Framework for denoising Monte Carlo photon transport simulations using deep learning

Area of Science:

• Biomedical Optics
• Computational Imaging
• Medical Physics

Background:

• Monte Carlo (MC) simulations are crucial for modeling light propagation in turbid media like human tissues.
• However, MC simulations suffer from inherent stochastic noise, requiring extensive photon counts and leading to high computational costs.
• Accelerating MC simulations is vital for improving efficiency in various biomedical applications.

Purpose of the Study:

• To develop and evaluate deep learning (DL) based image denoising techniques for low-photon MC simulations.
• The goal is to enhance the quality of MC simulation results and accelerate the overall MC method.
• To compare the performance of different DL architectures against traditional denoising algorithms.

Main Methods:

• Developed a cascade-network combining DnCNN and UNet architectures for MC data denoising.
• Evaluated established DL denoising networks: DnCNN, UNet, DRUNet, and ResMCNet (deep residual-learning for denoising MC renderings).
• Created a novel approach for generating synthetic datasets to train DL-based MC denoisers and compared DL methods against model-based algorithms.

Main Results:

• DL-based denoising significantly outperformed traditional model-based methods in image quality.
• The proposed cascade network achieved 14-19 dB SNR improvement, equivalent to 25-78x more photons.
• The cascade network excelled in complex domains (brain, mouse atlases), while ResMCNet and DRUNet performed better with high-photon inputs.

Conclusions:

• State-of-the-art DL denoising techniques can reduce MC simulation computation time by one to two orders of magnitude.
• DL-based denoising offers a powerful approach to accelerate MC simulations without compromising image quality.
• Open-source codes and data are available to facilitate further research and application.