Validation of the Acuros XB dose calculation algorithm versus Monte Carlo for clinical treatment plans.

PURPOSE: The two distinct dose computation paradigms of Boltzmann equation solvers and Monte Carlo simulation both promise in principle maximum accuracy. In practice, clinically acceptable calculation times demand approximations and numerical short-cuts on one hand, and modeling the beam characteristics of a real linear accelerator to the required accuracy on the other. A thorough benchmark of both algorithm types therefore needs to start with beam modeling, and needs to include a number of clinically challenging treatment plans.

METHODS: The Acuros XB (v 13.7, Varian Medical Systems) and SciMoCa (v 1.0, Scientific RT) algorithms were commissioned for the same Varian Clinac accelerator for beam qualities 6 and 15 MV. Beam models were established with water phantom measurements and MLC calibration protocols. In total, 25 patients of five case classes (lung/three-dimensional (3D) conformal, lung/IMRT, head and neck/VMAT, cervix/IMRT, and rectum/VMAT) were randomly selected from the clinical database and computed with both algorithms. Statistics of 3D gamma analysis for various dose/distance-to-agreement (DTA) criteria and differences in selected DVH parameters were analyzed.

RESULTS: The percentage of points fulfilling a gamma evaluation was scored as the gamma agreement index (GAI), denoted as G(ΔD, DTA). G(3,3), G(2,2), and G(1,1) were evaluated for the full body, PTV, and selected organs at risk (OARs). For all patients, G(3,3) ≥ 99.9% and G(2,2) > 97% for the body. G(1,1) varied among the patients. However, for all patients, G(1,1) > 70% and G(1,1) > 80% for 68% of the patients. For each patient, the mean dose deviation was ΔD < 1% for the body, PTV, and all evaluated OARs, respectively. In dense bone and at off-axis distance > 10 cm, the Acuros algorithm yielded slightly higher doses. In the first layer of voxels of the patient surface, the calculated doses deviated between the algorithms. However, at the second voxel, good agreement was observed. The differences in D(98%PTV) were <1.9% between the two algorithms and for 76% of the patients, deviations were below 1%.

CONCLUSIONS: Overall, an outstanding agreement was found between the Boltzmann equation solver and Monte Carlo. High-accuracy dose computation algorithms have matured to a level that their differences are below common experimental detection thresholds for clinical treatment plans. Aside from residual differences which could be traced back to implementation details and fundamental cross-section data, both algorithms arrive at identical dose distributions.