文献类型：期刊文献

英文题名：DermoSAM: Lightweight boundary-aware SAM distillation for dermoscopy lesion segmentation

作者：Che, Zhiyan[1];Chen, Ruyun[1];Li, Yonggui[1];Chen, Hao[1]

机构：[1]Guangdong Ocean Univ, Sch Math & Comp, Zhanjiang, Guangdong, Peoples R China

年份：2027

卷号：334

外文期刊名：EXPERT SYSTEMS WITH APPLICATIONS

收录：SCI-EXPANDED(收录号：WOS:001892543900001)、、WOS

语种：英文

外文关键词：Dermoscopic image segmentation; Knowledge distillation; SAM; Medical image processing

外文摘要：Accurate dermoscopic lesion segmentation requires high region agreement and reliable boundary delineation, while remaining efficient for deployment. Promptable foundation models have improved general segmentation, but dermoscopy workflows still often rely on manual prompts at inference. This reliance limits scalability and introduces operator variation. We present DermoSAM, a SAM framework for dermoscopic lesion segmentation that performs fully automatic inference without manual prompts and uses 10.41M total inference parameters, of which 6.35M are trainable. We evaluate DermoSAM on three in-domain benchmarks, ISIC2016, ISIC2017, and ISIC2018, and we further test zero-shot transfer on HAM10000 and PH2. On the ISIC test sets, DermoSAM achieves strong overlap and good boundary accuracy without user interaction. It also remains competitive under dataset shift in the zero-shot setting. DermoSAM adapts the SAM decoding interface to dermoscopic images through two internal components. The Boundary-aware Skin Prior Generator produces dense spatial priors that capture lesion extent and boundary evidence, and it reduces the influence of common artifacts. The Guided Region Tokenizer summarizes global context into a small set of region tokens. Token generation is guided by the dense priors so that global context supports local boundary cues during decoding. These designs enable accurate lesion segmentation with stable boundaries in a fully automatic workflow.