Coupling Deep Deformable Registration with Contextual Refinement for Semi-Supervised Medical Image Segmentation

Image segmentation is an important prerequisite of computer-aided diagnosis which has been applied in a wide range of clinical applications. Current learning-based methods mostly rely on sufficient annotated datasets which is expensive and time-consuming. In this study, we develop a semi-supervised learning paradigm integrating contextual refinement into deformable registration-based segmentation processes. By introducing deformable atlas prior, our method is capable to segment the image with no well-defined relation between regions and pixels intensities. A contextual refinement segmentation network is appended to further constrain unreasonable results. Inheriting from the merits of both prior knowledge and deep representation, our approach achieves a more satisfying performance than the state-of-the-art methods qualitatively and quantitatively on multiple medical image datasets.