Author: Yankui Chang, Shijun Li, Xi Pei, Ripeng Wang, Xuanhe Wang, X. George Xu, Qing Zhang, Jingfang Zhao 👨🔬
Affiliation: University of Science and Technology of China, Shanghai proton and heavy ion center, School of Nuclear Science and Technology, University of Science and Technology of China, Anhui Wisdom Technology Co., Ltd. 🌍
Purpose:
This paper describes disruptive methods using both GPU-based MC simulation and deep-learning (DL)-based MC denoising algorithms, as well as clinical tests involving more than 560 patient plans covering a range of proton and carbon ion treatment sites.
Methods:
The particle interactions and transport processes follow those in the GEANT4 code. The software parallelization was achieved through the CPU/GPU coprocessor architecture. Once validated against TOPAS for radiological physics, the codes were used first to generate an extremely large amount of training and testing data for DL-based convolutional encoder-decoder neural network (dCNN). The datasets consisted 560 cancer patient cases including the head-and-neck, lung, cervical, prostate and other sites. The GPU-based MC codes were used to produce the input dose distributions of 6 × 105 simulated particles and the reference dose (ground truth) of 1 × 108 simulated particles. Root mean squared error (RMSE) and Gamma Passing Rate (GPR, 2%/2mm) are used to quantitatively evaluate the performances. These MC code codes are finally integrated into the ARCHER dose engine to facilitate clinical studies.
Results:
The GPRs were found to exceed 99%, demonstrating the excellent accuracy of the GPU-based proton and carbon ion MC codes. In the denoising approach, the RMSE and the mean GPR are found to be 0.012 and 99.3%, respectively, suggest a strong agreement between the denoised dose and the reference dose. With an NVIDIA GeForce RTX 3090 card, the running time was found to be less than 0.2 seconds per patient case for protons and less than 0.4 seconds for carbon ions. Methods of GPU acceleration have been applied to helium ions to yield on the average of 10 seconds per patient case.
Conclusion:
We have demonstrated the feasibility of performing real-time proton, carbon ion and (possibly) helium ion MC dose calculations.