Classification Of Mental Diseases Based On Multimodal Magnetic Resonance Imaging

In recent years,with the acceleration of social rhythm and the increase of life pressure,the number of patients with mental diseases is increasing year by year.Schizophrenia is one of the most common mental diseases,but the diagnosis of schizophrenia patients is a difficulty in the medical field.At present,the diagnosis of schizophrenia mainly relies on the subjective experience of neurologists and clinical scales,but it brings a huge workload to neurologists.Meanwhile,it is very likely to cause missed diagnoses and misdiagnoses.Therefore,this thesis combines MRI with machine learning and deep learning to realize the diagnosis and classification of schizophrenic patients,which can assist the clinical diagnosis.The main research contents of this thesis are as follows:First of all,this thesis studies a classification method of mental diseases based on multimodal features and machine learning algorithms.In the experiment,the single-modal features and the multimodal features combined with both structural and functional MRI are extracted.The machine learning algorithms are used to classify schizophrenic patients.The experimental results show that SVM and KNN algorithm have better classification performance,and when using the same machine learning algorithm,the classification performance based on multimodal features is generally better than that of single-modal features.Secondly,this thesis investigates a classification method of schizophrenia patients based on gray matter volume image and CNN.This thesis constructs a CNN composed of three convolutional layers,one max-pooling layers and three fully connected layers.The network takes gray matter volume image as the input to solve the problem of binary classification between schizophrenia patients and normal people.The experimental results show that the classification performance of the CNN architecture is better than the traditional machine learning algorithms and classical deep learning networks.This thesis optimizes the CNN by depthwise separable convolution,and discusses the changes of classification performance,parameter and computation under different pooling scales.This method not only improves the classification performance,but also reduces the number of parameters in the model and the computation cost.