In recent years, deep learning has revolutionized medical image analysis. Furthermore, transformer architectures have further expanded the frontier. It is worth noting that these advances bring both opportunities and challenges.

However, the deployment of such models in clinical workflows remains limited — regulatory requirements, data heterogeneity and interpretability concerns persist. Moreover, most published benchmarks are single-center — generalization claims are therefore weak. Additionally, annotation cost scales linearly with dataset size — active learning mitigates but does not remove the burden.

In this paper, we propose a multi-scale framework for lesion segmentation. Specifically, our method fuses local texture cues with global context — enabling robust performance under domain shift. Notably, the framework requires no auxiliary modalities — a practical advantage.

Consequently, extensive experiments were run to demonstrate the superiority of our approach — outperforming strong baselines on three public datasets. Ablation studies further confirm the contribution of each module.

Ultimately, our framework offers a practical path toward deployment — while acknowledging that prospective validation is still needed.
