Reproduce “SynPo: Boosting Training-Free Few-Shot Medical Segmentation via High-Quality Negative Prompts” on OpenResearch

About this paper

Yufei Liu, Haoke Xiao, Jiaxing Chai, Yongcun Zhang, Rong Wang, Zijie Meng, Zhiming Luo

SynPo enhances training-free few-shot medical image segmentation by integrating DINOv2 and SAM-ViT features for robust confidence maps and generating informative negative prompts from a modeled distribution. The method achieves competitive Dice scores, outperforming other training-free approaches on both CT and MRI medical imaging datasets.

Key insights

• Fusing features from different LVMs (e.g., DINOv2's semantic understanding and SAM-ViT's spatial awareness) creates more robust and reliable confidence maps for medical image segmentation.
• Strategic selection of negative prompts within a specific confidence range, rather than trivially distant background pixels, provides more meaningful guidance to LVMs, enhancing boundary delineation and preventing over-segmentation.
• High-quality prompt engineering can significantly boost the performance of training-free LVMs in specialized domains, making them competitive with or superior to training-based methods while reducing computational and annotation costs.

Summary by alphaXiv.