(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: Monte Carlo (MC) simulation is considered to be the most-accurate of the available radiotherapy dose calculation methods, however, its clinical use is limited due to the extensive required computation times. We propose to train a Generative Adversarial Network (GAN) to predict high-resolution low-noise (HRLN) dose distributions from low-resolution high-noise (LRHN) MC dose distributions to circumvent the long calculation times without sacrificing accuracy of the MC calculations.
Methods: A model of a clinical 6MV photon beam was constructed using schematics and phase space data provided by the manufacturer. Simulations were performed in the dosxyznrc and BEAMnrc user codes of EGSnrc where dose distributions were generated at multiple depths for a 10x10cm2 field. Voxel resolutions of 4x4x4mm3 and 1x1x1mm3 in a homogeneous water phantom, with corresponding simulation uncertainties of 5.0% and 0.7%, respectively, were considered. The Enhanced Super-Resolution GAN was trained to generate HRLN dose distributions from the LRHN dose distributions at depths of 2, 5, and 10cm. Testing was performed on the trained depths plus additional depths. Each training sample consisted of an input-output pair of a LRHN and HRLN dose distribution, respectively.
Results: The trained model accurately predicted HRLN dose distributions from LRHN dose distributions, both at depths that were included and excluded in the training data. The mean relative error of the predicted dose distributions across the central 80% were between 1.2% and 2.4% for depths ranging from 2 to 20cm.
Conclusion: The model presented here is the first step in combining MC with a GAN for accurate dose calculations with significantly reduced computation times. Our experiments show that the generated dose distributions are comparable to those generated from analog HRLN MC simulations. Future research includes expanding the model to include additional radiation field sizes and asymmetric shapes.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
