DASAM: Medical Domain Adaptation of Segment Anything Model Without Further Pre-Training

Deep learning has revolutionized medical image segmentation, significantly enhancing accuracy and automation in identifying anatomical structures and pathologies crucial for diagnosis and treatment planning. While powerful vision foundation models pretrained on extensive natural image collections have transformed computer vision, adapting these models to medical image segmentation presents unique challenges due to limited data availability, diverse imaging protocols, ethical concerns, and task complexity. To this end, we introduce DASAM, a novel medical image segmentation framework that addresses these challenges through innovative lightweight domain adaptation mechanisms. DASAM employs a medical adapter and simplified prompt-tuning protocol to effectively adapt vision foundation models to medical imaging tasks without requiring extensive domain-specific pretraining. Our comprehensive evaluation demonstrates DASAM's superior performance, adaptability, and generalizability across 8 datasets and 5 imaging modalities, outperforming 7 baselines while reducing reliance on large-scale medical datasets and costly medical pretraining.