Source-free Domain Adaptation via Multicentric Prototype for Alzheimer's Disease Detection

The use of deep learning and transfer learning techniques for the early diagnosis of Alzheimer's Disease (AD) is of great significance for delaying its development. In the real world, due to different scanners, scanning protocols, and subject cohorts, structural Magnetic Resonance Imaging (MRI) often has the problem of domain shift. Conventional Domain Adaption (DA) methods need to access both source domain and target domain for feature alignment to achieve the generalization in target domain. However, medical image data usually need to concern data privacy and security, the source domain always cannot be accessed. Based on the above situation, we propose a Source-Free Domain Adaptation (SFDA) framework for AD detection. Firstly, we design a feature extraction module combining the advantages of CNN and Transformer, then we use the class-balanced multicentric prototype method to obtain robust pseudo labels. Finally, noise-robust loss function which based on Determinant based Mutual Information (DMI) is used to optimize the model. On the ADNI dataset, our method achieved 90.79%, 75.00% and 80.13% accuracy on the AD vs. CN, AD vs. MCI and MCI vs. CN tasks, respectively. Compared with the supervised learning methods, DA methods which can access to source domain and SFDA methods, our method achieves competitive results.