Diagnosis Of Alzheimer’s Disease Based On The Unsupervised Deep Neural Network

Deep learning technologies have played more and more important roles in computer aided diagnosis(CAD)in medicine.However,there is an always faced problem of insufficient labeled medical data for deep learning model training purpose.The problem will also affect the final classification accuracy.Alzheimer’s disease(AD)commonly occurs in the elderly.It brings a great burden on the family and society.Diagnosing AD early can limit the devastating physical on patient,which can make the great significance for the treatment of AD.However,there are always faced problem of insufficient labeled AD data because of the cost and specification for data collection,which seriously affects the training of deep learning models in AD prediction.In this paper,we propose three methods for computer aided Alzheimer’s disease prediction by using limited labeled MRI images.Specifically,we implement the transfer learning based Convolutional Neural Network(CNN)and unsupervised CNN into our work.The main research works are as following:(1)We proposed the transfer learning method for computer aided AD prediction.The method utilize only three views of MRI image in the form of three orthogonal planes(TOP)as a data source.We investigate a pre-trained Overfeat CNN by a dataset of neural image to learn the feature of TOP slices,and a two-stage classifier is followed for AD prediction.Experimental results demonstrated that,the framework of this proposed method can be thought as a learning-based approach to reduce the dimensionality of the 3D problem,and handle the 3D data analysis via a proper selection and combination of 2D views.(2)We proposed a PCANet method from aspects of unsupervised deep learning for computer aided AD prediction.We present the PCANet which the filter banks of this CNN are prefixed by the conventional unsupervised machine learning method,to learn the features of TOP slices.This is also followed by a two-stage classifier to achieve the final prediction score in AD prediction.Experimental results demonstrated that,the proposed method achieve promising performance in AD prediction.(3)We proposed a 3D-PCANet method from aspects of unsupervised deep learning for computer aided AD prediction.The 3D-PCANet method utilize full 3D data of MRI as input to classify each MRI image.The main steps of 3D-PCANet are tocompute the descriptors consist the first 3D convolution stage,the second 3D convolution stage,and the output stage(including Hashing and Histogram generating).We investigate a 3D-PCANet to learn the feature of each MRI image with full 3D data,and followed by a SVM classifier for AD prediction.Experimental results demonstrated that,the proposed methods achieve promising performance in AD prediction.Although TOP views of MRI image obtain acceptable prediction accuracy,the full 3-D MRI image provides more discriminative information for AD prediction.