Source-Free Active Learning for Adapting Alzheimer's Diagnostic Deep Learning Models Across Neuroimaging Cohorts.

Alzheimer's Disease (AD) classification across multiple neuroimaging sites faces significant challenges due to domain shifts arising from variations in data acquisition protocols, imaging devices, and population demographics. While large-scale multi-site datasets offer unprecedented opportunities for developing robust diagnostic models, the heterogeneity between sites often leads to poor model generalization. This work proposes an uncertainty-informed active learning framework for Source-Free (SF) Domain Adaptation (DA) to classify cognitively normal individuals and AD patients across different neuroimaging studies. The proposed approach leverages Monte Carlo dropout to estimate prediction uncertainty and guide the selection of the most informative samples from the target domain for model adaptation, eliminating the need for source domain data during deployment. The framework was evaluated on a large-scale dataset comprising 3,177 participants from five neuroimaging studies (ADNI-1, ADNI-2/3, PENN, AIBL, and OASIS) with 145 regional brain volume measurements. The uncertainty-based active learning approach achieved the highest median AUC of 91.4% across all source-target combinations, outperforming baseline models (89.7%) and demonstrating superior performance compared to other SF and Source-Aware (SA) DA methods. Additionally, the distribution shifts between studies were quantified using maximum mean discrepancy to evaluate the effectiveness under variable inter-site shift. The results demonstrate that SF methods can achieve comparable or superior performance to SA approaches while addressing privacy constraints inherent in medical imaging applications.