Comparison Of Different Convolutional Neural Networks In Recognition Of Stable And Progressive Mild Cognitive Impairment

The research object of this dissertation is patients with Alzheimer’s disease,commonly known as Alzheimer’s disease（AD）.my country is the country with the most AD patients in the world.AD has become one of the most serious public health and social problems,and it has caused a huge economic burden to the entire society.The diagnosis of AD and its precursor form of Mild Cognitive Impairment（MCI）requires a highly discriminative feature representation for classification.Deep learning technology can learn these representations from data,and the results of the research can help clinicians or communities to screen the elderly who may be sick,carry out related treatments earlier,delay the development of diseases,and reduce medical and nursing costs.Convolutional neural network recognition of AD and MCI images mainly focus on four methods: the method based on two-dimensional image slices,the method based on regions of interest,the method based on patch-level image blocks,and the method based on voxels.In the method based on two-dimensional image slices,most related studies have added other complementary information to improve the accuracy of classification.The effect of independent two-dimensional image slices on the classification task is not yet known.In many related studies based on patch-level image blocks and interesting regions,complex network models have been proposed.Although the accuracy of two-dimensional image slices is significantly improved,the computational cost is increased.This dissertation proposes a series of network improvement methods and multi-branch network structure in response to the problems in the above research.The main research work is as follows:（1）Aiming at the problem of whether the classic convolutional neural network can extract effective recognition features from MRI image slices,this thesis proposes to construct a three-view network structure based on the convolutional neural network to extract features from different perspective directions.The specific work is as follows:firstly obtain a part of the MRI slices with the largest brain tissue structure from the three perspectives,then train a convolutional neural network for the three perspectives,and finally merge the output features of the three networks into full connection to achieve classification.Through the comparative experiments of several convolutional neural networks,it is tested whether the classical convolutional neural network can effectively classify the two-dimensional image slices of stable and progressive MCI patients.（2）Aiming at the problem of losing three-dimensional spatial local relationship information when two-dimensional image slices are used as independent samples,it is proposed to use 34-layer 3D-Res Net to classify the three-dimensional MRI images of stable and progressive MCI patients.The specific work is as follows: The three-dimensional MRI image is divided into small blocks and input into 3D-Res Net,and the characteristics of residual blocks are used to solve the problem of information loss in the process of feature extraction.By optimizing the activation function and regular term to solve the possible over-fitting problem of the network.The experimental comparison proves that the optimization of Leaky Re LU and L2 regular term can effectively avoid the over-fitting problem,and it is verified that the addition of multi-modal data can effectively improve the accuracy of the network model.（3）Aiming at the problem that the three-dimensional MRI brain image contains information that is not helpful for the classification task,a 3D-Dense Net-based multi-branch network structure is proposed to realize the classification of stable and progressive MCI patients.The specific work is as follows: first extract the hippocampal tissue images of the left and right hemispheres and calculate the hippocampus volume according to the prior knowledge,then train the Dense Net network for the hippocampus tissue images of the left and right hemispheres respectively,and finally merge the hippocampus volume information of the left and right hemispheres to achieve classification.The experiment analyzed the classification effect of each branch,and the experiment of different input size images verified that the hippocampus is the main structure of the classification task.