VMAT Optimization: A Machine Learning Perspective
Abstract
Purpose
VMAT optimization is a non-convex problem with tightly coupled parameters and machine constraints, which limits the development of transparent and extensible frameworks outside commercial treatment planning systems. This work introduces a new perspective on VMAT optimization by leveraging modern machine learning (ML) frameworks and establishes a flexible platform for research.
Methods
VMAT optimization was reformulated as training a multi-layer neural network. Trainable parameters include multileaf collimator (MLC) leaf positions embedded in parameterized activation functions, and control-point weights represented by the final weighting layer. Machine-specific constraints, including maximum dose rate, MLC leaf-speed limits, and trajectory smoothness, were incorporated as regularization terms. The framework was implemented in PyTorch utilizing its L-BFGS optimizer with GPU acceleration and was evaluated with typical prostate (two-arc) cases and more complex head-and-neck (two- and four-arc) cases. Results were benchmarked against corresponding reference IMRT plans.
Results
All VMAT optimizations reached stable convergence of both dose objectives and machine-related regularization terms. All plans were validated in Eclipse and successfully delivered on the TrueBeam without interlocks. For prostate cases, two-arc VMAT plans achieved results comparable to benchmark IMRT, with PTV coverage showing D99% within 0.3% to 1.2% difference and similar hot-spot control (PTV D0.1% ≤ 103% for a prescription of 70 Gy), while maintaining comparable OAR sparing. For head-and-neck cases, four-arc VMAT plans likewise maintained overall plan quality comparable to benchmark nine-beam IMRT. Compared with two-arc configurations, four-arc plans demonstrated improved target coverage, with PTV D99% increased by approximately 1.1% to 1.8%, along with enhanced sparing of critical OARs. These dosimetric trends were consistent with clinical experience.
Conclusion
The implemented ML-based VMAT optimization framework achieves acceptable treatment plans and provides an alternative and robust platform for treatment plan optimization development and other research-driven innovations.