An arlstm-gcns-based autism diagnosis method

Abstract

The application discloses an autism diagnosis method based on ARLSTM-GCNs, relates to a medical pattern recognition diagnosis method, and discloses an autism spectrum disorder diagnosis method based on improved AR-LSTM and fusion attention graph convolution network, namely, ARLSTM-GCNs, which comprises the following steps: collecting fMRI imaging is divided into a plurality of ROIs through AAL, and BOLD sequence groups of all ROIs are extracted; linear normalization processing is performed on a time-frequency graph generated by wavelet transform of the BOLD sequence, and then two-dimensional sequence data obtained is used to estimate FCNs between each sample ROI by using Pearson correlation; the processed data set is divided into a training set, a verification set and a test set; an ARLSTM-GCNs diagnosis model is built, including an improved AR-LSTM embedding model and a GCN-SAGPool model; and based on the built ARLSTM-GCNs model, a staged training task is performed. The application realizes relatively accurate prediction of autism.

Description

Technical Field

[0001] This invention relates to medical pattern recognition diagnostic methods, specifically to an autism diagnostic method based on ARLSTM-GCNs. Technical Background

[0002] Autism spectrum disorder (ASD) is a serious mental disorder characterized by social impairment, language impairment, and repetitive and stereotyped behaviors, with lifelong effects. The average prevalence of ASD in Chinese children is 1.8%, and it is showing an increasing trend year by year. Internationally, the diagnosis of childhood autism relies on behavioral symptoms, which is time-consuming and prone to misdiagnosis and missed diagnosis. Resting-state functional MRI (rs-fMRI), as an objective neuromarker, has made significant contributions to research on brain function and disease diagnosis. By measuring blood oxygen level-dependent (BOLD) changes in brain structures using rs-fMRI and analyzing BOLD changes and synchronous activity levels in regions of interest (ROIs) of the brain network, key information about brain activity can be revealed. Unlike organic neurological damage, the causes of ASD involve the complex interaction of multiple factors, including genetics and environment. Therefore, the pathological factors shown in imaging vary greatly among different patients, posing a challenge to medical professionals in diagnosis.

[0003] Machine learning is playing an increasingly important role in ASD research. It provides a powerful tool for processing and analyzing complex brain imaging data. fMRI-based diagnosis of brain diseases typically employs two technical approaches: based on Regions of Interest (ROIs) information or Functional Connection Networks (FCNs) information. This single-modality analysis cannot fully capture the complex changes in brain activity in ASD patients. In recent years, researchers have attempted to fuse data from these two modalities to gain a more comprehensive understanding. Among these approaches, Generative Connectivity Networks (GCNs) have provided a new method for fusing and analyzing local structural and functional connectivity features of the brain, achieving great success in the field of brain disease diagnosis. However, in existing studies, the performance of modality fusion classification models using GCNs is generally lower than that of single-modality models. This is because GCNs have complex modeling requirements for time-series data in ROIs, making them prone to learning from noisy time-series data during training, leading to gradient explosion and overfitting problems. Summary of the Invention

[0004] The purpose of this invention is to propose an autism diagnosis method based on ARLSTM-GCNs. This method is based on an improved LSTM and a graph convolutional network fused with self-attention (GCN-SAGPool). It combines ARLSTM-GCNs—an improved adaptive residual long short-term memory network (AR-LSTM)—with a graph convolutional network fused with attention (GCN-SAGPool) for autism spectrum disorder (ASD) diagnosis—and its segmented training technique. This method aims to improve the accuracy and efficiency of autism diagnosis by analyzing patients' multimodal physiological data.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A diagnostic method for autism based on ARLSTM-GCNs, the method specifically includes the following steps:

[0007] 1) Collect and screen the fMRI data required for the experiment;

[0008] 2) The fMRI images were segmented into multiple ROIs using the AAL automated dissection template, and the blood oxygen level dependent BOLD signal sequence group of all ROIs was extracted.

