Image Diagnosis Of Alzheimer’s Disease Based On 3D Convolutional Neural Network And Ensemble Learning

The automatic neuroimaging diagnosis of Alzheimer’s disease(AD)has attracted a lot of attention in recent years,but so far there is no better technical means to accurately diagnose and identify the related diseases.Due to the development and breakthrough of image recognition technology,the image diagnosis technology of Alzheimer’s disease is facing the following problems:(1)The traditional image diagnosis technology needs to extract image features artificially,and then use machine learning classification algorithm,which has strong subjectivity;(2)The brain image of AD patients has three-dimensional space features,and the traditional two-dimensional image recognition algorithm can not better extract the pathological features of the brain.Aiming at the above problems,this study improves the whole process from image feature selection,data preprocessing,and classification modeling.It combines MRI and DTI,trying to solve the problem of poor recognition rate of traditional image diagnosis.Experiments show that the algorithm has achieved good results in the imaging diagnosis of Alzheimer’s Disease,and the main contributions are as follows :(1)A deep learning model M3 DCNN based on multi-mode data feature fusion framework is proposed.Aiming at the problems of insufficient representation ability and recognition accuracy of single data,a convolutional neural network framework M3 DCNN based on multi-modal data is proposed.It includes:(a)MRI data representing macroscopic changes;(b)DTI data representing microscopic changes;(c)Clinical supplementary data.Different 3D convolution structure blocks are designed to be combined as feature extraction network,and all kinds of high-level features are fused and classified by fully connected neural network.After five times cross validation experiment,the average classification accuracy of AD vs NC is 97.6%,AD vs MCI is 92.0%,MCI vs NC is 92.9%,LMCI vs EMCI is 93.2%.The average accuracy of the four classification tasks is higher than the current level.(2)A deep learning model framework based on ensemble learning is proposed.In view of the uncertainty of neural network weight initialization,multiple training networks may converge to different positions,which leads to the problem that the classification results are too different.Using the idea of ensemble learning,the same data set is trained and predicted by multi-modal data feature fusion network,and the final voting decision is made for each sample category.The average classification accuracy of LMCI and EMCI is about 97.2%,and other classification tasks are improved by 1-3%.(3)The design and implementation of AD recognition prototype system based on M3 DCNN algorithm is conducive to the visualization of imaging and clinical diagnosis data and computer-aided prediction of AD disease.