Program Information
Real-Time Error Estimation for Real-Time Motion Prediction
D Moore* and A Sawant, UT Southwestern Medical Center, Dallas, TX
Presentations
TH-AB-303-3 (Thursday, July 16, 2015) 7:30 AM - 9:30 AM Room: 303
Purpose:
A critical component of real-time MLC tracking is to incorporate motion prediction in order to account for system latency, the time between the occurrence of motion and MLC leaf repositioning, so as to avoid geometric and dosimetric errors. In this work, we address a major limitation of current motion prediction algorithms – estimating the inherent error in prediction in real-time. We propose using an ensemble of real-time predictors to reduce overall prediction error, and provide a quantitative metric to trigger beam holds based on a user-defined error threshold.
Methods:
An ensemble of eighteen kernel density estimation (KDE)-based estimators was implemented in parallel and deployed on motion traces recorded from ten lung cancer patients. For each trace, the ensemble generated a sequence of predictions for a look-ahead time of 250ms, typical for many real-time systems. Each predictor, and the ensemble as a whole, was evaluated based on their root-mean-square error (RMSE) with respect to the original data, and assigned a beam hold score between 0 (perfect) and 1 to quantify the number of false beam holds and false beam releases.
Results:
The average execution time of the KDE-ensemble was 16ms, well below the 250ms look-ahead. For all traces, the ensemble produced a lower RMSE and hold score than any of the individual constituent predictors. For the worst-case motion trace, the minimum RMSE(hold score) for individual predictors was 0.912cm(0.340) compared to 0.854cm(0.305) for the KDE-ensemble predictor. The hold score for individual estimators was ~0.6 when error estimates were not used to assert beam holds.
Conclusion:
The KDE-ensemble approach results in improved motion prediction. More importantly, it also provides real-time estimations of prediction error, thus addressing a major gap in current motion prediction algorithms. Such error prediction can be utilized to assert beam-holds during MLC tracking, preventing serious geometric and dosimetric errors.
Funding Support, Disclosures, and Conflict of Interest: Funding for this work was provided by Elekta Limited.
Contact Email: