Application Of Deep Learning In Diagnostic Assistance For Mental Illness Based On MRI Data

Depression,Alzheimer's disease(AD)and behavioral abnormal frontotemporal dementia(bvFTD)are mental disorders that are affected by many factors including biological,psychological,and social environments,and their incidence rate is upward trend.However,owing to lack of the uniform objective diagnostic criteria,the diagnostic accuracy of mental disorders is not high.The early cure rate of psychiatric diseases is high and the rehabilitation effect is good.Therefore,it is extremely important for research and clinical to further improve the accuracy of early diagnosis by constructing a psychiatric auxiliary diagnosis model.Domestic and foreign researchers are actively exploring the auxiliary diagnosis model of depression and dementia based on magnetic resonance imaging(MRI)data,but the existing auxiliary diagnosis still has the following problems:(1)The feature selection method has a significant influence on the results of auxiliary diagnosis;(2)The data collection of patient's MRI is difficult,and the number of available samples is insufficient,which makes some machine learning algorithms be unable to be used;(3)There are many clinical manifestations of AD and bvFTD overlap,and it is difficult to diagnose.At present,there is still lack of auxiliary diagnostic studies on these two types of dementia.In this context,this paper studies the diagnosis of depression and dementia(AD and bvFTD)based on MRI data.The main research contents are as follows.(1)The feature selection method has a significant impact on the results of auxiliary diagnosis.Against the problem,this paper proposed a hybrid feature selection algorithm and constructed a multivariate pattern analysis(MVPA)auxiliary diagnosis model for depression.Based on a full data-driven approach,the stationary and dynamic functional connection network of 41 subjects were established,and features were extracted.Then,different feature selection methods and different classification algorithms were used to classify patients with depression.The accuracy can reach 97.56% based on stationary and dynamic,which indicates that the proposed hybrid feature selection algorithm can effectively improve the classification accuracy of MVPA.Further analysis of the steady-state experimental results demonstrates that features with stronger discrimination are mainly located in the frontal network,default network,and visual network.(2)For the problem that the fMRI data is difficult to collect and the number of available samples is insufficient,the conditional deep convolution generation network(CDCGAN)was used to expand the dataset,and a MVPA model based on the hybrid feature selection algorithm was applied to study the auxiliary diagnosis of depression.The experimental results show that the accuracy of the model with CDCGAN is 92.68%,which is obviously better than the accuracy 68.29% without CDCGAN.It is also better than the classification result of the autoencoder.Further analysis shows that the features with high discriminatory power are roughly the same as previous research results,which shows that using CDCGAN can solve the problem of insufficient available samples to some extent.(3)Against the difficult of clinical differential diagnosis for AD and bvFTD,this paper proposed an assisted differential diagnosis model of AD and bvFTD based on deep learning.Firstly,based on the MRI data,the white matter and gray matter of the 132 subjects were separated.Then,the gray matter volume features of ROI brain regions were extracted based on different brain templates,and the three-classification was implemented by using convolutional neural network and deep forest respectively.Experimental results show that the accuracy of deep convolutional neural network is up to 81.80%,and the accuracy of deep forest is up to 74.24%.They are all better than the traditional method's three classification results 66.70%.This shows that the assisted diagnosis models of AD and bvFTD based on deep learning technology have better classification results and have potential clinical application value.It is worth further researching and exploring.This research mainly includes the following three innovations:(1)proposing a hybrid feature selection algorithm;(2)proposing a diagnosis aid model for depression based on generative confrontation networks;(3)proposing the AD/bvFTD three-class auxiliary diagnosis model based deep learning technology for the first time...