FMRI intelligent analysis method based on graph coarsening and three-dimensional dynamic space-time diagram attention network

Abstract

The invention discloses an fMRI intelligent analysis method based on graph coarsening and a three-dimensional dynamic space-time diagram attention network. The fMRI intelligent analysis method is used for accurately and efficiently recognizing sleep disorders. The method comprises the following steps: firstly, preprocessing an acquired human brain magnetic resonance image, mapping the human brain magnetic resonance image to a standard brain template for region division, averaging time sequence signals of each region of interest (ROI), and extracting a blood oxygen level dependence (BOLD) signal; and then, constructing a dynamic effective function connection matrix by utilizing Granger causal analysis, performing graph coarsening through a Gaussian mixture model, generating a simplified effective function connection matrix, and inputting the matrix into a constructed space-time graph convolutional neural network (ST-GCN) to perform deep feature extraction. Besides, in order to enhance the attention to a key space-time region in a dynamic effective connection matrix and effectively capture the dynamic change of a brain interval at a specific time point, a space and channel mixed attention mechanism module (CBAM) is introduced into the space-time diagram convolutional neural network. The module improves the discrimination capability of the model by adaptively distributing space and channel attention weights. And finally, the network maps the extracted spatial-temporal characteristics to a classification label space through a full connection layer (FC), and automatic identification of sleep disorders is realized. The method can be effectively applied to diagnosis of sleep disorders, and important technical support and clinical guidance are provided for neurological rehabilitation.

Description

[0001] The present invention belongs to the field of computer - aided diagnosis of brain diseases, and relates to a method for analyzing fMRI data based on deep learning. Specifically, an intelligent analysis method for fMRI based on graph coarsening and three - dimensional dynamic spatio - temporal graph attention network is proposed for the analysis of fMRI data and the accurate identification of sleep disorders. Background Art

[0002] Neurological function impairment is often accompanied by severe sleep disorder problems, especially more common in groups with impaired brain function. Sleep disorders may be related to the disorder of the brain neurotransmitter system, leading to disrupted biological rhythms, decreased sleep quality, and further affecting cognitive ability and emotional regulation. These problems not only affect the quality of an individual's daily life but may also exacerbate the further degradation of nervous system function. Therefore, accurately identifying sleep disorders is of great significance for formulating personalized intervention plans.

[0003] Functional magnetic resonance imaging (fMRI) is a non - invasive brain imaging technique that can be used to study the functional connectivity between different regions of the brain and is widely applied in the diagnosis of neurological diseases and the study of cognitive functions. However, fMRI data has characteristics such as high - dimensionality, strong noise, and dynamic changes. Traditional analysis methods are difficult to accurately capture the dynamic characteristics of the brain functional network, which limits the in - depth understanding of the neural mechanisms related to sleep disorders.

[0004] In recent years, deep learning techniques have made remarkable progress in the field of medical image analysis. Especially, methods based on graph neural networks (GNNs) can effectively model the complex functional connection patterns of the brain, providing new ideas for fMRI data analysis. Spatio - temporal graph convolutional neural network (ST - GCN) is a deep learning framework that combines spatial graph convolution and time - series modeling, capable of simultaneously learning the topological relationship between brain regions and their dynamic characteristics changing over time. However, when dealing with high - dimensional dynamic fMRI data, ST - GCN still faces problems of information redundancy and insufficient extraction of key features.

[0005] To more accurately detect sleep disorders, the present invention proposes an intelligent analysis method for fMRI based on graph coarsening and three - dimensional dynamic spatio - temporal graph attention network. This method can improve the recognition accuracy of sleep disorders, provide a scientific basis for personalized intervention and rehabilitation treatment, and at the same time promote the development of intelligent neuroimaging analysis technology. Summary of the Invention

[0006] The present invention discloses an intelligent analysis method for fMRI based on graph coarsening and three - dimensional dynamic spatio - temporal graph attention network for efficiently identifying sleep disorders. This method makes full use of the spatio - temporal characteristics of fMRI data, combines the topological learning ability of graph neural networks and the feature enhancement ability of the attention mechanism to achieve the accurate detection of sleep disorders.

[0007] The technical solution of the present invention is as follows:

[0008] Step 1, preprocess the experimental data and extract the BOLD signals of the ROI;

[0009] Step 1.1, collect resting-state fMRI (rs-fMRI) data and convert the DICOM format to the NIfTI format.

[0010] Step 1.2, perform data normalization, denoising, head motion correction, slice time layer correction, spatial normalization, spatial smoothing, temporal band-pass filtering, and covariate regression to eliminate motion artifacts and physiological noise.

[0011] Step 1.3, use the AAL brain segmentation template to divide the brain into 90 regions of interest, calculate the mean value of the time series signals of each ROI, and extract the blood oxygenation level-dependent BOLD signals.

