1. The study proposes a deep learning approach for automatically contouring organs at risk (OARs) on reduced field CT (RF-CT) images taken during radiation therapy based on an initial simulation CT and contours. 2. The approach was tested on lung, prostate, and brain cases and achieved average Dice coefficients of 0.944, 0.949 and 0.960 respectively when compared to expert contours on the RF-CTs. 3. The method shows potential as part of an online adaptive radiation therapy (ART) workflow by automating re-contouring and reducing planning time from hours to minutes.