Severe cross-modality domain shift poses a major obstacle to reliable medical image segmentation, especially when adapting from heterogeneous multi-source ultrasound to an unlabeled dermoscopy target. Such shifts introduce pronounced appearance variation and scale mismatch, which can aggravate semantic drift, boundary degradation, and miscalibration in unsupervised adaptation. We propose Pathology-Anchored Semantic Alignment and Cross-Scale Collaboration (PA-SACC), a closed-loop framework for multi-source unsupervised cross-modality lesion segmentation. PA-SACC introduces pathology-anchored semantic cues as weak anchors to stabilize alignment across domains, leverages bidirectional cross-scale collaboration to preserve global-to-local structural coherence under resolution mismatch, and employs uncertainty-aware, confidence-adaptive filtering to suppress artifact-induced noisy pseudo-labels and improve probabilistic calibration. Adaptation is performed only on the unlabeled ISIC training split, while the ISIC test split is reserved exclusively for evaluation. PA-SACC achieves a Dice of 0.9085 ± 0.0057 on ISIC and 0.8976 ± 0.0066 under strict zero-shot evaluation on PH2, and further shows promising performance under additional out-of-distribution endoscopy evaluations. Ablation studies further support the complementary roles of the proposed components.
Gou et al. (Wed,) studied this question.