DTSC-Net: Semi-supervised 3D Biomedical Image Segmentation through Dual-Teacher Simplified Consistency.

Xiaogen Zhou,Zhiqiang Li,Tong Tong

DOI: https://doi.org/10.1109/BIBM55620.2022.9995371

2022-01-01

Abstract:Although those deep learning networks have achieved remarkable performance in various biomedical image segmentation tasks, they rely on massive labeled data for training, which is time-consuming to acquire. To alleviate this challenging issue, semi-supervised learning (SSL) has shown the potential to simultaneously leverage between abundant unlabeled samples and limited labeled data. In this work, we propose a new SSL approach for left atrial and liver tumor segmentation called DTSC-Net. First, two teacher networks and one student network are optimized together through a supervised loss and two unsupervised consistency losses. Furthermore, pseudo labels are yielded by using the first teacher network on unlabeled data. In addition, the weights of the second teacher network are an exponential moving average of the student network’s weights. Moreover, the student network learns from the two teacher networks by minimizing a supervised segmentation loss, a pseudo-labeling unsupervised consistency (PUC) loss, and a dual-teacher unsupervised consistency (DUC) loss with respect to the targets of the two teacher networks. Experiment results show that our approach achieves state-of-the-art semi-supervised segmentation performance on the LA2018 and LiTS2017 datasets.