Source-free unsupervised domain adaptation (SFUDA) is crucial for medical image segmentation, enabling models trained on labeled source data to generalize to unlabeled target domains without accessing source data, but SFUDA faces challenges such as domain shift and a lack of reliable supervision. To address these issues, this paper proposes a novel SFUDA framework, EOPCNet, which integrates entropy optimization and multi-perspective conflict analysis. The framework incorporates multi-perspective feature extraction to capture diverse visual cues, alongside a pseudo-label generation strategy that leverages prediction conflicts across multiple perspectives to generate reliable supervision. Moreover, an entropy-optimized loss function is employed to focus on uncertain regions, enhancing segmentation precision. In addition, a weak-strong augmented mean teacher mechanism stabilizes training by transferring knowledge from a teacher model (processing weakly augmented data) to a student model (learning from strongly augmented data). Finally, evaluation experiments on several public fundus image datasets demonstrate that the EOPCNet outperforms other state-of-the-art SFUDA approaches, achieving segmentation performance close to the upper bounds of supervised learning, particularly in balancing region overlap and boundary accuracy. Ablation studies also confirm that entropy optimization and multi-perspective conflict-based pseudo-labels significantly enhance the performance of EOPCNet, making it effective for cross-domain medical image segmentation without accessing source data. Our code is available at: https://github.com/4Ezzhh/EOPCNet .
Chen et al. (Sun,) studied this question.