Predictor Model Training for Real-Time Motion Management of Lung Tumors
J Rottmann*, R Berbeco, Brigham and Women's Hospital, Dana-Farber Cancer Institute and Harvard Medical School, Boston, MAMO-F-WAB-5 Monday 4:30PM - 6:00PM Room: Wabash Ballroom
Purpose: Real-time image guided radiotherapy requires tumor motion prediction to compensate for system specific latencies caused by image acquisition, motion estimation, hardware control and other factors. The performance of prediction models depends strongly on the amount of patient specific data available for training. This may be problematic if additional imaging dose is associated with the training data acquisition. To solve this problem, we investigate the use of chest wall motion data as a surrogate to train a linear predictor for lung tumor motion.
Methods: A dataset containing both lung tumor and chest wall motion trajectories for 172 beams/fractions of lung radiotherapy (11 patients) is used. We implement a linear prediction model and estimate the average geometric tracking error as a function of latency. Each trajectory is split into a dataset used for training the predictor model and one used for verification. In a second study we compare two training scenarios for a fixed system latency value of 250 ms. In the first scenario tumor motion is used for training directly, while in the second scenario the external surrogate motion trajectory is used. Verification is performed in both scenarios with the tumor motion trajectory as the reference.
Results: For latencies of 0-500ms, the average error without using a predictor increases approximately linearly to 4 mm and at 250 ms we observe 1.8 mm. Using the linear predictor model the error is reduced by an average of about 50%. Training the predictor model with external surrogate data provides similar results.
Conclusion: A linear prediction model can reduce latency induced tracking errors by an average of 50% in real-time image guided radiotherapy systems with system latencies of about 250 ms. External surrogate data is suitable to train the prediction model yielding similar prediction performance as training with lung tumor motion data directly.
Funding Support, Disclosures, and Conflict of Interest: The project described was supported, in part, by Award Number R21CA156068 from the National Cancer Institute. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Cancer Institute or the National Institutes of Health.