Dual-Decoder Consistency via Pseudo-Labels Guided Data Augmentation for Semi-Supervised Medical Image Segmentation

Though supervised learning gains impressive success, the acquisition of indispensable large-scale labeled datasets are often impractical in biomedical imaging partially due to expensive costs and lengthy annotations done by experienced radiologists. Semi-supervised learning has been shown to be an effective way to address this limitation by leveraging useful information from unlabeled datasets. In this paper, we present a new semi-supervised learning method referred to as Dual-Decoder Consistency via Pseudo-Labels Guided Data Augmentation (DCPA) for medical image segmentation. We devise a consistency regularization to improve the semi-supervised learning. Specifically, to promote consistent representations during the training process, we use different decoders for student and teachers networks while maintain the same encoder. Moreover, to learn from unlabeled data, we create pseudo-labels generated by the teacher networks and augment the training data with the pseudo-labels. The two techniques contribute to the improved performance of the proposed method. We evaluate the performance of the proposed method on three representative medical image segmentation datasets. Extensive comparisons to the state-of-the-art medical image segmentation methods were carried out under typical scenarios with 10% and 20% labeled data. Experimental outcomes demonstrate that our method consistently outperforms state-of-the-art semi-supervised medical image segmentation methods over the three semi-supervised settings. Furthermore, to explore the performance of proposed method under extreme condition, we conduct experiments with only 5% labeled data. The results further verify the superior performance of the proposed method. Source code is publicly online at https://github.com/BinYCn/DCPA.git.