Automatic segmentation of organs at risk in thoracic CT scans by combining 2D and 3D convolutional neural networks

Segmentation of organs at risk (OARs) in medical images is an important step in treatment planning for patients undergoing radiotherapy (RT). Manual segmentation of OARs is often time-consuming and tedious. Therefore, we propose a method for automatic segmentation of OARs in thoracic RT treatment planning CT scans of patients diagnosed with lung, breast or esophageal cancer. The method consists of a combination of a 2D and a 3D convolutional neural network (CNN), where both networks have substantially different architectures. We analyse the performance for these networks individually and show that a combination of both networks produces the best results. With this combination, we achieve average Dice coefficients of 0.84± 0.05, 0.94± 0.02, 0.91± 0.02, and 0.93± 0.01 for the esophagus, heart, trachea, and aorta, respectively. These results demonstrate potential for automating segmentation of organs at risk in routine radiotherapy treatment planning.