Predicting 3D Gamma Passing Rates of Head-And-Neck Volumetric Modulated Arc Therapy Using Machine and Deep Learning Models
S Sivabhaskar1*, R Li1, J Buatti1, M de Oliveira1, M Bonnen1, A Roy2, N Kirby1, S Stathakis1, N Papanikolaou1, (1) University of Texas Health Science Center, San Antonio, TX, (2) The University of Texas at San Antonio, San Antonio, TX
Presentations
PO-GePV-M-56 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To evaluate and compare the performance of random forest regression, support vector regression, and artificial neural network (ANN) in predicting the 3D gamma passing rate (GPR) of head-and-neck volumetric modulated arc therapy (VMAT) plans delivered using Elekta linear accelerators.
Methods: In this study, 44 VMAT plans generated in Pinnacle treatment planning system were used. Patient-specific quality assurance was performed using the OCTAVIUS 4D phantom, and the GPRs were computed using a 3%/2 mm dose difference and distance-to-agreement gamma criterion. Sixteen plan complexity metrics (MU factor, number of arcs, total number of control points, arc length, dose rate, total MU, average collimator angle, plan modulation complexity score, total jaw travel, average jaw position, average jaw motion, leaf gap, total leaf travel, average leaf motion, gantry motion, and aperture area) were computed from DICOM RT-Plan files and used as model inputs to predict the GPR. The best hyperparameters were determined through grid search and 4-fold cross validation, and model performance was evaluated using the mean absolute error (MAE) and root mean square error (RMSE).
Results: The support vector achieved the following results (MAE = 1.746% and RMSE = 2.227%), and random forest achieved the following results (MAE = 1.620% and RMSE = 2.071%). ANN achieved the following results (MAE = 2.140% and RMSE = 2.629%). Random forest’s most important features are average leaf motion, aperture area, total jaw travel, average jaw position, plan modulation complexity score, total MU, total leaf travel, average jaw motion, leaf gap, MU factor, arc length, and total number of control points.
Conclusion: Support vector and random forest slightly outperform ANN in predicting the 3D GPRs of head-and-neck VMAT plans. With further inclusion of more plans, the prediction errors can be reduced.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email