Coarse-To-Fine Segmentation of Organs at Risk in Nasopharyngeal Carcinoma Radiotherapy

Accurate segmentation of organs at risk (OARs) from medical images plays a crucial role in nasopharyngeal carcinoma (NPC) radiotherapy. For automatic OARs segmentation, several approaches based on deep learning have been proposed, however, most of them face the problem of unbalanced foreground and background in NPC medical images, leading to unsatisfactory segmentation performance, especially for the OARs with small size. In this paper, we propose a novel end-to-end two-stage segmentation network, including the first stage for coarse segmentation by an encoder-decoder architecture embedded with a target detection module (TDM) and the second stage for refinement by two elaborate strategies for large- and small-size OARs, respectively. Specifically, guided by TDM, the coarse segmentation network can generate preliminary results which are further divided into large- and small-size OARs groups according to a preset threshold with respect to the size of targets. For the large-size OARs, considering the boundary ambiguity problem of the targets, we design an edge-aware module (EAM) to preserve the boundary details and thus improve the segmentation performance. On the other hand, a point cloud module (PCM) is devised to refine the segmentation results for small-size OARs, since the point cloud data is sensitive to sparse structures and fits the characteristic of small-size OARs. We evaluate our method on the public Head&Neck dataset, and the experimental results demonstrate the superiority of our method compared with the state-of-the-art methods. Code is available at https://github.com/DeepMedLab/Coarse-to-fine-segmentation.