Research On SMRI Image-aided Diagnosis Methods For Mental Diseases Based On Deep Learning

Mental diseases such as autism spectrum disorder and major depressive disorder have brought great burdens to patients,families and society.Therefore,early diagnosis,intervention and treatment of mental diseases play a vital role.Based on deep learning and s MRI images,the objective auxiliary diagnosis of mental diseases can be carried out and the corresponding biomarkers can be identified.At present,some methods have achieved preliminary results,but there are still some shortcomings such as the mismatch between model and data structure,and lack of interpretability of biomarkers.In view of the deficiencies of existing research,this thesis proposes three auxiliary diagnosis and potential biomarkers(brain regions)identification methods for mental diseases based on deep learning and s MRI images.(1)In Chapter 3,an auxiliary classification method for autism spectrum disorder(ASD)is proposed based on the self-attention model.Aiming at the problems of locality of CNN features,mismatch between data and model and lack of interpretability of biomarker identification existing in current machine learning and deep learning methods,a classification network based on self-attention mechanism is proposed.The model replaces the local feature extraction of CNN with the global feature extraction of the self-attention model.In the data preprocessing stage,the brain network data matching the self-attention model is constructed through the single-subject morphological brain networks.Finally,the potential biomarkers were identified based on the self-attention map of the model.The classification model in Chapter 3 achieved the best classification performance on the ASD dataset ABIDE Ⅰ,and identified the corresponding potential biomarkers of ASD.(2)In Chapter 4,an auxiliary classification method for major depressive disorder(MDD)is proposed based on graph neural networks.The complete graph data built in Chapter 3 is further extended to general graph data,and a classification network is constructed based on the Graph SAGE networks,which achieves the best classification performance of the major depressive disorder dataset REST-meta-MDD.Subsequently,the model interpretation algorithm was improved based on the GNN Explainer model to identify potential biomarkers of MDD.(3)In Chapter 5,an unsupervised subtype clustering method for MDD is proposed based on graph neural networks.Based on the Encoder-Decoder model and T-distribution dimensionality reduction,the unsupervised clustering of major depressive disorder was achieved,and the results of unsupervised clustering of three subtypes of depression were obtained.Subsequently,based on the improved GNN Explainer model interpretation method,model interpretation of unsupervised subtype clustering was achieved and potential biomarkers for three depression subtypes were identified.In summary,this thesis proposes three auxiliary diagnosis and biomarker identification methods for mental identification based on deep learning,which solves the shortcomings of current research,achieves the best classification performance on the corresponding dataset,finds potential biomarkers with interpretability,and provides a reference for the clinical diagnosis,treatment and research of mental diseases.