Local minima in anatomic aperture-based IMRT optimization

Purpose: An anatomic aperture‐based IMRT optimization program, named Ballista, was developed at our institution. Even though studies previously published concluded local minima in full‐IMRT optimization were not problematic, early observations with Ballista revealed their nuisance. The purpose of the present study was to evaluate the extent of local minima and their impact on the optimization. Method and Materials: In Ballista beam weights are optimized by quasi‐Newton algorithm, which cannot escape local minima, even with a quadratic dose‐based objective function. Therefore, a large number of descents were launched with random initial weights to explore the solution space for a varying number of beams. Treatment plan DVHs of different local minima were also analyzed. Results: When four beam weights were optimized, only a few but very distinctive local minima were found. For a case of 20 beam weights, the optimization revealed an astonishing number of local minima. DVH analysis showed local minima generally favor one or more organs‐at‐risk (OARs) while the other objectives are less than optimal compared with the global minimum. It was found that limiting the initial beam weights to small values eliminates the majority of the solution space containing local minima. Conclusion: With Ballista local minima proved to be a major problem. Plans corresponding to different minima differed drastically. In order to give the optimization a “clear shot” at the global minimum, initial beam weights must be limited to small values. The optimization thus focuses on improving the target volume objectives since all OARs objectives are initially met. © 2005, American Association of Physicists in Medicine. All rights reserved.