AD Classification Methods Based On Multi-Modal Neuroimaging

The Alzheimer's disease(AD)classification is a medical aid diagnosis method that extracts valid features from different medical images or biomarkers and classifies the features using machine learning methods.Early detection and timely treatment of AD and mild cognitive impairment(MCI)is crucial for suppressing the progression of the disease.In the past few years,medical imaging technology has made great strides,so we can use a variety of medical images for analysis.Experiments show that the effect of combining multiple modalities for classification is better than using just one modality.In this dissertation,feature fusion methods and classification algorithms are combined into two AD classification algorithms.These two methods can effectively utilize the complementary information in multi-modal data.1.The initial features used in the two methods are 90 regional volumes of MRI,90 regional average intensities of PET,and 3 CSF biomarker measures.MRI images and PET images need to be preprocessed first.Since some information in the whole brain is irrelevant to AD,We use t-test and Fisher's criterion for feature selection of the initial features.Then,the kernel matrices of each modalities are calculated respectively.Finally,we use the multiple-kernel SVM which can be used to combine multiple modalities of features for the classification of AD and healthy controls.2.In this paper,we will use another feature fusion method to fuse the features of multiple modalities.Firstly,we use the random forest method to measure the similarity between the samples and construct the feature graph,and each of the three modalities will construct a feature map.Secondly,the method of nonlinear graph fusion is used to fuse the similarity matrix of multiple modalities to generate a unified graph of final classification.Finally,the final unified graph will be classified by the random forest classifier.