Reproduce “Self-Loop Uncertainty: A Novel Pseudo-Label for Semi-Supervised Medical Image Segmentation” on OpenResearch

About this paper

Yuexiang Li, Jiawei Chen, Xinpeng Xie, Kai Ma, Yefeng Zheng

Researchers from Tencent Jarvis Lab developed a semi-supervised learning framework for medical image segmentation that leverages 'self-loop uncertainty' to efficiently generate robust pseudo-labels from unlabeled data. The method enables segmentation performance comparable to fully-supervised approaches while significantly reducing the need for manual annotations, achieving an F1 score of 79.10% on MoNuSeg with only 50% labeled data.

Key insights

• A self-supervised task, specifically Jigsaw puzzles, can be leveraged to generate diverse "states" of a single model's encoder, effectively approximating the uncertainty estimation of a multi-model ensemble.
• Recurrent optimization of the encoder within a self-loop, guided by a self-supervised loss, allows for efficient generation of robust pseudo-labels for unlabeled data without needing to train or infer with multiple distinct models.
• Weighting the individual segmentation predictions based on the confidence of the self-supervised task (e.g., lower Jigsaw puzzle loss) improves the reliability and quality of the final self-loop uncertainty pseudo-label.

Summary by alphaXiv.