Automated Segmentation of Organs at Risk for Nasopharyngeal Carcinoma by Dropout U-Net

Radiotherapy is one of the most important treatments for cancers. The first step in radiotherapy is accurately segmenting the organs at risk (OARs), which are later used in the design of the plan to minimize the dose received by the healthy tissues. Manual segment OARs are a time-consuming and laborintensive procedure, especially for nasopharyngeal carcinoma (NPC), which requires hours to accomplish the segmentation for a single patient. Recently, U-Net has been extensively applied in automated segmentation tasks, which incorporate shortcut connections between the encoder and decoder to facilitate the information flow. In this paper, a novel dropout-based regularization method is leveraged to augment the representation capacity of shortcut connections in U-Net. Experimental results show that dropout can substantially improve the dice similarity coefficient of U-Net on multiple OARs, including the brain stem, eyes, parotid, pituitary, and TMJ.