[0009] 3) The BOLD sequence is subjected to wavelet transform using the continuous wavelet transform method. The resulting time-frequency plot is then linearly normalized to obtain the required two-dimensional sequence data. Based on the two-dimensional sequence data, Pearson correlation is used to estimate the functional connectivity (FCN) between each sample's ROI.

[0010] 4) The processed sample ROI groups and their corresponding FCNs are used as datasets, and divided into training set, validation set and test set;

[0011] 5) Construct an ARLSTM-GCNs network algorithm model as the main classification model. This algorithm model consists of a multi-step training method.

[0012] 6) Diagnose ASD in subjects based on classification models.

[0013] The aforementioned autism diagnosis method based on ARLSTM-GCNs involves collecting and screening the fMRI data required for the experiment. Sufficient and accurate data should be collected, and then subjects with similar characteristics should be screened through subject phenotypic information.

[0014] The aforementioned autism diagnosis method based on ARLSTM-GCNs obtains the information required for diagnosis through data preprocessing.

[0015] The correlation coefficients between ROIs were estimated using Pearson correlation analysis, resulting in the correlation coefficient matrix for the corresponding samples. It contains all correlation coefficients between ROIs; for any pair of ROI two-dimensional sequences , Correlation coefficient The calculation is as shown in formula (8); subsequently, the obtained correlation coefficient matrix is... Thresholding is performed to obtain the adjacency matrix. As in formula (9):

[0016]

[0017]

[0018] in: and For sequence , Expectations is the sequence length, where A pair of brain ROI numbers , Therefore, we analyzed the connection information and correlation coefficients between ROIs;

[0019] Apply Morlet continuous wavelet transform to the ROI, based on the scaling factor. Translation factor and Morlet wavelet basis functions Calculate the corresponding sequence Time-frequency information, as shown in formula (10);

[0020] (10)

[0021] The time-frequency graph will be used as a dataset and injected into AR-LSTM for learning.

[0022] The autism diagnosis method based on ARLSTM-GCNs is described above. The method constructs an ARLSTM-GCNs classification algorithm framework, which consists of cascaded AR-LSTM and GCNs networks. The input model data is the ROI wavelet time-frequency sequence group of a batch of samples and the FCNs information of the corresponding samples in this batch. The output is the ARLSTM branch classification result and the GCNs branch classification result of the sample.

[0023] An improved adaptive residual long short-term memory network AR-LSTM model was constructed. The model includes a layer normalization layer, an ARLSTM layer and a self-normalized activation layer. The three layers are cascaded to form a basic AR-LSTM unit. The forward propagation process of the unit is shown in formulas (1)-(4).

[0024]

[0025] in: This is a linearly normalized wavelet time-frequency plot. The time-frequency matrix after layer normalization. , The normalized weight parameters are defined by equations (2) and (3), which define the forward propagation formula and output parameter matrix of the AR-LSTM time unit. Let t be the output hidden vector for time unit t. for The hidden vector at time t is used as the input parameter. Hidden parameters for time t-1 For time t, the state parameter, and Equation (4) represents the weight bias parameters for the ARLSTM adaptive residual connection and the weight bias parameters for the output gate; Equation (4) is the self-normalized activation function. for Activation output, scaling factor negative slope constant ;

[0026] A graph convolutional network model, GCN-SAGPool, integrating graph attention pooling mechanisms, is constructed. The model consists of multiple cascaded units, each including both GCN and SAGPooling algorithms. For the ... The forward propagation formulas for the unit are shown in (5)-(7):

[0027]

[0028] in, Indicates the first Layer node feature groups, For the first The weight matrix of the layer, Add a self-loop adjacency matrix to equation (7). yes The degree matrix, It is a sigmoid activation function; attention score matrix Corresponding node features Each node in, This refers to the weights of the SAGPool layer, where the attention score is calculated using the attention weights. The calculated score is used to retain the higher-scoring pool. Node combination Connect with it This reduces the complexity of the graph.

