Feasibility study of a machine learning inspired approach for VMAT optimization

Despite the widespread clinical adoption of volumetric modulated arc therapy (VMAT), advances in its fundamental optimization methodology have remained relatively limited, particularly with respect to open and researcher-accessible optimization frameworks.

This study introduces a novel machine learning (ML) inspired approach for VMAT optimization, reformulating the problem as a multilayer neural network solvable with modern ML toolkits.

In this framework, multileaf collimator (MLC) leaf positions and control-point weights are optimized. They are represented as trainable parameters embedded within parameterized activation functions and the final weighting layer, respectively. The dose-deposition matrix provides a fixed linear mapping. Optimization was performed using PyTorch's built-in L-BFGS optimizer with GPU acceleration. Machine-specific constraints, including maximum dose rate, gantry speed, MLC motion limits, and trajectory smoothness, were incorporated as regularization terms. The framework was evaluated using prostate cases with two arcs and head-and-neck (HN) cases with two and four arcs, with results compared against corresponding benchmark IMRT plans.

All VMAT optimizations converged successfully, with stable reduction of total objective values and reasonable trends in machine-related regularization terms. The optimized plans were successfully imported into Eclipse TPS and delivered on a TrueBeam linac without interlocks, confirming deliverability. For prostate cases, two-arc VMAT plans achieved planning target volume (PTV) coverage and organ-at-risk (OAR) sparing comparable to benchmark IMRT plans with similar DVH characteristics. For HN cases, four-arc VMAT plans provided plan quality comparable to benchmark IMRT, and consistently improved target dose conformity and OAR sparing compared with two-arc plans, particularly in regions adjacent to complex target geometries. All observations and comparisons are consistent with established clinical experience on VMAT optimization.

The proposed ML based VMAT optimization framework bridges modern machine learning optimization with treatment plan optimization and demonstrates strong potential as a flexible and extensible platform for future algorithmic development and research-driven innovations.