W-net: A Network Structure for Automatic Segmentation of Organs at Risk in Thorax Computed Tomography

Abstract

Accurate segmentation of Organs at Risk (OAR) on Computed Tomography (CT) images is a crucial step in radiotherapy treatment planning. In this paper, we propose a novel W-Net structure combining a U-Net segmentation network and an adversarial network (GAN) to reconstruct the OAR. With the reconstruction loss, W-Net can better learn effective features and get more accuracy segmentation result than U-Net. We test our method in the SegTHOR challenge which focus on 4 thoracic OAR: esophagus, heart, trachea and aorta. The average Dice Similarity Coefficient (%) of W-Net and U-Net on these 4 OAR are 80.6 versus 79.6, 93.8 versus 93.4, 88.3 versus 88.1, and 91.5 versus 90.6. The Hausdorff Distance (HD) are 0.5905 versus 0.6923, 0.2055 versus 0.2215, 0.7162 versus 0.7374, and 0.8061 versus 0.9290.