Skip to main content
Open Access Publications from the University of California


UCLA Electronic Theses and Dissertations bannerUCLA

Accelerating Radiation Dose Calculation with High Performance Computing and Machine Learning for Large-scale Radiotherapy Treatment Planning

  • Author(s): Neph, Ryan
  • Advisor(s): Sheng, Ke
  • et al.

Radiation therapy is powered by modern techniques in precise planning and execution

of radiation delivery, which are being rapidly improved to maximize its benefit to cancer

patients. In the last decade, radiotherapy experienced the introduction of advanced methods

for automatic beam orientation optimization, real-time tumor tracking, daily plan

adaptation, and many others, which improve the radiation delivery precision, planning ease

and reproducibility, and treatment efficacy. However, such advanced paradigms necessitate

the calculation of orders of magnitude more causal dose deposition data, increasing the time

requirement of all pre-planning dose calculation. Principles of high-performance computing

and machine learning were applied to address the insufficient speeds of widely-used dose

calculation algorithms to facilitate translation of these advanced treatment paradigms into

clinical practice.

To accelerate CT-guided X-ray therapies, Collapsed-Cone Convolution-Superposition

(CCCS), a state-of-the-art analytical dose calculation algorithm, was accelerated through its

novel implementation on highly parallelized GPUs. This context-based GPU-CCCS approach

takes advantage of X-ray dose deposition compactness to parallelize calculation across

hundreds of beamlets, reducing hardware-specific overheads, and enabling acceleration by

two to three orders of magnitude compared to existing GPU-based beamlet-by-beamlet

approaches. Near-linear increases in acceleration are achieved with a distributed, multi-GPU

implementation of context-based GPU-CCCS.

Dose calculation for MR-guided treatment is complicated by electron return effects

(EREs), exhibited by ionizing electrons in the strong magnetic field of the MRI scanner. EREs

necessitate the use of much slower Monte Carlo (MC) dose calculation, limiting the clinical

application of advanced treatment paradigms due to time restrictions. An automatically

distributed framework for very-large-scale MC dose calculation was developed, granting

linear scaling of dose calculation speed with the number of utilized computational cores. It

was then harnessed to efficiently generate a large dataset of paired high- and low-noise MC

doses in a 1.5 tesla magnetic field, which were used to train a novel deep convolutional

neural network (CNN), DeepMC, to predict low-noise dose from faster high-noise MC-

simulation. DeepMC enables 38-fold acceleration of MR-guided X-ray beamlet dose

calculation, while remaining synergistic with existing MC acceleration techniques to achieve

multiplicative speed improvements.

This work redefines the expectation of X-ray dose calculation speed, making it possible

to apply new highly-beneficial treatment paradigms to standard clinical practice for the first


Main Content
Current View