Tseung, H. Wan Chan

Abstract Background Conventional proton dose calculation methods are either time‐ and resource‐intensive, like Monte Carlo (MC) simulations, or they sacrifice accuracy, as seen with analytical methods. This trade‐off between computational efficiency and accuracy highlights the need for improved dose calculation approaches in clinical settings. Purpose This study aims to develop a deep‐learning‐based model that calculates dose‐to‐water ( D W ) and dose‐to‐medium ( D M ) using patient anatomy and proton spot map (PSM), achieving approaching MC‐level accuracy with significantly reduced computation time. Additionally, the study seeks to generalize the model to different treatment sites using transfer learning. Methods A SwinUNetr model was developed using 259 four‐field prostate proton stereotactic body radiation therapy (SBRT) plans to calculate patient‐specific D W and D M distributions from CT and projected PSM (PPSM). The PPSM was created by projecting PSM into the CT scans using spot coordinates, stopping power ratio, beam divergence, and water‐equivalent thickness. Fine‐tuning was then performed for the central nervous system (CNS) site using 84 CNS plans. The model's accuracy was evaluated against MC simulation benchmarks using mean absolute error (MAE), gamma analysis (2% local dose difference, 2‐mm distance‐to‐agreement, 10% low dose threshold), and relevant clinical indices on the test dataset. Results The trained model achieved a single‐field dose calculation time of 0.07 s on a Nvidia‐A100 GPU, over 100 times faster than MC simulators. For the prostate site, the best‐performing model showed an average MAE of 0.26 ± 0.17 Gy and a gamma index of 92.2% ± 3.1% in dose regions above 10% of the maximum dose for D W calculations, and an MAE of 0.30 ± 0.19 Gy with a gamma index of 89.7% ± 3.9% for D M calculations. After transfer learning for CNS plans, the model achieved an MAE of 0.49 ± 0.24 Gy and a gamma index of 90.1% ± 2.7% for D W computations, and an MAE of 0.47 ± 0.25 Gy with a gamma index of 85.4% ± 7.1% for D M computations. Conclusions The SwinUNetr model provides an efficient and accurate method for computing dose distributions in proton therapy. It also opens the possibility of reverse‐engineering PSM from D W , potentially speeding up treatment planning while maintaining accuracy.

AbstractBackgroundAlthough intensity‐modulated radiation therapy and volumetric arc therapy have revolutionized photon external beam therapies, the technological advances associated with electron beam therapy have fallen behind. Modern linear accelerators contain technologies that would allow for more advanced forms of electron treatments, such as beam collimation, using the conventional photon multi‐leaf collimator (MLC); however, no commercial solutions exist that calculate dose from such beam delivery modes. Additionally, for clinical adoption to occur, dose calculation times would need to be on par with that of modern dose calculation algorithms.PurposeThis work developed a graphics processing unit (GPU)‐accelerated Monte Carlo (MC) engine incorporating the Varian TrueBeam linac head geometry for a rapid calculation of electron beams collimated using the conventional photon MLC.MethodsA compute unified device architecture framework was created for the following: (1) transport of electrons and photons through the linac head geometry, considering multiple scattering, Bremsstrahlung, Møller, Compton, and pair production interactions; (2) electron and photon propagation through the CT geometry, considering all interactions plus the photoelectric effect; and (3) secondary particle cascades through the linac head and within the CT geometry. The linac head collimating geometry was modeled according to the specifications provided by the vendor, who also provided phase‐space files. The MC was benchmarked against EGSnrc/DOSXYZnrc/GEANT by simulating individual interactions with simple geometries, pencil, and square beam dose calculations in various phantoms. MC‐calculated dose distributions for MLC and jaw‐collimated electron fields were compared to measurements in a water phantom and with radiochromic film.ResultsPencil and square beam dose distributions are in good agreement with DOSXYZnrc. Angular and spatial distributions for multiple scattering and secondary particle production in thin slab geometries are in good agreement with EGSnrc and GEANT. Dose profiles for MLC and jaw‐collimated 6–20‐MeV electron beams showed an average absolute difference of 1.1 and 1.9 mm for the FWHM and 80%–20% penumbra from measured profiles. Percent depth doses showed differences of <5% for as compared to measurement. The computation time on an NVIDIA Tesla V100 card was 2.5 min to achieve a dose uncertainty of <1%, which is ∼300 times faster than published results in a similar geometry using a single‐CPU core.ConclusionsThe GPU‐based MC can quickly calculate dose for electron fields collimated using the conventional photon MLC. The fast calculation times will allow for a rapid calculation of electron fields for mixed photon and electron particle therapy.

PurposeProvide an adaptive multicriteria optimization (MCO) method for intensity‐modulated proton therapy (IMPT) utilizing GPU technology. Previously described limitations of MCO such as Pareto approximation and limitation on the number of objectives were addressed.MethodsThe treatment planning process for IMPT must account for multiple objectives, which requires extensive treatment planning resources. Often a large number of objectives (>10) are required. Hence the need for an MCO algorithm that can handle large number of objectives. The novelty of the MCO method presented here lies on the introduction of the adaptive weighting scheme that can generate a well‐distributed and dense representation of the Pareto surface for a large number of objectives in an efficient manner. In our approach the generated Pareto surface is constructed for a set of DVH objectives. The MCO algorithm is based on the augmented weighted Chebychev metric (AWCM) method with an adaptive weighting scheme. This scheme uses the differential evolution (DE) method to generate a set of well‐distributed Pareto points. The quality of the Pareto points’ distribution in the objective space was assessed quantitatively using the Pareto sampling metric. The MCO algorithm was developed to perform multiple parallel searches to achieve a rapid mapping of the Pareto surface, produce clinically deliverable plans, and was implemented on a GPU cluster. The MCO algorithm was tested on two clinical cases with 10 and 18 objectives. For each case one of the MCO‐generated plans was selected for comparison with the clinically generated plan. The MCO plan was randomly selected out of the set of MCO plans that had target coverage similar to the clinically generated plan and the same or better sparing of the organs at risk (OAR). Additionally, a validation study of the AWCM method vs the weighted sum method was performed.ResultsThe adaptive MCO algorithm generated Pareto points on the Pareto hypersurface in a fast (2–3 hr) and efficient manner for 2 cases with 10 and 18 objectives. The MCO algorithm generated a dense and well‐distributed set of Pareto points on the objective space, and was able to achieve minimization of the Pareto sampling metric. The selected MCO plan showed an improvement of the DVH objectives in comparison to the clinically optimized plan in both cases. For case one, the MCO plan showed a 48% reduction of the 50% dose to OARs and a 16% reduction of the 1% dose to OARs. For case 2, the MCO plan showed a 72% reduction of the 50% dose to OARs and a 42% reduction of the 1% dose to OARs. The comparison of AWCM to WS showed that the AWCM method has a dosimetric advantage over WS for both patient cases.ConclusionWe introduced an adaptive MCO algorithm for IMPT accelerated using GPUs. The algorithm is based on an adaptive method for generating Pareto plans in the objective space. We have shown that the algorithm can provide rapid and efficient mapping of the multicriteria Pareto surface with clinically deliverable plans.