C2FC: Coarse-to-fine Contour-Based Method for Interactive Medical Image Segmentation

Existing contour-based methods for interactive segmentation of medical images have achieved great success. However, these methods neglect the large shape error between the initial contour and the ground truth. The details of the contour are focused on too early, which can easily lead to unsatisfactory segmentation quality. To address this problem, we propose a coarse-to-fine contour-based method for interactive medical image segmentation. This method includes automatic segmentation and interactive segmentation. We propose a coarse deformation module (CDM) to generate a coarse contour for automatic segmentation to reduce the shape error. In addition, we propose a fine deformation module (FDM) to refine the coarse contour for automatic segmentation. The FDM is also presented to adjust the local contour for interactive segmentation. Our experimental results show that the proposed method outperforms the state-of-the-art segmentation methods on the PROMISE12 and our in-house nasopharyngeal carcinoma (NPC) datasets.