Automatic Classification And Individualized Prediction Of Schizophrenia Studied With Brain Imaging Features Using Multi-modal MRI

Schizophrenia(SZ)is a severe mental disorder with hallucination,delusion,apathy and working memory dysfunctions.The lack of explicit biomarkers in the diagnosis of schizophrenia increases chances of misdiagnosis rate and missed diagnosis rate.In this paper,we made systematic comparisons among different classification models and regression models.In addition,we discussed the most discriminative features and features contributing to the prediction.The main contents of this paper are:1)This study recruited 61 first-episode schizophrenia(FESZ)patients,79 chronic schizophrenia(CSZ)patients,and 205 normal controls,and acquired their structural MRI,resting-state functional MRI data and diffusion tensor imaging.We calculated gray matter volume,degree centrality,amplitude of low frequency fluctuation,and regional homogeneity of 246 brain regions,based on Human Brainnettome Atlas,and fractional anisotropy,mean diffusivity,axial diffusivity and radial diffusivity of 50 white matter regions,based on white matter parcellation map.Using the above multi-modal brain image features as the input of the classifier,based on the nested cross-validation method,the k-nearest neighbor method,the linear support vector machine,the logistic regression,the linear discriminant analysis and the random forest are used to classify the FESZ group,the NC group,and the CSZ group.We found that a classifier combining three modal features effectively improves the accuracy of classification compared to a classifier that uses only a single modality.Moreover,it is found that the classification of the linear support vector machine and the random forest is better than other classifiers,the maximum accuracy of classification is 85%.Further,it was found that functional MRI features made a great contribution to distinguish SZ patients from NC,and structural MRI features play the key role in discriminating FESZ patients from CSZ patients.2)In the predictive analysis of Positive and Negative syndrome scales(PANSS)scores,we applied the same features used in the discriminative analyses to compare the five pattern regression models,including least squares regression,ridge regression,LASSO regression,linear support vector regression and elastic network regularized linear regression.We found that elastic network regularized linear regression achieved the best prediction accuracy(r=0.43),linear support vector regression and ridge regression achieved significant prediction results.The features contributing to the prediction were primarily located in the frontal and temporal lobes.In this paper,based on the multi-modal MRI,we compared different pattern classification and pattern regression algorithms,achieved good classification performance in discriminative analysis of schizophrenia,and significant predicted PANSS total scores.The findings in the thesis may provide reference for the algorithm selection of SZ assisted diagnosis system and disease state evaluation of schizophrenia.