Reproduce “Balancing Multi-Target Semi-Supervised Medical Image Segmentation with Collaborative Generalist and Specialists” on OpenResearch

About this paper

You Wang, Zekun Li, Lei Qi, Qian Yu, Yinghuan Shi, Yang Gao

The Collaborative Generalist and Specialists (CGS) framework addresses scale imbalance in multi-target semi-supervised medical image segmentation, where organs of vastly different sizes coexist, by employing a dual-branch architecture and refining pseudo-labels. The method consistently outperforms state-of-the-art techniques across three medical imaging datasets, achieving notable accuracy gains for small and challenging organs while maintaining efficient inference.

Key insights

• Scale imbalance, characterized by the coexistence of vastly different-sized anatomical structures, is a critical yet underestimated challenge in multi-target semi-supervised medical image segmentation.
• Training individual specialized segmentation heads with a re-contextualized label definition (target, remaining classes, background) effectively balances the training signal across all classes, preventing larger targets from monopolizing optimization.
• Exploiting the inherent logical consistency between predictions from multiple specialized segmentation heads allows for robust detection and correction of erroneous pseudo-labels, significantly improving the quality of self-generated supervision.

Summary by alphaXiv.