[0029] The aforementioned autism diagnosis method based on ARLSTM-GCNs, wherein the method is based on a constructed ARLSTM-GCNs model and performs the following... Figure 2 The algorithm is shown in the segmented training method and application method. The method is completed by two-step training: ARLSTM training and GCNs training. The training adopts 5-fold cross-validation. The learning rate change in the ARLSTM training process uses the cosine annealing strategy. In the diagnostic application, ASD diagnosis is realized by forward propagation calculation.

[0030] The advantages and effects of this invention are:

[0031] This invention solves the problem that traditional GCN cannot represent time series, as well as the gradient smoothing problem caused by traditional LSTM after multi-layer cascading. Attached Figure Description

[0032] Figure 1 ARLSTM-GCNs algorithm framework;

[0033] Figure 2 Improved AR-LSTM algorithm structure;

[0034] Figure 3 Model training and diagnostic application methods;

[0035] Figure 4 GCN-SAGPool model structure. Detailed Implementation

[0036] To clearly illustrate the purpose, technical details, and effective applications of this invention, and to facilitate understanding and implementation by those skilled in the art, a further detailed description will be provided below in conjunction with the embodiments and accompanying drawings. Obviously, the embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the invention.

[0037] This invention provides an autism diagnosis method based on ARLSTM-GCNs, including the overall algorithm framework and the improved AR-LSTM algorithm structure. Figure 1 , 2 Model training and diagnostic application methods (see appendix) Figure 3 This includes the following steps:

[0038] Collect and screen the fMRI data required for the experiment;

[0039] fMRI images were segmented into multiple regions of interest (ROIs) using the AAL automated anatomical template, and the BOLD sequence set of all ROIs was extracted.

[0040] The BOLD sequence is subjected to wavelet transform using the continuous wavelet transform method. The generated time-frequency graph is then linearly normalized to obtain the required two-dimensional sequence data. Pearson correlation is then used to estimate the FCN between each sample ROI based on the two-dimensional sequence data. The dataset is divided into training, validation, and test sets, and a 5-fold cross-validation training strategy is used on the training set. An AR-LSTM model is trained, and the model output encoding is classified using a multilayer perceptron. Cosine annealing is used to adjust and optimize the step size, and the cross-entropy loss function is used to evaluate the output loss. After training the AR-LSTM model, the AR-LSTM network parameters are frozen, and the multilayer perceptron is removed. The output time-frequency graph node encoding is generated. The graph node encoding is combined with the FCN data to form graph data, which is then injected into the GCN-SAGPool model for training (GCN-SAGPool model structure attached). Figure 4 The output encoding is classified using a multilayer perceptron, and the loss is evaluated using the cross-entropy loss function. After completing GCN-SAGPool training, the ARLSTM-GCNs training is finished. The trained ARLSTM-GCNs autism diagnostic model is obtained. This model can be used for autism diagnosis and risk screening of suspected patients or subjects. Example

[0041] This invention discloses an autism diagnosis method based on ARLSTM-GCNs networks, which includes the following steps:

[0042] Collect and screen the fMRI data required for the experiment;

[0043] fMRI images were segmented into multiple regions of interest (ROIs) using the AAL automated anatomical template, and the BOLD sequence set of all ROIs was extracted.

[0044] The BOLD sequence is subjected to wavelet transform using the continuous wavelet transform method. The resulting time-frequency plot is then linearly normalized to obtain the required two-dimensional sequence data. Finally, Pearson correlation is used to estimate the FCN between each sample ROI based on the two-dimensional sequence data.

[0045] The processed sample ROI groups and their corresponding FCNs are used as datasets, which are then divided into training, validation and test sets.

[0046] An ARLSTM-GCNs network algorithm model was built as the main classification model. This algorithm model consists of a multi-step training method.

[0047] ASD diagnosis was performed on the subjects based on a classification model.