[0012] Step 2, use Granger causality analysis to calculate the dynamic effective functional connections between the BOLD signals of different brain regions and construct a time-varying connection matrix; in Step 2, the specific steps are as follows:

[0013] Step 2.1; use the sliding time window method to capture dynamic features by dividing the time series data into multiple consecutive and overlapping windows. For a time series of length T, divided according to a window width of h and a sliding step size of s, the signals of each brain region of interest are divided into n = (T - h) / s + 1 time windows, and the time series within each window is regarded as a short-term stable signal subset.

[0014] Step 2.2, for each window, use Granger causality analysis to calculate the causal relationships between brain regions, thereby constructing the dynamic effective connection matrix F corresponding to the window w , the matrix size is m×m, where F i,j,w represents the causal intensity of brain region i on brain region j in window w.

[0015]

[0016] Step 3, use the Gaussian mixture model to coarsen the connection matrix. Through maximum likelihood estimation and posterior probability calculation, allocate the spatio-temporal feature vectors of the nodes to different Gaussian components to generate an optimized coarsened matrix to reduce the computational complexity. In Step 3, the specific steps are as follows:

[0017] Step 3.1, for each sample, extract its dynamic effective connection matrix

[0018] D (n) ∈R T×H×W , n = 1, 2,..., N

[0019] Among them, D (n) at the t-th layer (t ∈ {1,..., T}) represents the dynamic effective connectivity matrix of the n-th sample at time step t. Concatenate the values of each node (x, y) over T time steps into a spatio-temporal vector of length T:

[0020] x n,x,y =(D (n) (1, x, y), D (n) (2, x, y),..., D (n) (T, x, y)) T ∈ R T

[0021] The set of node features of the entire sample is represented as:

[0022]

[0023] Step 3.2, for the node feature matrix X (n) of each sample, perform soft clustering using the Gaussian mixture model, assuming it is clustered into supernodes:

[0024]

[0025] where π k is the weight of the k-th Gaussian component, μ k , ∑ k are its mean and covariance respectively. For a single sample n, train X (n) with shape H×W×T through maximum likelihood to obtain the parameters {π k , μ k , ∑ k}. Subsequently, calculate the posterior probability of each node feature vector x n,x,y on each component k:

[0026]

[0027] That is, each node (x, y) obtains a clustering membership vector γ n,x,y =(γ n,x,y,1 ,..., γ n,x,y,K ).

[0028] Step 3.3 Define a mapping function k(α, β) that maps the new matrix coordinates to the component k ∈ {1,..., K} of the mixture model, and coarsen the original matrix to

[0029]

[0030] Step 3.4, further define the node scoring function using the maximum contribution degree to measure the importance of nodes.

[0031]

[0032] Use the maximum contribution degree to independently complete clustering and coarsening for each sample, generating a coarsening matrix that matches the sample characteristics. Use the coarsened matrix as the input for the subsequent classification task to improve classification performance and reduce computational complexity.

[0033] Step 4, construct a three-dimensional dynamic spatio-temporal graph convolutional network for training;

[0034] Step 4.1, use the coarsened dynamic functional connectivity matrix as the input, and first capture the spatial structure between brain regions and the complex local dependencies in the time series through the 3D-CNN module.

[0035] Step 4.2, use the spatio-temporal graph convolutional network to extract the topological relationships between different brain regions through spatial graph convolution, and combine temporal convolution to model the temporal dependence of signals to fully explore deep spatio-temporal features.

[0036] Step 4.3, the model further introduces channel attention and spatial attention mechanisms, and uses the channel attention module to dynamically allocate weights for feature channels to enhance the expression ability of key features:

[0037]

[0038] Among them, G represents the feature matrix weighted by the channel attention mechanism, G' is the input feature matrix, sigmod represents the activation function F ex (·), MLP represents the serialization operation F sq (·).

[0039] Use the Spatial Attention module to increase the attention to key brain regions and important time points,

[0040]

[0041] Among them, G represents the feature matrix weighted by the spatial attention mechanism, G' represents the original input feature matrix, MaxPool and AvgPool respectively represent the max pooling and average pooling operations, and Conv represents the convolution operation.

[0042] Use the attention module to adaptively enhance the model's attention to important information, enhance the expression ability of key features, and ultimately assist in more accurate classification and analysis tasks.

[0043] Step 5, classification and evaluation;

[0044] The fully connected layer is used to map the extracted spatio-temporal features, and finally the sleep disorder classification results are output.

[0045] The performance of the model is evaluated through cross-validation, and ablation experiments are conducted for comparative verification to prove the effectiveness of the proposed method in identifying sleep disorders.

[0046] Advantages and innovations of the present invention compared with other existing technologies:

[0047] 1. Combining dynamic functional connectivity with an adaptive spatio-temporal graph convolutional neural network to improve the recognition accuracy. Traditional static functional connectivity analysis methods are difficult to capture the temporal dynamic characteristics of fMRI data. In this invention, Granger causality analysis is used to construct a dynamic functional connectivity matrix, and the CBAM attention mechanism is incorporated into the ST-GCN to enhance the expression ability of spatial channel features, enabling the model to automatically focus on important brain regions related to sleep disorders and improving the model's temporal modeling ability.

