Multi-transformation Consistency Regularization for Semi-supervised Medical Image Segmentation

Segmentation algorithms based on supervised learning have achieved good results, but supervised learning algorithms require a large amount of labeled data for training. The scarcity of labeled data has greatly limited the development of medical image segmentation. In this paper, we propose a novel semi-supervised segmentation method, where a multi-transformation consistency regularization model is established for medical image segmentation. The model is trained with abundant unlabeled data of medical images. Our semi-supervised framework consists of two modules: supervised training and transformation consistency training. A variety of complex spatial and non-spatial transformations and perturbations are used for consistency regularization, including affine transformation, Euclidean transformation, similarity transformation, color transformation and virtual adversarial training. Multiple transformations are carried on unlabeled data randomly during the training process, and the network is trained by optimizing the supervised loss of labeled data and the multi-transformation consistency regularization loss of unlabeled data. Experimental results on two public datasets show that our method improves the performance in terms of segmentation accuracy to a great extent and achieves state-of-the-art results.