SU‐E‐T‐655: Dose Matrix Sampling Heuristic Applied to Improving IMRT Optimization Efficiency

Abstract

Purpose: To introduce a novel heuristic that increases the efficiency of IMRT fluence optimization, by employing sampled dose values in evaluation of the objective function. Methods: Pencil beam dose matrices were computed using CERR, for 1‐cm square beamlets. These matrices were then uniformly sampled at various rates (1–10%), and the resulting approximate dose‐volume values compared to the true values for various fractional volumes. By varying the beamlet weights for 10,000 random sets, we estimated the error in these approximations to be <1%. For IMRT optimization, a least squares objective function was used, based on clinical dose‐volume objectives. In the initial stage of optimization, the method utilizes approximate dose values, determined by 1% or 5% dose matrix sampling, in objective function evaluation. This results in a significant reduction in computation time, since the manipulation of large dose matrices is avoided. The optimized beamlet weights are subsequently used as the starting point in the final optimization. This method was evaluated on prostate bed and esophageal patients, by comparing the CPU time required with that of a standard optimization method. The downhill simplex and simulated annealing algorithms were used for the evaluations. Results: The prostate bed patient was optimized with both a 1% and 5% sampling rate resulting in patient treatment plans of equal score, and a time savings of 18% and 42%, respectively. The esophageal patient was optimized with a 5% sampling rate and provided a score improvement of 23%, with a time savings of 17%. Conclusion: We have developed a method to be used in conjunction with existing optimization algorithms, which dramatically decreases the number of standard objective function evaluations and computation time. This technique may be applied to any type of optimization algorithm and set of clinical variables requiring pre‐computed dose matrices and objective functions involving dose‐volume objectives (eg. multi‐criteria optimization).