[0048] 2. Introducing the Gaussian mixture model for graph coarsening to reduce the computational complexity. The dynamic functional connectivity matrix is optimized by the GMM to reduce the data dimension while retaining the key topological information, enabling the ST-GCN to process large-scale fMRI data more efficiently.

[0049] 3. Applicable to sleep disorder diagnosis and facilitating personalized intervention. The present invention can provide a scientific basis for neurorehabilitation assessment, help formulate personalized intervention strategies, and thus improve the success rate of treatment. Description of the Drawings

[0050] Figure 1 Schematic diagram of effective matrix coarsening of the present invention at a certain time point;

[0051] Figure 2 Schematic diagram of the CBAM structure involved in the present invention;

[0052] Figure 3 Flowchart involved in the present invention;

[0053] Figure 4 Comparison of ablation experiment results involved in the present invention. Detailed Embodiments

[0054] The present invention will be further described below in conjunction with the drawings and embodiments.

[0055] The experimental data of this embodiment includes a total of 29 healthy subjects and 90 sleep disorder patients. The sleep disorder patients are divided into two groups according to the PSQI score: 35 are mild sleep disorder patients (5 ≤ PSQI score ≤ 10) and 55 are severe sleep disorder patients (PSQI score ≥ 10). In order to apply the fMRI intelligent analysis method based on graph coarsening and three-dimensional dynamic spatio-temporal graph attention network, the following steps are carried out:

[0056] Step 1: Use the data collected by the magnetic resonance scanner to perform data preprocessing to extract the BOLD signals of the disease population and the healthy control population. The main preprocessing steps include format conversion, removal of time points, slice time correction, head motion correction, spatial normalization, smoothing, removal of linear drift, etc. The data preprocessing of each subject is based on the Gretna software and mapped into the AAL90 template, dividing the whole brain into 90 regions of interest (ROIs), and extracting the BOLD signals of each ROI.

[0057] Step 2: Use Granger causality analysis to calculate the causal relationships between the BOLD signals of each brain region, thereby constructing the corresponding dynamic effective connectivity matrix.

[0058] Step 3: Use the Gaussian mixture model to perform graph coarsening on the connectivity matrix. Through maximum likelihood estimation and posterior probability calculation, assign the spatio-temporal feature vectors of the nodes to different Gaussian components to generate an optimized coarsened matrix. Figure 1 It is a schematic diagram of the optimized coarsened matrix.

[0059] Step 4: Construct a three-dimensional dynamic spatio-temporal graph attention network and introduce the CBAM mechanism into the ST-GCN structure. The schematic structure of the CBAM module is as Figure 2 shown;

[0060] Step 5: Input the coarsened matrix into the constructed deep learning network for training. The output of the network is the classification category to which the subject belongs. The overall process is as Figure 3 shown. The dataset is divided into an 80% training set and a 20% test set, and by optimizing the global parameters (such as learning rate, learning strategy, maximum number of iterations, batch size, step size, weight decay, etc.), the value of the network loss function gradually decreases and converges. Finally, evaluate on the test set, and the accuracy rate can reach 91%;

[0061] Step 6: Remove the coarsening step, ST-GCN, and CBAM module respectively to conduct ablation experiment comparative analysis to verify the contribution of each module to the model performance. The experimental results are as Figure 4 shown.

[0062] In summary, the present invention proposes an fMRI intelligent analysis method based on graph coarsening and three-dimensional dynamic spatio-temporal graph attention network, which effectively improves the diagnostic accuracy and reliability of sleep disorders.

Claims

1. An fMRI intelligent analysis method based on graph coarsening and three-dimensional dynamic spatiotemporal graph attention network, characterized by: The method specifically comprises the following steps: Step 1, preprocessing the original resting-state fMRI data to obtain standardized and denoised resting-state fMRI data; Step 2: Extract the BOLD signals of 90 brain regions (excluding the cerebellum) at each time point based on the AAL template; Step 3: Use Granger causality analysis to calculate the effective connections between BOLD signals in different brain regions and construct a dynamic effective connection matrix; Step 4: Use the Gaussian mixture model to cluster and coarsen the data of each sample. Through the maximum likelihood estimation and posterior probability calculation of the Gaussian mixture model, the spatiotemporal feature vector of the node is assigned to a Gaussian component to generate a coarsened graph of each sample; Step 5: Divide the obtained matrix data into a training set and a test set, input them into the constructed 3D dynamic spatiotemporal graph attention network, train them on the training set, adjust the model parameters, and finally evaluate the final classification results in the test set; Step 6: Remove each module and conduct ablation experiment comparative analysis to verify the contribution of each module to the model performance.

Images