Evaluating Segmentation Accuracy with Diverse Prompt Strategies in Medsam

Segment Anything Model (SAM) shows robustness in prompt-based segmentation of natural images. However, adapting SAM to the medical domain as a foundation model, MedSAM, presents unique challenges arising from clinical settings. This study evaluates MedSAM's performance across five medical imaging modalities (MRI, CT, ultrasound, endoscopy, and fundus IR) using four different prompting strategies: standard bounding boxes (SBB), rotated bounding boxes (RBB), contour points (CP), and simulated human annotations (SHA). MedSAM's performance is assessed under precise and perturbed conditions for each prompt type to evaluate its segmentation capabilities comprehensively. We empirically evaluate MedSAM's abilities and adaptability by experimenting with zero-shot learning and fine-tuning approaches. The results indicate that performance varies markedly with prompt type, accuracy, and imaging modality. The study concludes that MedSAM is sensitive to prompt precision and anatomical complexity. Fine-tuning improves performance, but further refinements are necessary to make it robust in various clinical settings. This work highlights the need for prompt engineering and provides practical insights to enhance segmentation accuracy in SAM-like foundation models.