Study Of Deep Learning Model For Alzmeiher’s Disease

Alzheimer’s disease is a fatal neurodegenerative disease of unknown etiology,and there is still no effective treatment to prevent or reverse the occurrence of Alzheimer’s disease.Early cognitive impairment is the early stage of AD,and diagnosis of MCI is very important.At present,the diagnosis and follow-up treatment of AD require the help of medical imaging.The application of non-radiation MRI technology in brain diseases is becoming more and more extensive,and MRI has the advantages of high soft tissue resolution and multi-plane and multi-sequence imaging,which has important significance in the diagnosis of AD.This work mainly uses MRI image to perform AD/MCI classification and the work mainly includes the following three aspects:1.According to characteristics of brain MRI data of three-dimensional structure,this work builds the 3D-CNN.Factors affecting performance is analyzed carefully,including number of layers,the pooling method,activation function,the optimizer settings,according to the experimental results,when selecting four convolution layers,using max pooling,Re Lu activation function and Adam optimization method,the 3D-CNN model model achieves optimal performance.2.Three different classification methods were compared in the design experiment:(1)Using the 3D-CNN for classification;(2)Using the statistical features of the brain extracted by Freesurfer and other clinical diagnostic information for classification by XGBoost;(3)Using a 2D-CNN with the same depth,parameter setting of 3D –CNN for classification.The experimental results show that the 3D-CNN performs well in AD/NC and MCI/NC classification prediction,the classification accuracy is 93.3% and 78.3% separately,which is much higher than the traditional machine learning method and the 2D-CNN.3.This work uses the combination of different parallel slices of MRI to generate RGB images,and the Inception V3 and Res Net network models are fine-tined to predict the transformation of MCI samples.The classification accuracy of f MCI/s MCI is 76.4% and 72.7%.After adding 12,18 and 24 months of image data,the classification accuracy of the two models is improved again.At the same time,the immune-related gene expression data are selected and the image features extracted from the full connection layer of Inception V3 are fused for f MCI/s MCI classification.The classification accuracy is improved to 83.2%,greatly improving the classification effect.This paper explored the deep learning methods such as three-dimensional convolutional neural network for the analysis of MRI images of AD,and the results showed that the classification accuracy could be improved.Fusion of genomics data can further improve the prediction performance of MCI transformation.