Author: Lu Jiang, Shusen Jing, Qihui Lyu, Dan Ruan, Ke Sheng, Qifan Xu 👨🔬
Affiliation: Department of Radiation Oncology, University of California, Los Angeles, University of California San Francisco, Department of Radiation Oncology, University of California at San Francisco, Department of Radiation Oncology, University of California, San Francisco 🌍
Purpose: In radiotherapy, the conformity and compactness of dose distribution are vital to patient outcomes. The introduction of highly complex planning, such as 4π radiotherapy, has provided a systematic approach to optimally utilizing non-coplanar beams to achieve superior dosimetry. However, the computational cost of 4π treatment planning remains high due to the need to precompute a large dose-loading matrix of all candidate beams and handle such matrices during optimization. This work presents an ultra-high-performance parallel (UHPP) framework designed to expedite high-dimensional treatment planning.
Methods: We adopted synchronized parallel ray tracing in the non-voxel-based (NVB) beamlet dose calculation. Additionally, we sparsified dense dose-loading matrices using the cuSPARSE library native to CUDA. We then adopted a collapsed-cone convolution superposition (CCCS) exponential kernel calculation module to account for varying LINAC spectra. We implemented the group sparsity algorithm for beam orientation optimization using the cuSPARSE library to accelerate its execution on GPUs. We benchmarked dose calculation accuracy against Monte Carlo (MC) methods and compared planning performance on five pancreas and eight head-and-neck patients to dose distributions from state-of-the-art (SOTA) 4π non-coplanar and clinical VMAT coplanar plans.
Results: UHPP consistently achieved better gamma passing rates than 98.83% at 2%/2 mm compared to MC in water and slab phantoms with varying beamlet widths. While maintaining comparable plan quality for pancreas and head-and-neck patients, UHPP achieved speedups of 8.521× in dose calculation and 12.719× in treatment plan optimization, respectively, alongside additional and more significant gains in data I/O and preprocessing efficiency compared to the SOTA 4π method. Both 4π plans were consistently superior to clinical VMAT plans in terms of OAR sparing and PTV coverage.
Conclusion: The proposed framework provides high dose accuracy and significant speedups while preserving the dosimetric advantages of 4π planning over standard coplanar VMAT plans, facilitating integration into responsive clinical practice.