Room: AAPM ePoster Library
Purpose: To introduce a novel treatment planning algorithm based on a split-feasibility formulation of the inverse problem for intensity-modulated proton therapy (IMPT) and demonstrate its capabilities for handling multiple dose-volume constraints.
Methods: A dynamic string-averaging CQ-method has been developed for fluence-based optimization in IMPT inverse planning, allowing flexibility in the weighting and order in which projections are executed onto constraint sets. This includes dynamic adjustment of over- or under-relaxation of projections onto individual sets. The algorithm was implemented in MATLAB for a variety of test cases, including a head-and-neck case containing 23 delineated structures with both linear and nonlinear constraints imposed. An accelerated version of the algorithm was written in Python, making use of multithreading and just-in-time compilation techniques. Spatial dose distributions, dose-volume histograms, Wilcoxon signed-rank tests and constraint violation counts have been used to compare dose solutions.
Results: All test cases showed convergence of the algorithm to feasible solutions. The most complex case achieved target coverage within 0.5 Gy of the accepted clinical dose solution, and offered the same or better sparing in 11 out of 13 organs at risk. Accelerated implementation on up to 20 CPU threads reduced runtime by a factor of 15.4, down to 0.145 seconds per cycle for the most complex clinical case. Suitable results were achieved after approximately 300 cycles (43.5 seconds). Optimizing only on linear constraints reduced this time to only 15.6 seconds.
Conclusion: This work provides proof of concept for a projection-based IMPT inverse planning algorithm that is mathematically robust to the order and weighting with which projections are executed, even in non-convex settings imposed by multiple dose-volume constraints. Numerical convergence can be achieved well within clinically imposed time restrictions, however finer spatial considerations such as hot and cold dose spots may still need judicious correction post-optimization.
Funding Support, Disclosures, and Conflict of Interest: MB is supported by Cancer Research UK Grant No. C2195/A25197, through a CRUK Oxford Centre DPhil Prize Studentship, YC by ISF-NSFC joint research program grant No. 2874/19, AG by the COST Action CA16228 European Network for Game Theory, FVdH by Cancer Centre grant CRUK C2195/A25014 and CRUK OIRO Number C5255/A23755.