Vision Foundation Model for MRI Segmentation Through Training-free Few-shot Adaptation

Motivation: Vision foundation models have demonstrated impressive capabilities in natural image segmentation. However, their application to MRI remains challenging due to the lack of domain-specific adaptation. Goal(s): Introduce an adaptation method that effectively adapts vision foundation models for MRI image segmentation, without model training or fine-tuning. Approach: We propose a training-free adaptation pipeline that emulates human analogical thinking, leveraging few-shot examples to rapidly adapt the Segment Anything Model (SAM) to medical imaging. Results: The proposed approach demonstrates superior performance over both the original SAM and its advanced variant, SAM2, in terms of quantitative metrics and qualitative analysis. Impact: Our training-free adaptation method circumvents the need for laborious data collection and labeling, providing a generalizable solution for applying vision foundation models to medical image segmentation.