Research On Prototype Domain Adaptation Algorithm For Alzheimer’s Disease Detection

Alzheimer’s disease(AD)is a neurodegenerative disease characterized by memory impairment and cognitive decline.Structural Magnetic Resonance Imaging(s MRI)can clearly reveal the brain tissue structural lesions caused by AD,making it possible to assist in the diagnosis of AD.Deep learning technology can automatically extract features from massive data and achieve classification recognition.However,most methods ignore two problems:(1)domain shift problem caused by factors such as collection equipment,software protocols,and environment;(2)medical data often faces restrictions such as privacy protection and data security,making it difficult to access directly.To address the first problem,this thesis proposes a prototype-guided multi-scale domain adaptation framework.Firstly,multi-scale feature extraction and fusion are achieved using 3D convolution and self-attention mechanism,and spatial attention mechanism is added to enhance the localization of sensitive areas.Attention consistency loss is applied to achieve semantic information sharing between domains.Secondly,a new prototype maximum density divergence loss is proposed,which can align domain features and enhance the constraint on feature outlier samples.Finally,two different domain discriminators are used for cooperative training,and weight difference loss is used to prevent overfitting of a single domain discriminator.Through classification experiments on ADNI database using AD,MCI,and Cognitively Normal(CN)data,the proposed method achieves accuracies of 92.11%,76.01%,and 82.37% for AD vs.CN,AD vs.MCI,and MCI vs.CN tasks,respectively.A large number of experimental results show that the proposed method is superior to existing domain adaptation methods in the case of data domain shift.This thesis proposes a Source-free domain adaptation algorithm based on a classbalanced multi-center prototype strategy for the second problem.Firstly,a feature extraction module based on 3D convolution and hierarchical Transformer Encoder is designed to achieve multi-scale feature and fusion.Secondly,the class-balanced sampling based on Top-K and the intra-class multicentric prototype method based on K-Means clustering are used to effectively represent class features,and corresponding pseudo-labels are assigned to samples based on the distance between prototypes and samples.Finally,a noise-robust loss function is used to constrain both static and dynamic pseudo-labels for model forward optimization.Ultimately,the proposed algorithm achieves accuracies of90.46%,74.32%,and 82.69% on three classification tasks,outperforming both supervised learning and unsupervised domain adaptation methods.Compared to domain adaptation methods that can directly access the source domain,it still achieves competitive results.In summary,the methods proposed in this thesis can effectively address the domain shift and privacy security issues in MRI data.By addressing these two problems,the method has improved the model’s generalization ability and increased the efficiency and quality of assisted diagnosis,thus alleviating the problem of data isolation.