequency feature extractor for feature extraction. The extractor mainly comprises a wavelet convolution layer, which is a continuous recursive convolution layer network used to replace the wavelet convolution algorithm. After the layer, frequency features of five frequency bands of delta (1-3HZ), theta (4-7HZ), alpha (8-13HZ), beta (14-30HZ) and gamma (31-50HZ) are extracted.

[0048] S4: A local functional feature extractor is used to process the frequency features and the local functional adjacency matrix to obtain local functional features.

[0049] The extracted frequency features and the local functional adjacency matrix are sent to the local functional feature extractor to ensure that the influence information of different functional areas on autism is correctly extracted. The local functional feature extractor comprises a graph convolution layer, a RELU activation layer and a Dropout layer. The graph convolution layer uses the local functional adjacency matrix to perform a message aggregation operation on the output of the frequency feature extractor, and then performs a node information dimension reduction operation; the Dropout layer is used to process the dimension-reduced data to avoid the problems of gradient explosion and gradient disappearance; the RELU activation layer is used to enhance the non-linear relationship of the model. The calculation process of the graph convolution is as follows:

[0050] First, the local functional adjacency matrix is used to calculate the Laplacian matrix L:

[0051] L = D - D -1 / 2 (I n +A)D -1 / 2

[0052] where D represents the degree matrix, I n represents the identity matrix, and A is the local functional adjacency matrix. Using the Laplacian matrix for graph convolution operation involves eigenvalue decomposition, which is computationally intensive. Therefore, Chebyshev polynomials are used to approximate the solution:

[0053] The Chebyshev convolution operation is as follows:

[0054]

[0055] where σ represents the Sigmoid activation function, H (l) is the feature matrix of the l-th layer, T k (L) is the approximation of the Chebyshev polynomial, W k is the weight matrix of the k-th polynomial coefficient, and K is the polynomial coefficient, which controls the accuracy of the approximation.

[0056] S5: The global dynamic feature extractor is used to process the frequency features and the global dynamic adjacency matrix to obtain the global dynamic features.

[0057] The global dynamic adjacency matrix is set as a trainable weight matrix with random initial values, and the adjacency matrix is updated according to the backpropagation algorithm, thereby dynamically adjusting the connection relationship between the functional areas. In this way, the global dynamic adjacency matrix can change according to the change of feature information, thereby ensuring the correct reflection of the global dynamics between the different functional areas of the brain of autistic patients and normal people. The global dynamic feature extractor includes a graph dynamic convolution layer, a RELU activation layer, and a Dropout layer. The graph dynamic convolution layer uses the global dynamic adjacency matrix to perform message aggregation on the output of the frequency feature extractor and then performs node information dimensionality reduction. The Dropout layer is used to process the dimensionality-reduced data to avoid the problems of gradient explosion and gradient disappearance. The RELU activation layer is used to enhance the non-linear relationship of the model. The graph dynamic convolution process is the same as the graph convolution process.

[0058] S6: The local functional features and global dynamic features are fused to obtain the fused features.

[0059] Concate feature fusion of local functional information and global dynamic information can comprehensively reflect the brain information difference between autism patients and normal people, and increase the performance of the model.

[0060] S7: input the fusion features into the classifier to obtain autism recognition results.

[0061] The process of the classifier on the fusion features includes: inputting the fusion features into a first full connection layer to reduce dimension to obtain a reduced dimension result; inputting the reduced dimension result into a RELU activation layer to obtain intermediate features; inputting the intermediate features into a second full connection layer to reduce dimension to obtain autism-related features, and inputting the autism-related features into a Sofmax layer to normalize to obtain a probability value of whether suffering from autism, i.e., the autism recognition result.

[0062] S8: calculate the total loss of the model and adjust the model parameters according to the total loss of the model, and when the pre-set training times are reached, the model training is completed.