[0048] The specific implementation steps of the ASD diagnosis method provided by this invention are as follows:

[0049] Step 1:

[0050] The collected data came from the open database ABIDE and included raw fMRI data.

[0051] Step 2:

[0052] The fMRI images were segmented into 116 regions of interest (ROIs) using the AAL automated anatomical template, and the BOLD signal sequence of each ROI was extracted.

[0053] Step 3:

[0054] The correlation coefficients between ROIs were estimated using Pearson correlation analysis, resulting in the correlation coefficient matrix for the corresponding samples. It contains all correlation coefficients between ROIs; for any pair of ROI two-dimensional sequences , Correlation coefficient The calculation is as shown in formula (8); subsequently, the obtained correlation coefficient matrix is... Thresholding is performed to obtain the adjacency matrix. As in formula (9):

[0055]

[0056]

[0057] in: and For sequence , Expectations is the sequence length, where A pair of brain ROI numbers , Therefore, we analyzed the connection information and correlation coefficients between ROIs;

[0058] Apply Morlet continuous wavelet transform to the ROI, based on the scaling factor. Translation factor and Morlet wavelet basis functions Calculate the corresponding sequence Time-frequency information, as shown in formula (10);

[0059] (10)

[0060] The time-frequency graph will be used as a dataset and injected into AR-LSTM for learning.

[0061] Step 4:

[0062] An ARLSTM-GCNs classification algorithm framework is constructed, consisting of a cascaded ARLSTM and GCNs network. The input model data consists of wavelet time-frequency sequences of the ROI of a batch of samples and the FCNs information of the corresponding samples in the same batch. The outputs are the ARLSTM branch classification result and the GCNs branch classification result for that sample.

[0063] An improved adaptive residual long short-term memory network AR-LSTM model was constructed. The model includes a layer normalization layer, an ARLSTM layer and a self-normalized activation layer. The three layers are cascaded to form a basic AR-LSTM unit. The forward propagation process of the unit is shown in formulas (1)-(4).

[0064]

[0065] in: This is a linearly normalized wavelet time-frequency plot. The time-frequency matrix after layer normalization. , The normalized weight parameters are defined by equations (2) and (3), which define the forward propagation formula and output parameter matrix of the AR-LSTM time unit, where... Let t be the output hidden vector for time unit t. for The hidden vector at time t is used as the input parameter. Hidden parameters for time t-1 For time t, the state parameter, and Equation (4) represents the weight bias parameters for the ARLSTM adaptive residual connection and the weight bias parameters for the output gate; Equation (4) is the self-normalized activation function. for Activation output, scaling factor negative slope constant ;

[0066] A graph convolutional network model, GCN-SAGPool, integrating graph attention pooling mechanisms, is constructed. The model consists of multiple cascaded units, each including both GCN and SAGPooling algorithms. For the ... The forward propagation formulas for the unit are shown in (5)-(7):

[0067]

[0068] in, Indicates the first Layer node feature groups, For the first The weight matrix of the layer, Add a self-loop adjacency matrix to equation (7). yes The degree matrix, It is a sigmoid activation function; attention score matrix Corresponding node features Each node in, This refers to the weights of the SAGPool layer, where the attention score is calculated using the attention weights. The calculated score is used to retain the higher-scoring pool. Node combination Connect with it This reduces the complexity of the graph.

[0069] Step 5:

[0070] Based on the completed ARLSTM-GCNs model, a phased training task is performed.

