Parameter-Efficient Fine-Tuning of SAM for Label-Efficient Interactive Medical Image Segmentation

抄録

Accurate segmentation of medical images often requires substantial expert annotations, making the labeling process costly and time-consuming. In this study, we present a parameter-efficient fine-tuning strategy for the Segment Anything Model (SAM) to enable label-efficient interactive segmentation in medical imaging. Unlike conventional full-model fine-tuning, our approach utilizes lightweight adaptation techniques—spatial prior adapter (SPA)—that can be trained with only a small fraction of additional parameters. By leveraging SAM’s powerful, general-purpose segmentation capabilities and tailoring it to domain-specific characteristics through minimal parameter updates, we achieve high-quality interactive segmentation results with significantly reduced labeling effort. Experimental evaluations on two medical imaging datasets demonstrate that our method significantly enhances SAM’s adaptability to medical images. Even with a limited number of labels, it achieves superior performance compared to other specialized interactive segmentation models.