Reproduce “HPR-SAM: Hierarchical Probabilistic Representation Learning for Prompt-free SAM-based Medical Image Segmentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yingzhen Hu, Yiheng Zhong, Keying Zhu, Zimu Zhang, Zihan Ye, Sifan Song, Jionglong Su, Xiaofeng Liu
HPR-SAM introduces a hierarchical probabilistic framework for prompt-free medical image segmentation, learning expressive anatomical representations to overcome the limitations of deterministic prototypes. The approach achieved mean Dice coefficients of 85.09% on Synapse CT, 84.65% on LA (few-shot), and 81.26% on PROMISE12 (few-shot), surpassing prior SAM-based methods.
Key insights
- Learning richer, probabilistic anatomical representations (global priors, intra-structure diversity, local reliability) fundamentally enhances the quality of automatically generated prompts for SAM.
- Modeling anatomical properties as distributions or multi-component mixtures provides a more nuanced understanding, capturing inherent uncertainty and morphological variations missed by deterministic approaches.
- Hierarchical integration of predictions derived from these distinct probabilistic perspectives (global, diverse, reliable) leads to comprehensive improvements in both overall segmentation accuracy and precise boundary localization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.