Research On Multi-modal Alzheimer’s Disease Prediction Method Based On Deep Learning

Alzheimer’s disease(Alzheimer Disease,AD)is currently one of the most common senile diseases in the world.The morbidity and fatality rate of this disease in the world are gradually increasing,but it has not been paid attention to by people.The symptoms are discovered in time,and they cannot be recovered when they attack.There are several manifestations in the early stage of the complete onset of AD,including: Early Mild Cognitive Impairment(EMCI)and Late Mild Cognitive Impairment(LMCI).Therefore,correct diagnosis and early treatment of patients with cognitive impairment play a vital role in delaying the onset of AD.At present,with the gradual maturity of medical imaging technology,it has become the main tool in the field of human brain research,which mainly includes Magnetic Resonance Imaging(MRI)and Positron Computed Tomography(PET).Aiming at the problems of artificially extracting features and the low classification accuracy of classification models,two methods for the classification and prediction of Alzheimer’s disease multimodal data are proposed,which are based on the depth double deep learning method of linear attention network(DBAR)and the machine learning method based on multi-core support vector machine(DBARM-SVM).In this thesis,twoclass baseline experiments,two-class comparison experiments,and four-class baseline experiments are designed respectively based on the MRI and PET multi-modal data sets.In the two-classification experiment,three groups were set: AD vs NC,MCI vs NC,AD vs MCI,and the accuracy of DBAR in the three experiments were 97.34%,95.01%,and93.82%.MCI divided into EMCI and LMCI as the classification of four-class experiments.The accuracy of DBAR reached 97.15%.In the course of the experiment,it was found that the softmax classifier has problems such as weak generalization ability and large fluctuations in accuracy.Aiming at the problems found on the softmax classifier,this thesis proposes the DBARM-SVM method based on DBAR algorithm.The feature of the experimental data extracted though the DBAR network,and the single kernel function are optimized by PSO.The DBARM-SVM composed by the single kernel function with the optimal parameters which though the a linear combination method.A series of two-class experiments including single-core experiment,multi-core experiment and comparative experiment were set up.Three experimental groups were AD vs NC,MCI vs NC,AD vs MCI,and the accuracy of DBARM-SVM were 98.59%,96.53% and 95.77%.It shows that this method has better classification performance in the early diagnosis of AD,and can achieve better classification results.In summary,this thesis designs the DBAR algorithm and the DBARM-SVM algorithm for the diagnosis of Alzheimer’s disease at various stages based on the multimodal image data of Alzheimer’s disease.The design ideas of different models and the classification results of medical imaging data at various stages are discussed.A large number of experiments show that the above two algorithms can obtain ideal classification results better than other deep learning or machine learning methods.