Program Information
Prediction and Tracking of Lung Tumor Motion Using Calypso Electromagnetic Transponders and RPM Gating Device
A Omotayo*, S Venkataraman , B McCurdy , CancerCare Manitoba, Winnipeg, MB
Presentations
SU-H2-GePD-J(B)-4 (Sunday, July 30, 2017) 3:30 PM - 4:00 PM Room: Joint Imaging-Therapy ePoster Lounge - B
Purpose: Prediction and tracking of lung tumor motion using Calypso™ electromagnetic transponders and RPM gating device
Methods: Tumor position data was obtained via implanted electromagnetic transponder beacons inside a CIRS dynamic thorax phantom. The beacons provide 3-dimensional position of the internal tumor. For modeling, we implemented a non-parametric and probabilistic learning algorithm, using Gaussian kernels/covariance functions to capture the complex non-linear internal-external correlations between both signals. The model was trained and tested with previously obtained respiratory motion using an external surrogate (RPM™ system), in batch/offline and online modes. This signal was used to predict the internal tumor position beyond the look-ahead time or latency (33 – 167ms), and dynamically adapt in real-time, to the measured tumor position. The model parameters were optimized, while the expected prediction error (RMSE) was quantified as a function of the latency. We used breathing profiles of six different lung SBRT patients representing varying patterns from regular to irregular breathing. For real-time online operation, the algorithm was trained in a sequential manner as each time series arrives i.e. receive 1 input-output, train k-steps-ahead, in a recursive manner.
Results: The average ±one standard deviation RMSE for all patients was within a range of (0.12±0.26)mm to (2.44±1.71)mm over all investigated prediction latencies. As expected, the prediction accuracy decreases as the latency duration increases, however we can predict 167ms ahead within a 1mm prediction error, safe for one patient. The Gaussian kernel adaptive filter was deemed optimal in terms of computational complexity (memory and floating operations) and accuracy, as all training and predictions over 120s duration was computed in 3.5s.
Conclusion: In this study, we have demonstrated the prediction and tracking of respiratory motion in real time. We have also been able to use our predictive model to estimate the correlation between internal and external surrogates.
Contact Email: