Computer-aided Diagnosis Of Alzheimer’s Disease Based On Small Functional-MRI Datasets

The incidence of Alzheimer’s disease(AD)is on the rise with the development of an aging society,so it is of great significance to build a computer-aided diagnosis(CAD)system which is based on machine learning to help the diagnosis of AD.The development of neuroimaging technology,especially the functional-Magnetic Resonance Imaging(fMRI)technology helps us analyze the Alzheimer’s disease,and construct the CAD systems.Researchers have developed various machine learning methods for the diagnosis of Alzheimer’s disease with fMRI images.However,the collection of AD cases is difficult,so how to classify AD with small fMRI datasets is our research focus.In this paper,based on machine learning,we have studied the AD classification algorithms with some small fMRI datasets.The main research contents are as follows:Firstly,for the AD classification task with a small fMRI dataset,we analyzed the property of the fMRI and proposed an unsupervised deep feature learning algorithm,Local Binary Covariance based-3D-PCANet(LBC-3D-PCANet).We first get a series of feature maps through the filter learned by PCA,then pooling.Finally,features are learned from the feature maps using the Local Binary Covariance matrix.Experimental results show that LBC-3D-PCANet has achieved good performance with an AD fMRI dataset.Compared with the existing PCANet algorithm,LBC-3D-PCANet can reduce the consumption of computing resources effectively,and this algorithm is more suitable for small datasets than CNN.Secondly,to identify AD from an ultra-small dataset based on functional brain network,we proposed the Domain Weighted-Joint Distribution Adaptation(DW-JDA)algorithm to build the classifier with an auxiliary dataset.The DW-JDA algorithm project the source and target domain samples into a new feature space to reduce the distribution differences between domains,and then build a classifier that works better in the target domain by assigned larger weights to the target domain samples in the classifier.Experimental results show that this algorithm improves the performance of the classifier on an ultra-small dataset using shared data by reducing the distribution differences between datasets.Thirdly,to identify AD fMRI images from an ultra-small dataset,we analyzed the learning process of deep feature learning,and propose a deep transfer learning model,Subspace Alignment-PCANet algorithm.First,with the help of auxiliary data,we extract features with LBC-3D-PCANet from 3D neuroimaging datasets and get the domain shared information and domain special information.Then it reduces the distribution differences between source and target domain by Subspace Alignment algorithm for transfer learning.Experimental results show that this algorithm improves the performance of the classifier on an ultra-small fMRI dataset by making better use of the auxiliary data.Finally,based on the similarities and differences between neuroimaging data and natural images,we discussed the possibility of transfer learning with cross-domain data to improve the accuracy of CNN on a small fMRI dataset.And we explored the impact of auxiliary datasets with different fields and different scales on the pre-training of CNN models.Experimental results show that cross-domain images can significantly improve the accuracy of CNN on AD fMRI datasets.Besides,with more auxiliary samples and more similar modalities,the pre-trained CNN works better on fMRI datasets.However,pretraining with a too-small dataset may have a negative effect on the transfer learning of CNN.