Uncertainty-guided cross teaching semi-supervised framework for histopathology image segmentation with curriculum self-training

Rui Xu, Nan Zhou, Siyang Feng, Zhenbing Liu, Huahu Deng, Yajun An, Jian Li, Chu Han, Zaiyi Liu, Rushi Lan, Cheng Lu, Xipeng Pan

Abstract: Highlights•We propose an uncertainty-guided cross teaching consistency learning framework (UT) that effectively extracts knowledge from noisy pseudo-labels by quantifying disagreement between two peer models.•Built on our UT, to alleviate the performance degradation caused by incorrect pseudo labels, we further propose a curriculum-based self-training framework named UTCS that performs selective re-training via prioritizing reliable images based on holistic prediction stability in the entire training course.•Utilizing the advantages of consistency and self-training, we propose an efficient semi-supervised histopathology image segmentation framework that outperforms several state-of-the-art methods on two publicly available datasets.