A time series-based autism spectrum disorder identification method

Abstract

The application belongs to the field of autism spectrum disorder auxiliary diagnosis, and particularly relates to an autism spectrum disorder identification method based on time series, which comprises the following steps: constructing and training an autism identification model, inputting to-be-identified data into the trained autism identification model to obtain an identification result; the autism identification model comprises a graph attention spatial network feature extractor, a bidirectional long short-term network time feature extractor and a multilayer perception network classifier; the application can comprehensively utilize potential graph structure information and time sequence dependence between regions of a brain region template, and has an advantage in capturing important features in time sequence data and complex relationships between nodes.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of auxiliary diagnosis of autism spectrum disorder, and particularly relates to an autism spectrum disorder identification method based on time series. BACKGROUND

[0002] Autism spectrum disorder (ASD) is a serious neurological disease. According to the latest large-scale survey, the prevalence of autism spectrum disorder is about 1-2%. The clinical manifestations mainly include difficulties in social interaction, language communication and repetitive behavior. Due to the overlapping symptoms and the lack of clear biological tests that can be used as clear measurement standards, the diagnosis of autism spectrum disorder is mainly made by professionals using tools such as autism diagnostic observation and autism diagnostic interview. First, a large amount of high-quality medical resources is essential. Second, this test lacks objectivity. Therefore, there is an urgent need for computer-aided diagnosis (CAD) technology to assist medical personnel in auxiliary diagnosis. Neuroimaging technology provides a powerful technical means for understanding brain mechanisms. Resting-state functional magnetic resonance imaging (fMRI) is a commonly used non-invasive neuroimaging technique that can provide better spatial resolution. It can observe the functional changes of brain activity during task execution by measuring the changes in blood oxygen level dependent (BOLD) contrast images.

[0003] In the research of ASD auxiliary diagnosis based on fMRI, there are mainly two research routes: machine learning and deep learning. Due to the lack of data and the limitation of machine performance, early studies on the FC data of a single international neuroimaging laboratory used traditional machine learning techniques, including supervised methods such as SVM, RF, logistic regression, and unsupervised methods such as PCA. In recent years, with the emergence of the ABIDE dataset and the widespread use of end-to-end deep learning applications, there has been a paradigm shift in ASD diagnosis research. Most deep learning methods have been applied to the task of identifying ASD patients using fMRI data. These methods include DNN, CNN, LSTM, etc. In recent years, with the rise of Transformer and GCN, their corresponding applications have also been used to identify ASD patients. Generally, different brain partition templates are used to partition the regions of interest (ROI) of fMRI data, thereby obtaining different ROI average time series. However, existing methods do not fully consider the potential spatial relationship between different ROIs, thereby ignoring the possible spatial features of the time series data of fMRI data.

[0004] To solve these problems, the application proposes an ASD autism spectrum disorder identification method based on a graph attention network of time series, which can effectively extract the spatial-temporal features in the time series data after extracting the time series fMRI data according to the brain partition template. SUMMARY

[0005] To solve the above problems, the application provides a time series-based autism spectrum disorder identification method, which comprises constructing and training an autism identification model, inputting to-be-identified data into the trained autism identification model to obtain an identification result; the autism identification model comprises a graph attention spatial network feature extractor, a bidirectional long short-term network time feature extractor and a multilayer perception network classifier.

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

[0007] S1. Obtain fMRI data and perform preprocessing to obtain preprocessed data;

[0008] S2. Divide the whole brain into 200 brain regions according to the CC200 brain region template, and extract the average time series data of each brain region according to the preprocessed data;

[0009] S3. Calculate the Pearson correlation coefficient between each brain region according to the average time series data, and construct a graph structure according to the Pearson correlation coefficient;

[0010] S4. Input the graph structure into the graph attention spatial network feature extractor to update the node features of each node;

[0011] S5. Input the node features of the graph structure into the bidirectional long short-term network time feature extractor to obtain a global context vector;

[0012] S6. Input the global context vector into the multilayer perception network classifier to obtain an identification result;

[0013] S7. Calculate the loss according to the identification result, train the model parameters, and stop until the model parameters converge.

