Machine Learning for Intra-Fraction Tumor Motion Modeling with Respiratory Surrogates

Advances in radiation therapy for cancer have made it possible to deliver conformal doses to the tumor while sparing normal healthy tissues. However, one of the difficulties radiation oncologists face is targeting moving tumors, such as those in the lung, which can change position during normal respiration. Tumor motion can be determined by directly monitoring tumor position using continuous xray imaging or electromagnetic transponders placed in the tumor that emit a signal. These approaches require potentially unnecessary radiation to the patient or acquisition of expensive technology. Alternatively, one can image the patient intermittently to determine tumor location and external markers placed on the patient´s torso. The external surrogates can then be used to determine an inferential model that would determine the tumor position as a function of external surrogates. These external surrogates can be monitored continuously in order to determine the real-time position of the tumor. In this study, we evaluate a machine learning algorithm for inferring intra-fraction tumor motion from external markers using a database of Cyberknife Synchrony^TM system.