Study On Functional Magnetic Resonance Disease Diagnosis And Cognitive State Prediction Based On Graph Neural Network

Functional Magnetic Resonance Imaging(f MRI)is an emerging neuroimaging technique with non-invasive,radiation-free,and high spatial and temporal resolution properties.Resting-state functional MRI(rs-f MRI)is widely used for the diagnosis of neurological disorders such as Alzheimer’s disease and autism because it does not require subjects to complete a task.Task functional MRI(t-f MRI)is an image scanned when the brain performs specific tasks such as memory,recognition,and movement,and is widely used to study the cognitive functions of the brain.In recent years,automatic diagnosis of neurological disorders and prediction of cognitive status of the brain combined with deep learning techniques have received increasing attention.In this thesis,we investigate the problems and challenges in diagnosing neurological disorders and predicting cognitive state based on rs-f MRI and t-f MRI,respectively,in combination with graph neural networks.The main work of this thesis contains three parts.First,to address the problem of data heterogeneity under multi-site,we propose a novel end-to-end denoising population graph convolutional network for multi-site disease prediction.Large-scale collaborative initiatives to collect and share brain imaging and behavioral data from laboratories or hospitals around the world pose a multi-site problem.Differences in acquisition protocols and MRI scanner types result in sites with different data distributions,and this inhomogeneity can affect biomarker extraction and even mislead the diagnosis.The framework we proposed aims to construct a robust population graph and use the denoising information for disease diagnosis.The framework disentangles raw rs-f MRI into site-invariant and site-specific information through disentangled representation learning,constructs edges of the population graph using site-specific information,and represents nodes using site-invariant information.Under this strategy,the framework can both connect multi-site samples and avoid the propagation of unfavorable information between two heterogeneous samples when they are connected.Second,to address the privacy preservation problem under multi-site,we propose a novel federated learning framework based on population subgraph augmentation.The centralized sharing of subject data in a multi-site collaborative project faces privacy and ownership issues.In the privacy-preserving scenario,federated learning can federate multiple sites for model training,but the population graphs constructed within each site suffer from graph structure corruption,specifically missing information across sites.The framework we proposed synthesizes nodes and edges with real data distribution through the population subgraph augmentation module,and uses the enhanced local population graph to federate other sites to centrally train a global graph convolutional network for privacypreserving distributed disease diagnosis.Third,in order to address the problem of adaptive brain regions in brain cognitive state prediction,a novel end-to-end individual graph adaptive convolutional network is proposed.The study of brain cognitive function is beneficial to promote the pathological study of neurological diseases and provide antecedents for the early diagnosis.Different regions of interest(ROI)play different roles in achieving a particular cognitive function,and the framework we proposed adaptively predicts the cognitive state of the brain based on ROI.The framework generates functional brain network representations of a series of t-f MRIs and designs an adaptive graph convolution module and an adaptive graph pooling module to predict the corresponding cognitive states of the brain,where the adaptive graph convolution module adopts a multi-headed attention mechanism and the adaptive graph pooling module adopts an LSTM-based attention mechanism.The proposed denoising population graph convolutional network is trained,validated and tested on ABIDE,a shared dataset of autism brain imaging containing 20 sites,and the experimental results show that the method achieves better performance in multi-site disease diagnosis with test accuracy of 80.77%,which is 3.34% improvement compared to the state of art.Meanwhile,the federated learning framework based on population subgraph augmentation is trained,validated and tested on the ABIDE I&II dataset containing 20 sites,and the experimental results show that this method outperforms the existing federal learning methods in terms of accuracy,AUC,precision and recall,and the precision rate is improved by 1.1% compared with the start of art,which is 65.8%.In addition,the individual graph adaptive convolutional network is applicable to label brain activities on 21 cognitive states under six experimental task settings,and the method is trained,validated,and tested on the NMA dataset in the Human Connectome Project,and the experimental results show that the method improves the accuracy of cognitive state classification and reduces the training time,demonstrating the excellent performance of the method,and in the case study The ROI scores acquired by the model were found to be consistent with the results of cognitive state literature studies.