Room: Karl Dean Ballroom A2
Purpose: Recent implementations of multi-objective optimization in treatment planning estimate the Pareto frontier by generating anchor plans where one objective is optimized per plan and a balance plan which equally weighs all objectives, connecting the frontier through interpolation. However, balance plans have been shown to suboptimally represent the frontier center, leading to inaccurate interpolation. Here, we present a method of computing a more central plan by incorporating anchor plan information as constraints.
Methods: 50 anonymized prostate cases were retrospectively used for analysis. The three objectives of interest were the bladder and rectum gEUD (alpha = 3), and the PTV D95. To ensure clinical relevance, higher-weighted constraints were applied to each plan optimization, including PTV D95 > 95% of prescription and bladder/rectum gEUD < 60% of prescription. Anchor plans and two central plans were computed, including a balance plan and a constrained central plan optimizing PTV D95 under additional higher-weighted constraints. The values for these constraints were the averages between the minimum and maximum values that the objectives took over all anchor plans. The central plans were compared by examining their distances from the central line spanned by the mean of the anchor plans. Central line distance distributions were summarized using descriptive statistics and Studentâ€™s paired t-test.
Results: The central line distances of the mean anchor plans, balance plans, and constrained central plans had means 19.75%, 7.39%, and 2.80%, and standard deviations 3.90%, 3.23%, and 1.08%, respectively. The Studentâ€™s paired t-test yielded p < 1E-12, indicating that the constrained central plan distances are significantly shorter than the balance plan distances.
Conclusion: Preliminary results suggest that the proposed technique is capable of more accurately estimating the Pareto frontier. This may lead to more efficient use of treatment planning time, since interpolated solutions are more likely to accurately approximate achievable solutions.