Framework for Denoising Monte Carlo Photon Transport Simulations Using Deep Learning

Abstract

Significance: The Monte Carlo (MC) method is widely used as the gold-standard for modeling light propagation inside turbid media like human tissues, but combating its inherent stochastic noise requires one to simulate large number photons, resulting in high computational burdens. Aim: We aim to develop an effective image denoising technique using deep-learning (DL) to dramatically improve low-photon MC simulation result quality, equivalently bringing further acceleration to the MC method. Approach: We have developed a cascade-network combining DnCNN with UNet, in the meantime, extended a range of established image denoising neural-network architectures, including DnCNN, UNet, DRUNet, and ResMCNet, in handling three-dimensional (3-D) MC data and compared their performances against model-based denoising algorithms. We have also developed a simple yet effective approach to create synthetic datasets that can be used to train DL based MC denoisers. Results: Overall, DL based image denoising algorithms exhibit significantly higher image quality improvements over traditional model-based denoising algorithms. Among the tested DL denoisiers, our Cascade network yields a 14 - 19 dB improvement in signal-to-noise ratio (SNR), which is equivalent to simulating 25x to 78x more photons. Other DL-based methods yielded similar results, with our method performing noticeably better with low-photon inputs, and ResMCNet along with DRUNet performing better with high-photon inputs. Our Cascade network achieved the highest quality when denoising complex domains, including brain and mouse atlases. Conclusion: Incorporating state-of-the-art DL denoising techniques can equivalently reduce the computation time of MC simulations by one to two orders of magnitude. Our open-source MC denoising codes and data can be freely accessed at http://mcx.space/.

Efficient Photoacoustic Image Synthesis with Deep Learning

“…An alternative approach for fast photon propagation models might be performing Monte Carlo simulations with lower resolution or fewer photons and, respectively, upsample or denoise afterwards. Ardakani et al presented a deep neural network that denoises low photon simulations to achieve more accurate, fast simulations [42]. Their model is, however, not differentiable and thus not suitable for gradient-based learning-to-simulate approaches for photoacoustic images.…”