Author: Jiankui Yuan, Dandan Zheng, Tingliang Zhuang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center, University of Rochester, Varian Medical Systems, Advanced Oncology Solutions 🌍
Purpose: In Monte Carlo (MC) radiation therapy dose calculations, latent variance exists when directly applying phase-space files (PSF) with a finite number of source particles, while the latter is practically limited by computational, data handling, and time costs. This novel work applies a generative adversarial network (GAN) to generate potentially unlimited source particles to eliminate this limitation.
Methods: A GAN was developed with three hidden layers, each containing 400 neurons, for both the generator and discriminator. It was trained on a 2x109 particle dataset from the Varian TrueBeam PSF for 6MV photons. Each source particle is described by energy (E), spatial coordinates (x and y), and direction (u and v). The Wasserstein loss function was applied to enhance the training stability. The outputs of the final layer of the model are unnormalized logits. The training process used a batch size of105 and the latent space dimension was set to 6. To evaluate the effectiveness of the model, we calculated dose distributions in water using GAN-generated particles compared with those calculated using the original PSF.
Results: The network contained about 0.6 million weight parameters, resulting in a model size of ~5 MB for the 6MV photon dataset, significantly smaller than the original PSF, which was ~60 GB. Averaged similarities between the original PSF and the GAN-generated one were greater than 0.99 on a validation dataset group of 107 particles. The PDD difference for a 10x10 field using the original vs. GAN particle sources was less than 1%.
Conclusion: We demonstrated that source particle distributions could be well learned by the novel application of a GAN architecture to achieve accurate MC dose calculations with increasing data and computation efficiency. This work paved the way for future work focused on bixel-wise source generation for IMRT/VMAT MC dose calculations.