[0014] The application has the following beneficial effects:

[0015] The application captures the functional connectivity between different brain regions through fMRI data, adopts a graph attention network and a bidirectional LSTM network to obtain the spatial and temporal relationships between different brain regions, realizes the information enhancement between each region under the brain region template partition, and thus improves the accuracy of the model in ASD autism spectrum disorder identification based on fMRI time series data. In addition, the application optimizes the feature extraction capability of the model through the attention mechanism, and increases the generalization capability of the model during classification by combining dropout and the like. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The method flowchart of the application;

[0017] Figure 2 The model structure schematic diagram of the application. DETAILED DESCRIPTION

[0018] With reference to the accompanying drawings: clearly and completely describe the technical solutions in the embodiments of the present application, obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0019] The present application provides a time series-based autism spectrum disorder identification method, comprising constructing and training an autism identification model, inputting to-be-identified data into the trained autism identification model to obtain an identification result; the autism identification model comprises a graph attention spatial network feature extractor, a bidirectional long short-term network time feature extractor and a multilayer perception network classifier.

[0020] As shown in Figure 1 , Figure 2 The training process of the autism identification model comprises:

[0021] S1. Obtain fMRI data and pre-process to obtain pre-processed data.

[0022] Specifically, the present application uses the ABIDE I dataset as experimental data, and the ABIDE I dataset comes from 17 different international brain imaging laboratories, which contains a total of 1112 subjects, of which 539 are ASD patients and 573 are normal.

[0023] The fMRI data of each subject is pre-processed, including:

[0024] Delete the brain slice data of the first 10 time points in the fMRI data to avoid the influence of the subjects not being able to enter the resting state immediately at the beginning of the test;

[0025] According to the form of interlaced scanning, the brain slice data obtained at different time points is time-corrected; all brain slice data is arranged in time sequence, and when interlaced scanning is performed, the brain slice data in the middle of the sequence is generally selected as a reference;

[0026] In order to prevent the interference of the subject's head movement on the fMRI data, rigid transformation is used to fix the brain in all images at the same target position, and then MNI image is selected as a configuration template to reposition the structural image and functional image and remove the scalp.

[0027] Use noise regression and spatial standardization to map the brains of different subjects into the same space, then select 6mmx6mmx6mm to further remove artifacts, and then select low-frequency signals in the range of 0.01-0.08Hz to filter, to remove the influence of high-frequency noise such as breathing and heartbeat.

[0028] S2. Divide the whole brain into 200 brain regions according to the CC200 brain partition template, and extract the average time series data of each brain region according to the preprocessed data.

[0029] S3. Calculate the Pearson correlation coefficient between each brain region according to the average time series data, and construct a graph structure according to the Pearson correlation coefficient.

[0030] Specifically, the specific process of step S3 for constructing a graph structure includes:

[0031] Each brain region is taken as a node, and the average time series data is taken as the node feature. The Pearson correlation coefficient between each two nodes is calculated according to the time series data. If the Pearson correlation coefficient is greater than 0.5, an edge is connected between the two nodes, and the Pearson correlation coefficient is taken as the edge value. If the Pearson correlation coefficient is not greater than 0.5, there is no connecting edge between the two nodes. Thus, the adjacency matrix of the graph structure is obtained.

[0032] S4. Input the graph structure into the graph attention spatial network feature extractor to update the node feature of each node.

[0033] Specifically, the adjacency matrix is normalized to obtain a normalized matrix, denoted as

[0034]

[0035] wherein A represents the adjacency matrix, D represents the degree matrix of the graph structure.

[0036] Specifically, there is a potential spatial relationship between different brain regions. The graph attention spatial network spatial feature extractor not only aggregates the potential information of each brain partition, but also determines the influence of different brain regions on ASD recognition through the attention mechanism that can be learned, and further extracts effective spatial features. The graph attention spatial network feature extractor includes multiple attention graph convolution layers, and step S4 specifically includes

[0037] For each node, the node feature thereof is taken as the initial input, and after passing through multiple attention graph convolution layers, the output of the last attention graph convolution layer is taken as the updated node feature. The processing process of the lth attention graph convolution layer includes:

[0038] S31. Take the output of node i in the (l-1)th attention graph convolution layer as the input of the lth attention graph convolution layer;

[0039] S32. For the input of node i in the lth attention graph convolution layer The neighbor node feature fusion processing is performed by using the attention mechanism to obtain the output of node i in the lth attention graph convolution layer is denoted as

[0040]

[0041] wherein, is the input of node i in the l-th attention graph convolution layer, W (l) is a trainable parameter, is the linear transformation of node i in the l-th attention graph convolution layer, [||] denotes the concatenation process,

[0042] is the weight vector of the l-th attention graph convolution layer, LeakyReLU is an activation function, and N(i) denotes the neighbor node set of node i, is the attention weight between node i and node j, is the updated feature representation of node i in the l-th attention graph convolution layer.

