Click here to

Session: Applications of AI in Radiotherapy Planning and Adaptation [Return to Session]

Investigation of Computation Time and Storage Saving Using Generative Adversarial Network (GAN) Source Models for Dose Simulation of a Binary MLC Linac

M Shi1*, S Cui1, C Chuang1, O Oderinde2, N Kovalchuk1, K Bush1, M Surucu1, L Xing1, B Han1, (1) Stanford University, Palo Alto, CA, (2) RefleXion Medical, Hayward, CA


SU-J-BRC-6 (Sunday, 7/10/2022) 4:00 PM - 5:00 PM [Eastern Time (GMT-4)]

Ballroom C

Purpose: To investigate the feasibility of using Generative Adversarial Network (GAN) in modeling RefleXion X1 Linac by replacing each collimated beam with a GAN generator for Monte Carlo simulation. The potential benefits including calculation time and storage saving are estimated.

Methods: A series of GAN generators were created for RefleXion, a novel image-guided radiotherapy machine equipped with a binary multi-leaf collimation (MLC) system. Each generator serves as a beam source of a MLC aperture in Monte Carlo simulation, similar to a phase space source. Dose distributions in water were simulated using both the GAN sources and the conventional phase space sources. Agreement between those simulated dose distributions were investigated. A total of 32 apertures from single leaf opening to full leaf opening with two jaw settings at field center and field edge were tested. Each GAN model was trained from a phase space dataset different from the verification dataset. Training parameters were fine-tuned to optimize the water dose agreement with the conventional simulation. Computation time and file storage space saved by implementing this deep neural network method was estimated.

Results: PDD10, penumbra, full-width half maximum, and beam profile consistency of the GAN simulation agree with conventional simulation with differences of 0.4% ± 0.2%, 0.23 ± 0.19 mm, 0.31 ± 0.51 mm, and 1.98% ± 1.20%, respectively. Gamma passing rate (1%/1mm) of the planar dose at 10 cm depth is greater than 90%. The estimated time saving for the simulation of an IMRT plan delivering 100 cGy per fraction using 5766 beamlets is 5300 CPU hours. Storage usage is reduced by 42 folds.

Conclusion: The GAN model can simulate the RefleXion Linac accurately and efficiently. The storage-saving and time-saving features as compared with phase space simulation make this technique potentially useful in the simulation and cloud-based applications of a new Linac.

Funding Support, Disclosures, and Conflict of Interest: This research is supported by a research grant from RefleXion Medical Inc and a departmental trainee research grant.


Monte Carlo, Phase Space


TH- External Beam- Photons: Computational dosimetry engines- Monte Carlo

Contact Email