Segmentation of Organs at Risk in Chest Cavity Using 3D Deep Neural Network

Cancer ranks second in the world in terms of morbidity and mortality. Radiotherapy is one of standard treatment procedures in case of cancer. In lung and esophageal cancer, the irradiation planning begins with the delineation of the target tumor and healthy organs. In this paper we propose solving Organs at Risk (OAR) segmentation task with deep learning approach. Proposed network is based on U-Net-like architecture with ResNet-50 backbone. Multi-resolution approach is used for localization of volumes of interest (VOIs) for each individual organ at inference stage. Generalized Dice Loss is used for neural network weights optimization. Experiments demonstrate competitive performance on a dataset of 60 computed tomography (CT) scans (40 CT scans are used for network training, 20 CT scans - for evaluation). The following metric scores were obtained on the test split: dice score - 0.77 (esophagus), 0.933 (heart), 0.913 (trachea), 0.908 (aorta); Hausdorff distance - 0.605 (esophagus), 0.225 (heart), 0.223 (trachea), 0.326 (aorta).