Reproduce “Bi-Level Collaborative Learning for Few-Shot Scribble-Supervised Medical Image Segmentation” on OpenResearch

About this paper

Xiang-Xiang Su, Yufan Ye, Yihang Zheng, Min Gan, Guang-Yong Chen

A bi-level collaborative learning framework (BiSCL) tackles medical image segmentation in few-shot, scribble-supervised settings by dynamically aligning learnable superpixel generation with segmentation learning. This method enhances pseudo-label reliability and boosts performance, achieving up to a 5.6 percentage point improvement in Mean Dice score on ACDC and Prostate datasets compared to existing baselines.

Key insights

• Dynamically aligning superpixel generation with the segmentation task via bi-level optimization significantly improves the relevance of structural priors and overall segmentation accuracy.
• An adaptive filtering mechanism, guided by spatial priors and expansion ratios, is crucial for mitigating pseudo-label noise amplification during superpixel-based label diffusion in few-shot, weakly supervised settings.
• Concentrating superpixel capacity in central image regions through non-uniform initialization, combined with semantic feedback, enables the generation of superpixels that are robustly aligned with anatomical structures.

Summary by alphaXiv.