Deep Learning Based Feature Representation For AD/MCI Classification

Alzheimer’s disease(AD)is a chronic disease with cognitive impairment as a major clinical feature and a common high incidence in elderly diseases.Mild cognitive impairment(MCI)is the prodromal stage of AD,being a state between normal aging and dementia.Due to the slow progression of the disease,there is no obvious clinical symptom at the beginning of the disease.When patients have obvious symptoms,most of them have been advanced and can not be treated.The early diagnosis and intervention of AD are critical for disease control and can delay the progression of disease,making it the focus of the study.For AD/MCI classification,unlike the previous methods that consider simple low-level features such as gray matter tissue volumes from MRI,mean signal intensities from PET,in this paper,we used a deep learning method to extract latent feature representations from these primitive features.Then,combine the latent feature representations with low-level features to classify and predict the different processes of AD and to improve the classification accuracy and model robustness,to assist in the diagnosis of disease.The main contents of this paper are as follows:(1)Select some sample from the ADNI database,and quality control of the data,filtering out the samples with MRI,PET and CSF baseline data simultaneously.Meanwhile,the demographic information including MRI data(organizations volume of 132 Regions of Interests),PET data(mean signal intensities of 132 Regions of Interests),CSF data(contents of three biomarkers Aβ42,T-tau,P-tau)and two types of clinical scores(MMSE and ADAS-Cog),age,gender,and educational level of each sample were collected.After the data quality control,788 samples were obtained,including 165 healthy elderly samples,97 subjective memory complained patients,259 patients with early mild cognitive impairment,141 patients with late mild cognitive impairment and 126 patients with Alzheimer disease.(2)Classify AD / MCI based on Stacked Auto-encoders and BP Neural Networks.Firstly,utilizing the method of randomly dividing the subset and stratifying the subset respectively,four 2-classification experiments,one 3-classification experiment and one 5-classification experiment were carried out under the MRI,PET and CSF feature set.Then,the paper analyzes the influence of batch size,iteration times and network structure on reconstruction error and recall rate in SAE,determining the optimal parameters and the optimal network structure.At the same time,the implicit features of three phenotypic data of MRI,PET and CSF were extracted respectively.(3)Classify AD / MCI based on Multi-task and Multi-kernel SVM Learning.First,combine MRI,PET and CSF original features and serve as input to MTL+BP network.And compare the optimal classification accuracy of each classification problem with Single-task Learning,which verifies the validity of Multi-task Learning.Next,by concatenating the SAE-learned feature representation with the original low-level features,we construct an augmented feature vector to conduct the experiment based on Multi-task and Multi-kernel SVM Learning.And analyze the optimal classification accuracy compared to Single-kernel SVM Learning.Finally,the optimal classification algorithm is determined and the higher classification accuracy of AD/MCI is obtained.