[0043] S5. Input the node features of the graph structure into the bidirectional long short-term network time feature extractor to obtain a global context vector.

[0044] Specifically, step S5 specifically includes:

[0045] S51. Obtain the time comprehensive representation of different time steps through the bidirectional LSTM network from all node features;

[0046] Specifically, each node feature processed by the graph attention spatial network feature extractor contains spatial features of different time steps, and the spatial feature sequence of different time steps is formed by extracting the spatial features corresponding to different time steps from all node features. Assuming that the updated node feature of node i is X i ={x 1i ,x 2i ,…,x Ti}, x ti denotes the spatial feature of node i at time step t, and T denotes the number of time steps; then the spatial feature sequence at time step t is denoted as H t =(x t1 ,x t2 ,…,x tM ), and M denotes the number of nodes. The spatial feature sequences of different time steps are obtained through the bidirectional LSTM network to obtain the time comprehensive representation of different time steps.

[0047] S52. Calculate the attention score of the time comprehensive representation of each time step, denoted as

[0048] e t =tanh(W a h t +b a )

[0049] where e t denotes the attention score at the t-th time step, W a denotes a learnable weight matrix, b a denotes a learnable bias vector, h t denotes a temporal summary representation at the t-th time step;

[0050] S53. Normalizing each attention score to obtain a scalar score, denoted as

[0051]

[0052] where a t denotes the scalar score at the t-th time step, and T denotes the number of time steps;

[0053] S54. Computing a global context vector according to the scalar scores and the temporal summary representation, denoted as

[0054]

[0055] where c denotes the global context vector.

[0056] S6. Inputting the global context vector into a multi-layer perceptron network classifier to obtain a recognition result.

[0057] Specifically, the multi-layer perceptron network classifier includes three hidden layers, wherein the first hidden layer includes a linear transformation layer, a Dropout layer and a batch normalization layer, the second hidden layer includes a linear transformation layer and a ReLU activation function layer, and the third hidden layer includes a linear transformation layer.

[0058] Specifically, the specific process of inputting the global context vector into the multi-layer perceptron network classifier includes:

[0059] S61. Inputting the global context vector into the first hidden layer to obtain a first output, denoted as

[0060]

[0061] where W1 denotes a learnable weight matrix, b1 denotes a learnable bias vector, c denotes the global context vector, BatchNorm(·) denotes a batch normalization layer, Dropout(·) denotes a Dropout layer, denotes the first output;

[0062] S62. Inputting the first output into the second hidden layer to obtain a second output, denoted as

[0063]

[0064] where denotes the second output, ReLU(·) denotes a ReLU activation function layer, W2 denotes a learnable weight matrix, and b2 denotes a learnable bias vector,

[0065] S63. The second output is input into a third hidden layer to obtain a third output, denoted as

[0066]

[0067] wherein z3 denotes the third output, W3 denotes a learnable weight matrix, and b3 denotes a learnable bias vector;

[0068] S64. The third output is normalized by a Sigmoid activation function to obtain a recognition result.

[0069] S7. A loss is calculated according to the recognition result, and model parameters are trained until the model parameters converge.

[0070] Specifically, a binary cross-entropy loss function is used to effectively measure the error between a predicted value and a true label, and the formula is as follows:

[0071]

[0072] wherein N denotes the number of samples, y i denotes a true label value of the i-th sample, denotes a model predicted label value of the i-th sample.

[0073] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "setting", "connecting", "fixing", "rotating" and the like should be understood in a broad sense, for example, can be fixedly connected, or can be detachably connected, or can be integrated; can be mechanically connected, or can be electrically connected; can be directly connected, or can be indirectly connected through an intermediate medium; can be the internal communication of two elements or the interaction relationship between two elements, unless otherwise explicitly limited, the above-mentioned terms in the present application can be understood according to the specific meaning of the above-mentioned terms in the present application by those skilled in the art according to the specific circumstances.

