Co‐teacher‐guided pseudo label supervision: A semi‐supervised learning framework for muscle and adipose tissue segmentation on chest CT scans

Abstract Automatic segmentation of muscle and adipose tissue in chest CT scans is essential for the diagnosis, treatment, and prognosis of various diseases. However, this task is hindered by the limited availability of annotated data. This study proposes a novel semi‐supervised learning framework, co‐teacher‐guided pseudo‐label supervision (CTGP), to address this challenge. CTGP combines co‐training and the Mean‐Teacher strategy, where predictions generated by teacher models are filtered and utilized as high‐quality pseudo‐labels to train the other student models, thus facilitating co‐training. Additionally, a medical image‐specific augmentation method, MIAugment, is introduced to better adapt to the unique characteristics of medical data. Experiments on a real‐life clinical dataset demonstrate that CTGP achieves high segmentation accuracy with minimal labelled data. Using only 10% of labelled data, the framework achieves a mean Dice Similarity Coefficient of 90.03% for four tissue types, a decrease of just 2.85% compared to fully supervised learning. This approach provides a promising solution for automated muscle and adipose tissue segmentation with limited annotations.