Weighted Semi-supervised MRI Segmentation with Paired Consistency and Entropy Minimization
Ao Ma, Jieyan Liu, Yimin Zhou, Ke Lü · 2022 IEEE Asia-Pacific Conference on Image Processing, Electronics and Computers (IPEC) · 2022
U-net is very popular in medical image segmentation. However, the labels of training sets are always lacking for supervised learning. Thus, semi-supervised learning which utilizes unlabeled data shows advantages when dealing with such label-scarce situations. In this paper, we utilize U-net to tackle the challenge of segmenting MR images in semi-supervised setting. Specifically, we introduce paired consistency (PC) as data augmentation method. Also, we propose an entropy minimization (EM) term which is flexible and easy for calculation to boost the semi-supervised learning. Specifically, experiments show that our semi-supervised framework with PC and EM outperforms supervised U-net by %0.3 and %0.5 on OASIS and Hammers datasets, respectively. Also, entropy minimization accelerates the convergence of training, namely, reducing nearly %50 of training time, which is beneficial to real-world application.