Research On Multimodal Image Analysis Algorithms Of Alzheimer’s Disease Based On Deep Learning

Alzheimer’s disease(AD)is a chronic neurodegenerative disease that manifests clinically as cognitive impairment and memory deficits,commonly seen in the aging population,and is characterized by a slow onset and incomplete cure.We can classify subjects as healthy controls(Healthy Controls,HC),mild cognitive impairment(Mild Cognitive Impairment,MCI)and AD according to the level of AD.By the time a subject is diagnosed with AD,the disease has already caused irreparable damage to the subject’s health.In practical clinical applications,early detection and correct differentiation of MCI and AD and effective intervention and treatment for them are important to prevent or delay the onset of Alzheimer’s disease.But diagnosis of AD is very difficult and requires highly discriminative features for accurate classification.To address the above problems,this paper uses the visual attention mechanism in deep learning to model the features of micro changes in the scanned brain images of AD subjects,and tries to improve the traditional residual network model to enhance its ability to extract features.The main contents and work of this paper are as follows:(1)We designed two deep residual network(Res Net)algorithms based on attentional mechanisms with channel shuffling for AD levels classification.For the imaging characteristics of MRI images,it is hoped that the attention mechanism can be introduced into the traditional Res Net residual blocks to get more detailed information focusing on image changes,and thus suppress other useless information,to improve the discriminative ability of the network.Therefore,the Contextual Transformer(Co T),Group Convolution,and Channel Shuffle mechanism are considered being introduced into the traditional Res Net residual blocks.The Co T module is used to replace the 3×3convolution in the residual block to enhance the feature extraction capability of the residual block,and the group convolution and channel shuffle mechanism are used to balance the relationship between the number of parameters and the complexity of the algorithm.One of the proposed algorithms is based on the improvement of the 2D Res Net-18 model by replacing the 3×3 convolutional layers in the basic residual block of 2D Res Net-18 with Co T modules to enhance the feature extraction capability of the basic residual block;Second algorithm is based on 2D Res Net-50 by replacing the 1×1 convolutional layer in the Res Net-50 residual bottleneck block with a 1×1group convolution,and then reorganizing the output features of the group convolution using the Channel Shuffle mechanism in order to enhance the communication of the output features from different groups,and also replacing the 3×3 convolutional layer with a Co T module.The CCS-Res Net-50 algorithm has a classification accuracy of96.23% in the AD and MCI classification tasks of the ADNI slice dataset,which is 0.93%higher than the comparison algorithm.(2)The two improved 3D convolutional neural network algorithms were designed.The 3D sample data in the ADNI database is sliced and then applied to the 2D convolutional neural network for classification,whose slicing operation will lose the3 D information of the original 3D data;second,it is more reasonable to classify the levels of Alzheimer’s disease subject,and it can make predictions for each subject’s brain scan images.Therefore,a 3D convolutional neural network model was used to classify AD.The overall architecture of the algorithm is to use 3D PET and MRI sample data as input,first preprocess them separately in different ways,then extract the respective features using the proposed 3D convolutional neural network model,and finally fuse the features and input the fused features to a multilayer perception for classification prediction.This paper proposes two 3D convolutional neural networks,following the design benchmarks of 2D Res Net-18 and 2D Res Net-50 model architectures,and reasonably transforming them into 3D network models,namely 3D Co T-Res Net-18 and 3D CCS-Res Net-50 models.Finally,the classification effects of the two algorithms were examined on the multimodal sample data of AD: HC,AD: MCI,MCI: HC and AD: MCI: HC on four tasks,respectively,using the PET and MRI multimodal data in 3D NIFTI format collated from ADNI data as input.Experiments show that 3D Co T-Res Net-18 and 3D CCS-Res Net-50 achieve better results on ADNI dataset.Especially in the multi-modal task of AD: MCI: HC,the 3D Co T-Res Net-18 algorithm achieved a recognition accuracy of 77.60%,which was 9.86% higher than the comparison algorithm.