Room: Room 207
Purpose: Currently in HDR brachytherapy planning, a manual fine-tuning of an objective function is necessary to obtain case-specific valid plans. This study intends to facilitate this process by proposing a multi-criteria optimization that incorporates prior knowledge, couple to a fast quasi-Newton optimizer enabling the generation of multiple-plans in a clinically acceptable time.
Methods: 252 previously treated prostate HDR cases were divided into training (16) and validation (236) sets. In the training set, Pareto surfaces were characterized: the clinically acceptable solution space was identified after generating 2400 solutions with a Pareto surface approximation algorithm. The relationship between this particular solution space and four anchor plans (three plans that favor each organ, and one plan that equally favors all organs) was illustrated with polynomial regression models. A knowledge-based multi-criteria optimization (kMCO) was established in four steps: (1) four anchor plans computation, (2) clinically acceptable solution space estimation with regression models, (3) ten alternative plans computation (including various trade-offs) and (4) final plan selection. In the validation set, kMCO plans were compared with the physician-approved IPSA (Inverse Planning Simulated Annealing) plans.
Results: The number of clinically valid plans was 195 (82.20%) for IPSA and 231 (97.88%) for kMCO. Comparable target coverage and urethra protection were observed between IPSA plans and kMCO plans (+0.66% in mean target V100, +0.23% in mean urethral D10). The number of plans with a > 1 cc bladder V75 was 38 (4) for IPSA (kMCO). The number of plans with a > 1 cc rectum V75 was 6 (1) for IPSA (kMCO). IPSA plan generation time was 15 s/plan, while kMCO generates 14 Pareto optimal plans in 15.4 s (or 1.1 s/plan).
Conclusion: kMCO is a fast and robust inverse planning algorithm that can improve treatment plan quality without any user interventions or increase in plan generation time.
Funding Support, Disclosures, and Conflict of Interest: The authors acknowledge a scholarship from the Chinese Scholarship Council, partial support by the National Sciences and Engineering Research Council of Canada (NSERC) via the NSERC-Elekta Industrial Research Chair (Grant number: 484144-15), and partial support by the CREATE Medical Physics Research Training Network grant of the NSERC (Grant number: 432290).
Brachytherapy, Inverse Planning, Optimization