Exploring baseline shift prediction in respiration induced tumor motion

Effective management of respiratory motion is essential for achieving the clinical goals of stereo tactic thoracic and abdominal radiotherapy, where highly potent radiation beams are precisely directed in order to ablate the tumor, while minimizing radiation damage to normal tissue and critical organs. Due to cycle-to-cycle variations in respiratory motion, it is important to be able to predict imminent anomalous or irregular tumor motion ahead of its occurrence. Such information can then be used to pause the radiation delivery, or to track the moving tumor. However, predicting tumor motion anomalies presents a challenge as the occurrence of these anomalies can vary from patient to patient and from day to day for the same patient. In this paper, we explore the use of observed data in predicting baseline trends, and baseline shifts, in particular. Using a tumor motion dataset obtained from 143 treatment fractions from 42 patients treated with Cyber knife Synchrony System, we execute multifaceted analyses, including offline and online scenarios. Given the variation in tumor motion patterns and the absence of standardized baselines and adequate personalized prior data, we compare performances of standard prediction algorithms with and without training on prior data. Our analyses yield promising results for baseline shift prediction, and real-time baseline trend estimation in general.