CER-SAM: a coarse-guided edge refinement network for skin lesion segmentation with SAM and its derivatives

Abstract

Skin lesion image segmentation plays an important role in computer-aided diagnosis. In recent years, the Segment Anything Model (SAM) and its derivatives have demonstrated clear advantages in medical image segmentation due to their strong feature representation capability and robust cross-domain generalization, achieving stable segmentation performance with limited task-specific design. Nevertheless, when applied to skin lesion images with blurred boundaries, irregular morphology, and low contrast against surrounding skin, SAM-based models often exhibit unstable local structures, fragmented boundaries, and spurious foreground predictions, which collectively degrade the accuracy and robustness of segmentation results. To address these issues, we propose a coarseguided edge refinement network, named CER-SAM, which is applicable to SAM and its derivatives. CER-SAM jointly feeds coarse segmentation masks generated by foundation models and original dermoscopic images into the network, and improves overall segmentation performance through edge-neighborhood enhancement and multiscale feature fusion. In the encoding stage, edge-neighborhood attention guides multi-scale feature extraction, enabling the network to focus on high-uncertainty regions and capture local lesion structures. In the decoding stage, enhancement of the final feature maps further ensures boundary continuity and accuracy. Experimental results on the PH2 dataset demonstrate that CER-SAM consistently improves the segmentation performance of SAM and its derivatives, significantly enhancing boundary quality and overall accuracy, while exhibiting strong generality and robustness.