Reproduce “PMT: Progressive Mean Teacher via Exploring Temporal Consistency for Semi-Supervised Medical Image Segmentation” on OpenResearch

About this paper

Ning Gao, Sanping Zhou, Le Wang, Nanning Zheng

The Progressive Mean Teacher (PMT) framework, developed at Xi’an Jiaotong University, introduces a semi-supervised approach for medical image segmentation by utilizing temporal consistency to foster diversity among co-training Mean Teacher architectures. It achieved state-of-the-art performance, improving Dice scores by up to 1.45% on the LA dataset (5% labeled data) and 7.17% on the Pancreas-NIH dataset (10% labeled data) compared to previous methods.

Key insights

• A progressive design that leverages temporal consistency through alternating model updates effectively creates and sustains diversity and performance gaps between co-training networks.
• Explicit pseudo-label filtering based on model performance, combined with sharpening, is crucial for generating high-fidelity supervisory signals for unlabeled data, preventing the propagation of errors.
• Targeted alignment of lagging models in regions of prediction discrepancy can accelerate learning and prevent unfavorable divergence among the co-training networks.

Summary by alphaXiv.