CAU: A Consensus Model of Augmented Unlabeled Data for Medical Image Segmentation

Abstract

Medical image segmentation plays an important role in medical diagnosis and treatment. However, medical image data are more expensive and time-consuming to obtain than ordinary image data. In this paper, we propose a novel semi-supervised method named CAU for medical image segmentation, which can easily use Convolutional Neural Networks (CNNs) to segment 2D images. The network learns through a combination of common supervision losses for labeled data and losses for unlabeled data. Specifically, we augment the unlabeled data strongly and weakly and send them to the student model and the teacher model respectively. We take full advantage of unlabeled data learning through a novel combination of minimizing the difference between network predictions for different data augmentation processing scenarios and using an unsupervised loss of min-entropy on the outputs of the two networks. In order to improve the regularization effect, we use the teacher-student model to optimize the teacher model by averaging the student model weights. Experiments show that our method in labeled data experiments with 5%, 10%,35% and 50% labeled data on Automated Cardiac Diagnosis Challeng(ACDC) dataset exceeds the fully supervised algorithm using the same amount of data and the existing popular semi-supervised learning algorithms 1.922% ~ 4.451%(Dice),0.841 ~ 5.031(Asd) and 0.06 ~ 1.116(HD95), respectively, and the Dice index exceeds the fully supervised algorithm 0.019% with 50% labeled data, which verifies its effectiveness in medical image segmentation.