Research On Multi-Modal Alzheimer’s Disease Computer-Aided Diagnosis Model Based On Intergrated Learning

Alzheimer’s disease(AD) is a kind of irreversible neurodegenerative disease, the clinical manifestations is the memory decline and other cognitive impairment. Mild cognitive impairment(MCI) is the precursor of AD, but has not yet reached dementia. There is still no effective treatment for AD, previous studies have researched on AD early diagnosis based on single modal brain image data, but the effect is not ideal. With the development of brain imaging technology, multi-modal brain imaging data may obtain more comprehensive information related to AD diseases, and has important clinical value for the early diagnosis of AD.Multi-modal brain imaging features are usually in high dimensions, it has high time complexity and poor classification effect with single classifier. Therefore, this study will base on the integrated learning method, using multi-modal characteristics to explore and analyze early MCI and late MCI, AD and normal elderly group,to implement the computer-aided diagnosis. The main work is as follows:(1) Data collection. The s MRI data, PET data and DTI data of the four groups have been collected, and the original data of each modality were preprocessed to reduce the image noise, to prepare for the next step of feature extraction.(2) Feature extraction. Two different feature extraction methods were used to process the preprocessed s MRI and PET data, the first one is based on the AAL template, and the second is based on the significance analysis. Based on the preprocessed DTI image, obtain the anisotropy(FA) and the average diffusion rate(MD) maps, then significant analysis can be used on the FA and MD maps among four groups, get the FA and MD values of significant differences brain regions between the four groups, extraction each participant FA and MD values of corresponding significant differences brain regions as classification features.(3) Study on the integrated learning method. In recent years, in the field of AD research, ensemble learning has become a hot research topic in the field of machine learning. This study proposes PCA-FLDA integrated classifier and its basic principle is through the PCA method obtained different energy feature subspaces, using each feature sub-space are trained to get the base classifier, and finally multiple base classifiers by weighted voting mechanism are the final classification results. Secondly, this study on the base of the previous research realized a kind of multiple classifiers weighted voting integrated classification method. The method is based on five classifiers including fisher linear discriminant analysis, Naive Bayes, SVM, BP neural network and Ada Boost integrated strategy is weighted voting.(4) Classification experiment. The results of the study showed that, using the same classification methods, the characteristics of significant brain area can get better classification accuracy than the whole brain characteristics; With the same features, the classification results of integrated classifier are better than the single classifier. The classification effect of combination of multi-modal features are better than single modal features; using multiple classifiers integrated voting method for classification experiments between four groups, classification accuracy rate was not significantly improved and time consuming far more than single classifiers and PCA-FLDA classifier. Using PCA-FLDA classifier between the four groups, the classification effect is obviously improved and the time complexity is lower than others.PCA-FLDA integrated classifier has been presented in this paper can improve the classification accuracy and reduce the dependence on the feature space selection. Using mulit-modal combination features of significant brain regions for the classification test, The classification accuracy of AD vs.NC, AD vs. EMCI, AD vs. LMCI, EMCI vs. LMCI, EMCI vs. NC and LMCI vs. NC are 95.65% and 88.64% and 82.35%, 86.05%, 60.53% and 77.78% respectively.