Graphics processing units-accelerated adaptive nonlocal means filter for denoising three-dimensional Monte Carlo photon transport simulations

Yaoshen Yuan, Leiming Yu, Zafer Doğan, Qianqian FangJournal of Biomedical Optics 2018 · DOI

TL;DR

This paper introduces a novel GPU-accelerated Adaptive Nonlocal Means (ANLM) filter designed to significantly reduce noise in 3D Monte Carlo (MC) simulations of light transport. By achieving over 5 dB SNR improvement, equivalent to running 3.5x more photons, this filter dramatically speeds up simulations while preserving crucial image details, making it a valuable tool for biophotonics and other volumetric imaging applications.

What Problem Does This Paper Address?

Accurately simulating light propagation in biological tissues is crucial for advancing biophotonics techniques, particularly in medical imaging and diagnostics. The gold standard for this is the Monte Carlo (MC) method, which simulates the path of individual photons. However, MC simulations are inherently noisy due to their stochastic nature. While increasing the number of simulated photons can reduce this noise, it drastically increases computation time, even with GPU acceleration. This creates a trade-off between simulation accuracy (low noise) and computational feasibility (speed). Existing denoising methods often struggle with the specific characteristics of MC photon transport data, such as spatially varying noise and the need to preserve sharp optical fluence changes, especially in complex heterogeneous media.

Methods

The core contribution is a GPU-accelerated Adaptive Nonlocal Means (ANLM) filter. The ANLM filter works by averaging the values of neighboring voxels, but unlike simpler filters, it considers a larger "neighborhood" of voxels and adaptively weights their contributions based on their similarity to the target voxel. This allows it to effectively reduce noise while preserving important image features like edges and sharp transitions.

Key innovations in this work include:

  • GPU Acceleration: The entire ANLM filtering process is implemented on GPUs, leveraging their massive parallel processing capabilities to achieve significant speedups.
  • Adaptive Noise Estimation: The filter is designed to handle spatially varying noise, a common characteristic of MC simulations where noise levels differ across the image.
  • Preservation of Sharp Features: The ANLM algorithm is inherently good at preserving edges and fine details, which is critical for accurately representing optical fluence changes in biological tissues.
  • Optimized GPU Implementation: The authors introduce several optimizations for the GPU implementation, including:
  • Loading preprocessing outputs to shared memory.
  • Utilizing 3D thread blocks for better cache utilization.
  • Performing preprocessing steps (variance, residual calculations) directly on the GPU.
  • Streamlining multi-step filtering processes on the GPU to minimize data transfer.
  • Support for Rician Noise: While primarily targeting Gaussian/Poisson noise in MC, the filter is also adapted to handle Rician noise, making it applicable to other volumetric imaging modalities like MRI.

Datasets

The authors used simulated datasets for evaluation. They created three benchmark scenarios:

  • B1: Homogeneous Domain: A simple, uniform medium.
  • B2: Absorbing Inclusion: A heterogeneous domain with a cubic region of higher absorption.
  • B3: Refractive Inclusion: A heterogeneous domain with a cubic region of higher refractive index.

These benchmarks were tested with varying photon counts (10⁵ to 10⁸) and included simulations of a realistic 19.5-year-old brain atlas.

Key Insights and Ideas

  • Denoising as a Practical Alternative to More Photons: The ANLM filter provides a significant SNR boost (over 5 dB) that is directly comparable to running substantially more photons (3.5x). This offers a practical way to achieve high-quality simulations without the prohibitive computational cost of simply increasing photon counts.
  • Preservation of Critical Features: The filter's ability to preserve sharp edges and fluence discontinuities is crucial for accurately modeling light propagation in complex biological structures with varying optical properties (e.g., inclusions). This is a key advantage over simpler smoothing filters.
  • GPU Acceleration is Key: The paper strongly demonstrates that GPU acceleration is essential for making advanced denoising techniques like ANLM computationally viable for the massive datasets generated by MC simulations.
  • Adaptive Nature is Crucial: The filter's adaptiveness to spatially varying noise ensures that it effectively targets noisy regions without over-smoothing areas that already have good signal-to-noise ratios.
  • Broad Applicability: The filter's design and demonstrated performance suggest its utility beyond MC photon transport, extending to other volumetric imaging data like MRI and CT scans.

Key Results

  • SNR Improvement: The GPU-ANLM filter achieved an average SNR improvement of over 5 dB, which is equivalent to running approximately 3.5 times more photons.
  • Speedup:
    • The GPU-ANLM filter showed a 2.5x speedup compared to previous GPU-ANLM implementations.
  • It achieved a 3-4x speedup compared to CPU-based ANLM implementations.
  • Overall runtime reductions were significant, with optimizations leading to nearly a 60% reduction in total filtering time.
  • Edge Preservation: Qualitative results showed excellent preservation of sharp features, even at inclusion boundaries, in heterogeneous media.
  • Robustness: The denoising performance and speedup were largely independent of the media's heterogeneity and photon counts.

Limitations and Future Work

  • Effective Region Limitation: The SNR improvement from the ANLM filter was observed to be limited to regions distal from the source, with effectiveness gradually reducing in regions with already high SNRs. This is a consequence of its adaptive nature but means the benefit is not uniform across the entire simulation volume.
  • Comparison with Other Advanced Filters: While the paper compares its GPU-ANLM to a CPU-BM4D filter, it acknowledges that exploring and accelerating other contemporary 3D adaptive filters like BM4D could be a promising future direction.
  • Optimization of Other Filters: The authors suggest that accelerating other advanced 3D adaptive filters could also yield significant benefits.
  • Broader Applications: The paper highlights the potential for using the GPU-ANLM filter for denoising other types of 3D volumetric data, such as MRI and CT scans, suggesting future work in these areas.

Summary updated

Abstract

The Monte Carlo (MC) method is widely recognized as the gold standard for modeling light propagation inside turbid media. Due to the stochastic nature of this method, MC simulations suffer from inherent stochastic noise. Launching large numbers of photons can reduce noise but results in significantly greater computation times, even with graphics processing units (GPU)-based acceleration. We develop a GPU-accelerated adaptive nonlocal means (ANLM) filter to denoise MC simulation outputs. This filter can effectively suppress the spatially varying stochastic noise present in low-photon MC simulations and improve the image signal-to-noise ratio (SNR) by over 5 dB. This is equivalent to the SNR improvement of running nearly 3.5 × more photons. We validate this denoising approach using both homogeneous and heterogeneous domains at various photon counts. The ability to preserve rapid optical fluence changes is also demonstrated using domains with inclusions. We demonstrate that this GPU-ANLM filter can shorten simulation runtimes in most photon counts and domain settings even combined with our highly accelerated GPU MC simulations. We also compare this GPU-ANLM filter with the CPU version and report a threefold to fourfold speedup. The developed GPU-ANLM filter not only can enhance three-dimensional MC photon simulation results but also be a valuable tool for noise reduction in other volumetric images such as MRI and CT scans.
BibTeX
@article{yuan2018graphics,
  author = {Yaoshen Yuan and Leiming Yu and Zafer Doğan and Qianqian Fang},
  title = {{Graphics processing units-accelerated adaptive nonlocal means filter for denoising three-dimensional Monte Carlo photon transport simulations}},
  journal = {Journal of Biomedical Optics},
  year = {2018},
  doi = {10.1117/1.jbo.23.12.121618},
}
Figure