The Segment Anything Model (SAM), a foundation model for promptable image segmentation, has shown strong performance on zero-shot natural image segmentation tasks. However, due to the significant differences between the natural and medical image domains, as well as the subtle lesions and blurred boundaries inherent in medical images, SAM suffers a marked performance drop when applied to medical images. To address these challenges, we propose DB-SAM, a domain-adaptive and boundary-enhanced framework for few-shot medical image segmentation. To better adapt to medical image segmentation, we fine-tune the image encoder using a hybrid feature representation that incorporates low-level features. During the fine-tuning process, a nonlinear mapping method is employed to update a small portion of weights, thereby capturing rich features in medical images. Furthermore, to supplement the multi-scale and boundary features required for medical image segmentation, a multi-scale feature extractor and a boundary-aware module are designed. The multi-scale feature extractor enhances the representation of local receptive field features, providing SAM with rich local multi-scale contextual information. The boundary-aware module leverages limited input data to capture enhanced boundary-representative features by encoding boundary and texture features. Experimental results on three datasets from different modalities demonstrate that our DB-SAM outperforms the state-of-the-arts methods, and remarkably, even with only one labeled training sample, it consistently surpasses various SAM and CNN-based methods. The code is publicly available at https://github.com/ZZY010411/DB-SAM.
Wen et al. (Wed,) studied this question.