Research On Stage Recognition Of Schizophrenia Driven By Magnetic Resonance Data

Magnetic Resonance Imaging(MRI)is a common clinical imaging examination method,and the use of MRI for clinical auxiliary diagnosis has become an important application direction in the medical field.Clinically high-risk patients with mental illness are a group of people who have not developed severe mental illness,and some may develop schizophrenia.Early identification of clinically high-risk patients and timely intervention are of great significance.At present,the diagnosis of mental illness mainly relies on the subjective judgment of doctors,which may lead to missed and wrong judgments,and the lack of objective biological indicators,resulting in patients missing the best opportunity for treatment.Deep learning technology has become one of the important means of medical image processing because of its high performance in object recognition and classification.Therefore,based on MRI and deep learning technology,this thesis studies the classification methods of patients with schizophrenia at different stages and healthy controls,to provide an objective,rapid and accurate basis for the clinical diagnosis and treatment of schizophrenia.The main work is as follows:(1)Construct a classification model for schizophrenia based on Convolutional Neural Network(CNN).Aiming at the problem that 3D MRI uses 2D slices to easily ignore the spatial information in the image,three models are proposed and constructed,including 2DResNet(R2D),3DCNN(C3D)and 3DResNet(R3D).In this thesis,MRI of Drug-naive First-episode Schizophrenia(DF-SZ),Clinical High-Risk Patients(CHR)and Healthy Controls(HC)collected by Nanjing Brain Hospital were used as data sources to classify them.And a data augmentation method called Mix up is introduced to improve the generalization of the models.Experimental results show that R3D performs best after enhancing the data set with Mix up,with an accuracy rate of 84.88%.(2)Develop an image key feature recognition method based on attention mechanism.Aiming at the unbalanced recognition accuracy of R3D for the three indiviuals,this thesis introduces attention mechanisms to guide the model to focus on the feature areas in the samples related to the classification task.On the basis of R3D,three classification models of DF-SZ,CHR and HC based on different attention mechanisms are built.Feature visualization method is used to observe the influence of the attention mechanisms on the image output features.And use the class activation map to observe the features that the model is most concerned about.The experimental results show that the MAR3D classification model based on the mixed attention mechanism performs the best,with an accuracy rate of 93.48%.The key features of the image that the model focuses on are temporal pole,orbitofrontal,anterior cingulate gyrus,and hippocampus.(3)Build a classification model for schizophrenia combining scale data with image feature.Aiming at the problem of unbalanced accuracy when only using MRI images,a method of using the subject’s scale data and image features to improve the performance of the model was proposed.In this thesis,MRI was used to predict the brain age of the subjects,and the actual-predicted brain age difference was combined with the cognitive scale,cortical thickness,local gyrification index and MRI for multimodal feature fusion,and then the fused data was used for schizophrenia classification.The experimental results show that the multimodal feature fusion can further improve the performance of the model,and the accuracy rate is increased by 2.3%.