[0074] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A time series based autism spectrum disorder identification method, characterized in that, The method comprises constructing and training an autism recognition model, and inputting to-be-recognized data into the trained autism recognition model to obtain a recognition result; the autism recognition model comprises a graph attention spatial network feature extractor, a bidirectional long short-term network time feature extractor, and a multi-layer perception network classifier; The training process of the autism recognition model comprises: S1. obtaining fMRI data and preprocessing to obtain preprocessed data; S2. dividing the whole brain into 200 brain regions according to the CC200 brain region template, and extracting average time series data of each brain region according to the preprocessed data; S3. calculating the Pearson correlation coefficient between each brain region according to the average time series data, and constructing a graph structure according to the Pearson correlation coefficient; The specific process of step S3 for constructing the graph structure comprises: Taking each brain region as a node and the average time series data as the node feature; calculating the Pearson correlation coefficient between each two nodes according to the time series data, and connecting an edge between the two nodes if the Pearson correlation coefficient is greater than 0.5, and recording the Pearson correlation coefficient as the edge value; if the Pearson correlation coefficient is not greater than 0.5, there is no connecting edge between the two nodes; S4. inputting the graph structure into the graph attention spatial network feature extractor to update the node feature of each node; The graph attention spatial network feature extractor comprises a plurality of attention graph convolution layers, and the processing process of the lth attention graph convolution layer comprises: S31. the output of the (l-1)-th attention map convolutional layer of the node i as the input of the l-th attention map convolutional layer S32. For the input of the i-th node in the l-th attention graph convolution layer , the neighbor node feature fusion processing is performed by adopting the attention mechanism to obtain the output of the i-th node in the l-th attention graph convolution layer ; S5. inputting the node feature of the graph structure into the bidirectional long short-term network time feature extractor to obtain a global context vector; Step S5 specifically comprises: S51. obtaining a time comprehensive representation at different time steps through the bidirectional LSTM network by using all node features; S52. calculating an attention score of the time comprehensive representation at each time step, which is represented as where e t denotes the attention score at the t-th time step, W a denotes a learnable weight matrix, b a denotes a learnable bias vector, h t denotes a temporal summary representation at the t-th time step; S53. normalizing each attention score to obtain a scalar score, which is represented as where α t denotes the scalar score at the t-th time step, and T denotes the number of time steps. S54. calculating the global context vector by weighting according to the scalar score and the time comprehensive representation, which is represented as Wherein, c represents the global context vector; S6. inputting the global context vector into the multi-layer perception network classifier to obtain a recognition result; The multi-layer perception network classifier comprises three hidden layers, wherein the first hidden layer comprises a linear transformation layer, a Dropout layer and a batch normalization layer, the second hidden layer comprises a linear transformation layer and a ReLU activation function layer, and the third hidden layer comprises a linear transformation layer; S7. calculating a loss according to the recognition result, training model parameters, and stopping until the model parameters converge. 2.The time series based autism spectrum disorder identification method of claim 1, wherein, The specific process of inputting the global context vector into the multi-layer perception network classifier comprises: S61. inputting the global context vector into the first hidden layer to obtain a first output, which is represented as wherein W1 represents a learnable weight matrix, b1 represents a learnable bias vector, c represents a global context vector, BatchNorm(·) represents a batch normalization layer, Dropout(·) represents a Dropout layer, represents the first output; S62. inputting the first output into the second hidden layer to obtain a second output, which is represented as wherein, denotes a second output, ReLU( ) denotes a ReLU activation function layer, W2denotes a learnable weight matrix, b2denotes a learnable bias vector, S63. inputting the second output into the third hidden layer to obtain a third output, which is represented as Wherein, z3 represents the third output, W3 represents a learnable weight matrix, and b3 represents a learnable bias vector; S64. normalizing the third output by a Sigmoid activation function to obtain a recognition result. 3.The time series based autism spectrum disorder identification method of claim 1, wherein, The loss is calculated by using a binary cross-entropy loss function, which is represented as where N represents the number of samples, y i represents the true label value of the i-th sample, represents the model predicted label value of the i-th sample.

Images