SF-UniDA: Unlocking abdominal multi-organ segmentation via source-free universal domain adaptation

Universal domain adaptation is essential for medical image segmentation across diverse imaging modalities and datasets, where labeled source data and unlabeled target data exhibit varying degrees of class overlap. In this paper, we present a novel framework that addresses abdominal segmentation under closed, partial, open, and universal adaptation settings. The framework is built upon a Source-Free Unsupervised Domain Adaptation baseline and introduces LEAD (Label-Efficient Adaptive Decomposition) for segmentation, which generates reliable pseudo-labels using confidence-based unknown class modeling. The framework is enhanced with anatomical zero-shot learning, a target-aware feature alignment module, and an uncertainty-guided curriculum strategy. Extensive experiments on four benchmark datasets – BTCV, CHAOS, FLARE22, and Synapse – demonstrate the superiority of our approach in closed-set, partial-set, open-set, and universal-set scenarios, showing strong generalization to novel organs in the target domain through quantitative and qualitative analysis.SF-UniDA achieves strong performance across all UniDA settings: 0.9172/0.1450 (Dice/ASSD) in closed-set, 0.9047/0.1820 in partial-set, 0.7850/0.2653 in open-set, and 0.7860/0.2632 in universal-set. These results establish SF-UniDA as a state-of-the-art UniDA method for medical segmentation, offering robust generalization across domains and label sets with minimal supervision.