ASFOSDA: First active open-set domain adaptation for medical image segmentation

Deep learning models in medical image segmentation often fail due to domain shifts—differences in scanners, protocols, or populations. Active domain adaptation helps by querying a few target samples for retraining. However, existing methods assume source and target share the same classes (closed-set) and require access to source data, which is often restricted by privacy regulations. A new paper from Jin Yang et al., accepted at MICCAI 2026, introduces Active Source-free Open-set Domain Adaptation (ASFOSDA), the first method to tackle all these challenges simultaneously. It operates without source data and can handle unknown classes (e.g., new pathologies) appearing in the target domain.

The core innovation combines two metrics for sample selection: Class-aware Decomposed Uncertainty (CDU) to measure aleatoric (data) and epistemic (model) uncertainty via test-time augmentation, and Class-agnostic Prototype Discrepancy (CPD) to capture cross-domain and self-domain diversity. Selected samples are labeled by an oracle; unselected ones get high-quality pseudo labels via a target-refined self-training strategy. This semi-supervised setup boosts adaptation performance. Evaluated on volumetric medical image segmentation tasks, ASFOSDA outperformed state-of-the-art adaptation methods, proving effective for real-world clinical scenarios where data privacy and unknown classes are common.