Leveraging deep learning for markerless motion management in radiation therapy Project Summary Organ motion is a predominant limiting factor for the maximum exploitation of modern radiation therapy (RT). Adverse influence of the organ motion is aggravated in hypofractionated treatment because of protracted dose delivery. Current image guided RT often relies on the use of implanted fiducial markers (FMs) for online/offline target localization, which is invasive and costly, and introduces possible bleeding, infection and discomfort of the patient. In this project, we harness the enormous potential of deep learning and investigate a novel markerless localization strategy by combined use of a pre-trained deep learning model and kV X-ray projection or cone beam CT images. We hypothesize that incorporation of deep layers of image information allows us to visualize otherwise invisible target in real-time and greatly reduce the uncertainties in beam targeting. Specific aims of the project are to: (1) Develop a DL-based tumor target localization framework for image guided RT (IGRT); (2) Apply the DL-based strategy to localize prostate target on 2D kV X-ray projection and 3D CBCT images; and (3) Evaluate the potential clinical impact of the DL strategy for pancreatic IGRT. This study brings up, for the first time, highly accurate markerless target localization based on deep learning and provides a clinically sensible solution for IGRT of prostate and pancreas cancers or other types of cancers. Successful completion of this investigation will significantly advance the current beam targeting technique and provide radiation oncology discipline a powerful way to safely and reliably escalate the radiation dose for precision RT. Given its significant promise to optimally cater for inter- and intra-fractional uncertainties, the study should lead to substantial improvement in patient care and enables us to utilize maximally the technical capability of modern RT such as IMRT and VMAT. Given the dose responsive nature of various cancers and that the proposed method requires no hardware modification, this research should lead to a widespread impact on the management of neoplasmic diseases affected by organ motion.