Research On Classification Of AD Stage With Structural MRI And Deep Learning Model

Alzheimer’s Disease(AD)is the most common case of dementia,which will be a global burden over the coming decades,due to the aging trend.The progression risks of AD allow the patients to take prevention measures before irreversible brain damages are shaped.Mild cognitive impairment(MCI)is a prodromal stage of dementia.Predicting MCI to AD discriminating and conversion play critical roles in preventing the progression of AD.Structural Magnetic Resonance Imaging(sMRI)has ability to capture subtle differences in the brain.However,the early AD diagnosis by sMRI also faced the difficult like high dimension and heterogeneity.Deep learning network,which transfers the original data into a higher level and more abstract expression.The method of deep learning can reduce the subjective factor,which have gained great popularity in diseases classification and prediction.In this research,we propose to a pre-trained deep convolutional neural network for AD、MCI and NC classification,and prediction of transition risk of MCI subjects can be further categorize as MCI converter(MCIc)or MCI non-converter(MCInc).The main contents of this paper are as follows:(1)The data selection and setup the deep learning environment.The big data is the foundation of deep learning.From the database of Alzheimer’s Disease Neuroimaging Initiative(ADNI)has launched,which provides material for the research by deep learning in the field of AD.But has a problem that the database is too big and complex.So first introduced the overview of ADNI website and the steps of data screening.Deep learning framework and platform are indispensable link in solve practical problems by deep learning algorithms.In this research,considering the experiment data,requirements and other factors,we choose building the platform of CAFFE under Linux,and configure complete interface.(2)The research classification of AD,LMCI and normal control(NC)with sMRI.Although the deep convolution network(CNN)has achieved high accuracy in the classification of nature images.There is the difference between medical image and nature image.And the number of these has a big difference.But the deep network usually has versatility,the method of transfer learning can efficient use pre-trained model to new data set.In this chapter,we use the method of transfer learning to the three stages of AD by AlexNet model.The result show that the feature by transfer learning,after feature reduction and selection by principal component analysis and sequential forward search,can play a better classification of AD,LMCI and NC in support vector machine model.(3)The research on MCI convert prediction.This chapter will classify of MCI converters and MCI non-converters,in order to prediction the risk of MCI convert.The early positive intervention with high-risk MCI converters can be positive effects on AD defense.This chapter make a prediction model to MCInc and MCIc base on the previous chapter,and classify the Late MCI(LMCI)and Early MCI(EMCI).This research base on sMRI data,using deep CNN model,classification of AD stages by feature extract.Then combined the machine learning methods to feature selection and reduction to the purpose of optimize classification model.The propose of this research is that classify the AD stage and screening the high-risk of MCI converters.