Zhu S, Maslowski A, Cunningham JM, Kuusela EH, and Chetty IJ. 3D Dose-Driven, Automatic VMAT Machine Parameter Generation with Deep Learning. Int J Radiat Oncol Biol Phys 2022; 114(3):e581-e582.
Int J Radiat Oncol Biol Phys
Purpose/Objective(s): Recent research efforts utilizing knowledge-based treatment planning for the prediction of 3D radiation dose distributions from planning structure sets have achieved positive results. Most ongoing efforts to generate deliverable plans from the predicted doses rely on full inverse optimizations using dose-volume histogram (DVH) objectives derived from these doses. In this study, we aim to leverage deep learning (DL) to rapidly generate machine delivery parameters for volumetric modulated arc therapy (VMAT) from predicted desired doses.
Materials/Methods: Data of 50 previously treated patients at our institution with prostate adenocarcinoma who received definitive radiotherapy were retrospectively obtained. All plans were generated with a one-arc VMAT technique, with conventional fractionation (78 Gy in 39 fx or 79.2 Gy in 44 fx to the prostate gland +/- seminal vesicles). A multi-task U-Net was constructed: it takes the 2D projections of the 3D dose and planning structures as inputs, and it predicts the numerical multi-leaf collimator (MLC) sequence and weights for the 178 control points. Five cases were randomly selected for testing only, and the remaining 45 formed the training set. The algorithm was implemented in Python 3.8 with PyTorch 1.7 as the DL framework. Model training was performed on a GPU. The DL-predicted plans underwent further inverse optimization with the 3D-dose-derived DVH objectives, utilizing only the last step of the Photon Optimizer (PO) in a treatment planning system. The optimization time and plan quality were compared to plans generated with one full PO optimization with the same objectives and clinical plans (all normalized to D95%=100% Rx dose).
Results: The DL model was trained for 200 epochs. On average, DL-predicted plans could be optimized in 22% (range, 18-26%) of the time required for full optimization plans. Dosimetric comparison (Table 1) demonstrated that the quality of the DL-predicted plans was comparable with clinical plans and full optimization plans, but the DL-predicted plans tended to have increased dose inhomogeneity within the PTVs.
Conclusion: We demonstrated the feasibility of rapidly generating deliverable VMAT plans from desired 3D doses with deep learning. Further work is needed to improve PTV dose homogeneity and generalize the method to multi-arc VMAT delivery.