[0063] The probability value output by the model is classified to obtain a predicted label, and the cross entropy of the real category is calculated to obtain a loss value, and the model parameters are updated by back propagation; the classification loss, i.e., the total loss of the model, is:

[0064]

[0065] wherein, L cls represents the total loss of the model, N t represents the number of training samples, and C represents the total number of categories, represents the real label of the i-th sample as a class, represents the predicted value of the i-th sample as a class.

[0066] In summary, the present application partitions the brain by DMN mode, constructs local function and global dynamic graph structure; inputs the original data into a space-time attention extractor to obtain preliminary space-time features; inputs the space-time features into a wavelet convolution layer to obtain different frequency information; inputs the frequency information, local function and global dynamic graph structure into a graph convolution structure respectively to obtain local feature information and global dynamic feature information of the brain; and fuses and classifies the obtained features. The present application reduces the data difference between different autism patients by combining the brain partition mode with the graph neural network, and can well improve the recognition effect.

[0067] The above examples further illustrate the objects, technical solutions and advantages of the present application. It should be understood that the above examples are only preferred embodiments of the present application and are not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made to the present application within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for autism recognition based on functional partition and graph convolutional neural network, characterized in that, The method comprises the following steps: Obtaining brain imaging data and inputting the brain imaging data into a trained autism recognition model to obtain an autism recognition result; The training process of the autism recognition model comprises the following steps: S1: obtaining autism brain imaging data, partitioning the brain using a DMN mode, and constructing a local functional adjacency matrix and a global dynamic adjacency matrix according to the partitioning; S2: processing the autism brain imaging data using a spatio-temporal feature extractor to obtain spatio-temporal features; the spatio-temporal feature extractor comprises a channel attention layer and a spatial attention layer; The calculation formula of the channel attention layer is: ； The calculation formula of the spatial attention layer is: ； wherein, denotes the channel attention layer output, denotes the spatial attention layer output, denotes a sigmoid activation function, and denote first and second training weight matrices, and denote average and max pooling, respectively, denotes a convolution with a kernel size of . S3: inputting the spatio-temporal features into a frequency feature extractor to obtain frequency features; the frequency feature extractor is a wavelet convolution layer, and the wavelet convolution layer uses a convolution recursive operation to simulate a wavelet function to perform frequency feature extraction on the spatio-temporal features to obtain the frequency features; S4: processing the frequency features and the local functional adjacency matrix using a local functional feature extractor to obtain local functional features; the local functional feature extractor comprises a graph convolution layer, a RELU activation layer and a Dropout layer; S5: processing the frequency features and the global dynamic adjacency matrix using a global dynamic feature extractor to obtain global dynamic features; the global dynamic feature extractor comprises a graph dynamic convolution layer, a RELU activation layer and a Dropout layer; S6: fusing the local functional features and the global dynamic features to obtain fused features; S7: inputting the fused features into a classifier to obtain an autism recognition result; inputting the fused features into the classifier comprises the following steps: inputting the fused features into a first fully connected layer to reduce dimensions to obtain reduced dimension results; inputting the reduced dimension results into a RELU activation layer to obtain intermediate features; inputting the intermediate features into a second fully connected layer to reduce dimensions to obtain autism-related features; inputting the autism-related features into a Sofmax layer to normalize to obtain the autism recognition result; S8: calculating a total loss of the model and adjusting model parameters according to the total loss of the model, and completing the model training when a preset number of training times is reached.

2. The autism recognition method based on functional partition and graph convolutional neural network according to claim 1, characterized in that, The process of fusing the local functional features and the global dynamic features comprises splicing the local functional features and the global dynamic features to obtain the fused features. 3.The autism recognition method based on functional partition and graph convolutional neural network according to claim 1, characterized in that, The formula for calculating the total loss of the model is: ； wherein, denotes the total loss of the model, denotes the number of training samples, denotes the total number of classes, denotes the true label of the i-th sample, denotes the predicted value of the i-th sample.

Images