Neuroimaging Analysis And Brain Disease Diagnosis Based On Machine Learning

With the rapid development of neuroimaging technology,neuroimages such as Magnetic Resonance Imaging?MRI?has been widely used in clinical diagnosis of Alzheimer's disease?AD?and mild cognitive impairment?MCI?because of the advantages of noninvasibility and high image resolution.In practical clinical application,the radiologists have to recognize and quantify some regions of inter-est by manual.However,with the increase of patients,recognition and quantiza-tion with manual become unrealistic.With the development of image recognition technology,it became possible to analyze the neuroimaging data and assist the diagnosis with the help of computer.However,due to the inherent high dimension of neuroimaging data,it is still an important problem to analyze and build an ef-fective and efficient brain-disease diagnosis model.Neuroimaging computing and analysis include the steps of image pre-pro-cessing,segmentation,registration,feature extraction and classifier design.This dissertation aims to address two key problems of neuroimaging computing and analysis,i.e.,feature extraction and the classifier design,based on machine learn-ing technologies.Then the developed algorithms are tested on ADNI database to differentiate the normal,MCI and AD subjects.The contributions of this paper are mainly focused on:?1?Based on the sparse representation model,we propose a group sparse representation based classifier by combining l₁-norm and l₂-norm regularizations,which can make use of the label information of training data.The group sparse representation based classifier performs better on the classification problem of MCI convertor?convert to AD?and MCI non-convertor than the traditional clas-sification methods.?2?Based on the deep learning framework,a feature extraction method is proposed with stack auto-encoder network to extract the non-linear high-level fea-tures from simple low-level features.The experiments have justified that under the same classifier,the features generated by stack auto-encoder perform better than the original feature vectors.