Room: AAPM ePoster Library
Purpose: To regularize planning by predicting a VMAT plan’s control points or linac/MLC parameters for a novel anatomy. The machine parameters’ models are learned by deep neural networks (DNNs).
Methods: A CycleGAN DNN was trained on 138 VMAT prostate plans with a separate 20 test plans. The target anatomy was encoded as beam’s-eye-view projections at fixed gantry angles, and the MLC apertures were encoded as graphical images scaled and registered to the projections and centered on the plan isocenter. A plan is represented by the stacked apertures and its model is learned paired with a unique corresponding anatomy. Supervised learning was implemented as a voxel-wise binary classifier to predict those voxels occurring in the apertures. Training performance was assessed by the mean absolute leaf position differences (MALD) and models were saved at epoch intervals. Models at MALD minima were translated back into RTPlans and input to a TPS from which the RTDoses and DVHs were exported.
Results: Models for several training regimes were examined. The 3D CycleGAN models routinely produced plan-estimates with MALDs of less than 2.5 mm for all control points and with target and OAR DVHs closely matching those of the ground truth plans. Mean absolute Dmean, Dmax, and other dosimetry metrics routinely lie within 3% of the ground truth plans.
Conclusion: These results suggest possible roles for DNN-predicted treatment plans: 1) to provide a lower bound to plan quality for planning guidance or for automated plan quality assurance, or 2) to warm-start plan refinement to generate a clinical plan. Since the DNN plans can be generated in 60 s or less, they may be useful in adaptive treatment.