Program Information
On the Fluence Map Delivery Problem in VMAT
K Van Amerongen1*, M Kelly2 , M Balvert3 , T Bortfeld4 , D Craft5 , (1) Tilburg university, Tilburg, Tilburg, (2) Rethink Robotics, Boston, Massachusetts, (3) Tilburg university, Tilburg, Tilburg, (4) Massachusetts General Hospital, Boston, MA, (5) Massachusetts General Hospital, Cambridge, AA
Presentations
TU-D-108-2 (Tuesday, August 1, 2017) 11:00 AM - 12:15 PM Room: 108
Purpose: The most efficient way to dynamically delivery a given fluence map remains an unsolved problem, but is a critical component of VMAT optimization. We present two new methods for optimal time delivery of a given fluence map.
Methods: The first method models the leaf trajectories and dose rate as polynomials, which regularizes the search space and thus reduces the number of local minima. This is a new technique that is borrowed from methods that have been applied in spacecraft trajectory optimization for decades. Rather than optimize individual leaf positions over time, we represent the leaf trajectories as polynomials and optimize over the coefficients of those polynomials, which is a much smaller space, and implicitly instructs the optimization routine to not search for wasted “back-and-forth” leaf trajectories, which are allowable but unlikely optimal. The second technique uses an outer (infeasible) projection method to “skip over” local minima in order to converge more directly on the global minimum. These methods are both wrapped in a parallel optimization framework since leaf trajectory optimizations for the fluence rows decouple once the dose rate is set (in the outer loop).
Results: Both methods show improved solution quality and marked improvement in solution time (>80% reduction) compared to the initial methods implemented to solve this problem. We discuss how both of the methods will extend to the full map VMAT case.
Conclusion: The key to solving the full VMAT problem, meaning finding the optimal plan in a given allotted delivery time, is solving the underlying optimal delivery of a single fluence map, which has largely been ignored (general case: variable dose rate, arbitrary leaf motion) even though it is a fundamental and challenging non-convex optimization problem. This work highlights the difficulty of the problem and describes two promising approaches that attack it from different angles.
Funding Support, Disclosures, and Conflict of Interest: D Craft and T Bortfeld are partially supported by RaySearch laboratories.
Contact Email: