Optimization of beam angles for intensity modulated radiation therapy treatment planning using genetic algorithm on a distributed computing platform

Abstract

Planning intensity modulated radiation therapy (IMRT) treatment involves selection of several angle parameters as well as specification of structures and constraints employed in the optimization process. Including these parameters in the combinatorial search space vastly increases the computational burden, and therefore the parameter selection is normally performed manually by a clinician, based on clinical experience. We have investigated the use of a genetic algorithm (GA) and distributed-computing platform to optimize the gantry angle parameters and provide insight into additional structures, which may be necessary, in the dose optimization process to produce optimal IMRT treatment plans. For an IMRT prostate patient, we produced the first generation of 40 samples, each of five gantry angles, by selecting from a uniform random distribution, subject to certain adjacency and opposition constraints. Dose optimization was performed by distributing the 40-plan workload over several machines running a commercial treatment planning system. A score was assigned to each resulting plan, based on how well it satisfied clinically-relevant constraints. The second generation of 40 samples was produced by combining the highest-scoring samples using techniques of crossover and mutation. The process was repeated until the sixth generation, and the results compared with a clinical (equally-spaced) gantry angle configuration. In the sixth generation, 34 of the 40 GA samples achieved better scores than the clinical plan, with the best plan showing an improvement of 84%. Moreover, the resulting configuration of beam angles tended to cluster toward the patient's sides, indicating where the inclusion of additional structures in the dose optimization process may avoid dose hot spots. Additional parameter selection in IMRT leads to a large-scale computational problem. We have demonstrated that the GA combined with a distributed-computing platform can be applied to optimize gantry angle selection within a reasonable amount of time.

A surrogate-based metaheuristic global search method for beam angle selection in radiation treatment planning

“…For fast algorithms, researchers tried simple geometric measures of patient anatomy (Potrebko et al 2007) and ranking of the beams (Vaitheeswaran et al 2010). Nazareth et al (2009) and Zhang et al (2009) used distributed computing environment to speed up the computation time. Our method takes advantage of both aspects.…”

Dose–volume histogram evaluation scheme for beam orientations in intensity-modulated radiation therapy

“…The overall time needed to find the intensity for every random selection of beam orientations becomes excessive, especially in the 3D situation. In order to improve the computational complexity, some studies have utilized branch-and-cut algorithms or distributed-computing techniques to solve the problem with an IP or MIP model in which binary and positive float variables are employed to represent candidates for beam orientations and beam intensity profiles (D'Souza et al 2004;Ehrgott et al 2008;Yang et al 2006;Nazareth et al 2009). Although such methods are able to gain some improvement in terms of computational time, the major drawback of having plan optimization does not completely capture clinical criteria for plan judgment remains, in either approach, resulting in a clinically sub-optimal or even unacceptable plan.…”

Automatic learning‐based beam angle selection for thoracic IMRT

“…Methods in the first category operate by exploring the search space of possible beam angles and running an IMRT optimization at every step. The beam angle search space is typically explored exhaustively [3][4][5] or using stochastic approaches, such as simulated annealing, 6-10 sampling, 11 genetic algorithms, 12,13 or sparse optimization. 14 Such approaches suffer from the overwhelming complexity of the search space, which grows exponentially with the number of beams.…”

“…To cope with the prohibitive computation time, the community has suggested better sampling approaches, 8,10 efficient surrogatebased beam scoring, 15 speeding up the dose computation, 7 or distributing the computation over multiple processors. 4,13 Nevertheless, these improvements are insufficient to make such methods clinically applicable at this point. Recently, hybrid approaches combining local and global minima search have been described.…”