Claims

1. A method for diagnosing autism based on ARLSTM-GCNs, characterized in that, The method specifically includes the following steps: 1) Collect and screen the fMRI data required for the experiment; 2) The fMRI images were segmented into multiple ROIs using the AAL automated dissection template, and the blood oxygen level dependent BOLD signal sequence group of all ROIs was extracted. 3) Perform wavelet transform on the BOLD sequence using the continuous wavelet transform method, and then perform linear normalization on the generated time-frequency graph to obtain the required two-dimensional sequence data; then use Pearson correlation to estimate the functional connectivity (FCN) between each sample ROI based on the two-dimensional sequence data. 4) The processed sample ROI groups and their corresponding FCNs are used as datasets, and divided into training set, validation set and test set; 5) Construct an ARLSTM-GCNs network algorithm model as the main classification model. This algorithm model consists of a multi-step training method. 6) Diagnose ASD in subjects based on classification models; An ARLSTM-GCNs classification algorithm framework is constructed, consisting of cascaded AR-LSTM and GCNs networks. The input model data consists of the wavelet time-frequency sequence of the ROI of a batch of samples and the FCNs information of the corresponding samples in the same batch. The output is the ARLSTM branch classification result and the GCNs branch classification result of the sample. An improved adaptive residual long short-term memory network AR-LSTM model was constructed. The model includes a layer normalization layer, an ARLSTM layer and a self-normalized activation layer. The three layers are cascaded to form a basic AR-LSTM unit. The forward propagation process of the unit is shown in formulas (1)-(4). ； in: This is a linearly normalized wavelet time-frequency plot. The time-frequency matrix after layer normalization. The normalized weight parameters are defined by equations (2) and (3), which define the forward propagation formula and output parameter matrix of the AR-LSTM time unit, where... Let t be the output hidden vector for time unit t. for The hidden vector at time t is used as an input parameter. Hidden parameters for time t-1 For time t, the state parameter, and Equation (4) represents the weight bias parameters for the ARLSTM adaptive residual connection and the weight bias parameters for the output gate; Equation (4) is the self-normalized activation function. for Activation output, scaling factor negative slope constant ; A graph convolutional network model, GCN-SAGPool, integrating graph attention pooling mechanisms, is constructed. The model consists of multiple cascaded units, each including both GCN and SAGPooling algorithms. For the ... The forward propagation formulas for the unit are shown in (5)-(7): ； in, Indicates the first Layer node feature groups, For the first The weight matrix of the layer, Add a self-loop adjacency matrix to equation (7). yes The degree matrix, It is a sigmoid activation function; attention score matrix Corresponding node features Each node in, This refers to the weights of the SAGPool layer, where the attention score is calculated using the attention weights. The calculated score is used to retain the higher-scoring pool. Node combination Connect with it This reduces the complexity of the graph.

2. The autism diagnosis method based on ARLSTM-GCNs according to claim 1, characterized in that, The collection and screening of fMRI data required for the experiment should involve collecting sufficient and accurate data, and then screening subjects with similar characteristics through subject phenotypic information.

3. The autism diagnosis method based on ARLSTM-GCNs according to claim 1, characterized in that, The method obtains the information needed for diagnosis through data preprocessing; The correlation coefficients between ROIs were estimated using Pearson correlation analysis, resulting in the correlation coefficient matrix for the corresponding samples. It contains all correlation coefficients between ROIs; for any pair of ROI two-dimensional sequences , Correlation coefficient The calculation is as shown in formula (8); subsequently, the obtained correlation coefficient matrix is... Thresholding is performed to obtain the adjacency matrix. As in formula (9): ； in: and For sequence , Expectations Sequence length, where A pair of brain ROI serial numbers , Therefore, we analyzed the connection information and correlation coefficients between ROIs; Apply Morlet continuous wavelet transform to the ROI, based on the scaling factor. Translation factor and Morlet wavelet basis functions Calculate the corresponding sequence Time-frequency information, as shown in formula (10); ； This time-frequency graph will be used as a dataset and injected into AR-LSTM for learning.

4. The autism diagnosis method based on ARLSTM-GCNs according to claim 1, characterized in that, The method is based on the constructed ARLSTM-GCNs model and executes a segmented training and application method. The method completes training in two steps: ARLSTM training and GCNs training. The training adopts 5-fold cross-validation. The learning rate change in the ARLSTM training process uses a cosine annealing strategy, and ASD diagnosis is achieved through forward propagation calculation in the diagnostic application.

Images