Enhancing Alzheimer's Disease Diagnosis via Open Set Domain Adaptation-Based Representation Learning on the Hypersphere

Structural magnetic resonance imaging (sMRI) is commonly employed for Alzheimer's disease (AD) diagnosis. However, sMRI datasets often suffer from domain shifts due to variations in scanning protocols, and real-world clinical scenarios frequently involve unknown categories, presenting challenges for conventional domain adaptation methods. To address these issues, we propose a novel Open Set Domain Adaptation (OSDA) framework that not only detects unknown class but also improves AD diagnosis. The method integrates demographic information as auxiliary features alongside sMRI-derived features to enhance the recognition of high-dimensional data. Features are projected onto a hyperspherical space via a spherical variational auto-encoder (S-VAE), enabling efficient feature representations. Domain alignment is achieved using a von Mises-Fisher kernel-based maximum mean discrepancy function, while A-softmax loss enables the model learn angularly discriminative feature representations. Additionally, an entropy minimization strategy ensures greater prediction confidence in the target domain. Experimental results show that the proposed method achieves OS of 92.29%, 85.28%, and 80.37% on AD vs. CN vs. Unk, MCl vs. CN vs. Unk, and AD vs. MCl vs. Unk tasks, respectively. This framework extends the applicability of OSDA techniques to medical imaging, providing enhanced diagnostic accuracy and enabling more robust AD detection.