Room: Exhibit Hall | Forum 4
Purpose: To evaluate a method for fast approximation of the achievable solution space in multi-criteria-optimization (MCO) driven volumetric-modulated-arc (VMAT) treatment planning using high-level composite optimization parameters and trilinear dose-interpolation.
Methods: This work is based on a research version of a cranial stereotactic radiosurgery (SRS) inverse treatment planning software where the optimization result is controlled by three high-level parameters representing the trade-off between target-coverage and organ-at-risk-sparing, the importance of healthy tissue sparing, and the allowed degree of dose-rate- and leaf-movement-variability during treatment. The functionality was enhanced by automatic pre-optimization of different parameter combinations and an interactive preview of the achievable solution space without the need to perform a high number of time-consuming complete optimizations. The approximation of the solution space spanned by all achievable optimization results was created via trilinear dose interpolation on a subset of pre-optimized treatment plans. The quality of the approximated dose distributions for a specific parameter choice was determined by comparison to the corresponding actual optimization results. Comparison was performed on a total of 3430 treatment plans.
Results: The current discretization of the parameter space allows for a total of 343 possible parameter combinations per patient, each optimization run taking ~2 min while trilinear dose interpolation is virtually instantaneous. For a choice of 9 characteristic dose/volume values, approximation of the full solution space with a subset of 27 pre-optimized plans resulted in a mean deviation of 1.8% between estimated and actual value, decreasing the subset to 9 pre-optimized plans resulted in an average approximation error of 2.8%.
Conclusion: If SRS treatment plans controlled by a small number of non-trivial composite objectives are created in well-defined cranial settings, a subset of a relatively small number of pre-optimized plans can provide a fast overview of the achievable dose distributions and facilitate the desired parameter choice for full-scale optimization.
Funding Support, Disclosures, and Conflict of Interest: The first and second author are employees of Brainlab AG, Munich, Germany