Analysis of relative error in perturbation Monte Carlo simulations of radiative transport

Mahsa Parsanasab, Carole K. Hayakawa, Jerome Spanier, Yanning Shen, Vasan VenugopalanJournal of Biomedical Optics 2023 · DOI

TL;DR

This paper introduces a novel method to predict the accuracy of perturbation Monte Carlo (pMC) simulations without running computationally expensive simulations for every possible optical property. By analyzing statistical properties of a single "reference" simulation, researchers can now estimate the reliable range of optical property variations for which pMC will yield accurate results, significantly streamlining inverse problem solving in diffuse optics.

What Problem Does This Paper Address?

Monte Carlo (MC) simulations are the gold standard for modeling light transport in biological tissues, offering high accuracy. However, running many MC simulations for complex problems like inverse scattering (e.g., determining tissue properties from light measurements) is computationally prohibitive. Perturbation Monte Carlo (pMC) offers a speedup by calculating changes relative to a reference simulation. The challenge is that pMC accuracy degrades with larger changes in optical properties, and it's unclear a priori how large a change is acceptable. This limits pMC's application, especially for multispectral imaging where optical properties can vary significantly. Researchers need a way to predict the accuracy of pMC before running it, to avoid wasted computation and ensure reliable results.

Methods

The core of this work is developing a predictive method for pMC relative error. Instead of running pMC for every perturbation, the authors leverage error propagation theory. They analyze statistical moments (variance, covariance, skewness) of photon weight, path length, and collision number from a single reference MC simulation. This information is then used to estimate the variance of the perturbed photon weights, which directly relates to the pMC relative error.

Key methodological contributions include:

  • Error Propagation on Log-Transformed Weights: To handle the wide dynamic range of photon weights, they apply error propagation to the logarithm of the weights.
  • Refined Variance Estimation: They adapt the standard error propagation formula to account for tallied and untallied photons in the simulation, providing a more accurate estimate of the overall variance.
  • Impact of Russian Roulette (RR): They investigate the use of the Russian Roulette technique in the reference simulations. RR is a variance reduction method that helps terminate low-weight photons, leading to more compact distributions of photon properties and, as they show, improving the accuracy of their predictive method.

Datasets

The authors used simulated data generated from conventional Monte Carlo (cMC) simulations. They created five distinct media with varying ratios of reduced scattering to absorption coefficients ($\mu'_s / \mu_a$), ranging from 5 to 100. For each medium, they simulated light transport with a directional point source and collected data at two detector locations: a proximal detector (0-0.2 mm from the source) and a distal detector (3-3.2 mm from the source). These simulations were performed both with and without the Russian Roulette technique. No new experimental datasets were introduced.

Key Insights and Ideas

  • Predicting pMC Accuracy is Possible: The fundamental insight is that you don't need to run pMC for every scenario to know its accuracy. Statistical properties of a single reference simulation hold the key.
  • Low $\mu'_s / \mu_a$ is Advantageous for Reference Simulations: Running reference simulations in media with low scattering-to-absorption ratios (i.e., more absorbing media) leads to more accurate predictions of pMC error. This is because the distributions of photon properties are less "spiky" and more amenable to the error propagation assumptions.
  • Russian Roulette is a Game Changer: Implementing RR in the reference simulation significantly improves the accuracy of the predictive method and expands the range of scattering perturbations that can be reliably analyzed. It effectively "tames" the long, sparse tails in photon property distributions.
  • Detector Location Matters: The predictive method works better for proximal detectors than distal ones. This is likely due to the more direct path of photons to proximal detectors, leading to less complex photon histories.
  • Conservative Estimates: The proposed method tends to overestimate the pMC relative error, which is a desirable characteristic for ensuring that the predicted reliable range is indeed safe.

Key Results

  • Proximal Detector Accuracy: The method can estimate pMC relative error within 5% of the true value for scattering perturbations up to ±20% when using RR and low $\mu'_s / \mu_a$ ratios.
  • Distal Detector Accuracy: For distal detectors, the method provides estimates within 20% for scattering perturbations up to ±15% (with RR and low $\mu'_s / \mu_a$).
  • Impact of $\mu'_s / \mu_a$: Lower $\mu'_s / \mu_a$ ratios consistently yield better accuracy for the predictive method, especially for the distal detector.
  • RR Improvement: The use of RR reduced the normalized variance and covariance of photon properties, leading to a significant improvement in the accuracy of the pMC error estimation. For example, the normalized variance of log(WR) for the distal detector dropped from 12.07 to 0.489 (for $\mu'_s / \mu_a = 5$) after applying RR.

Limitations and Future Work

  • Homogeneous Medium Assumption: The current method is demonstrated for homogeneous media. Extending it to heterogeneous media, where perturbations are localized, is a logical next step. The authors suggest that their methodology will likely perform even better in such cases.
  • Scattering Perturbations Only: The study focused on scattering perturbations. While absorption perturbations can be handled exactly with continuous absorption weighting (CAW), further investigation into combined perturbations might be beneficial.
  • "Worst-Case" Scenario: The paper focuses on perturbations applied to the entire medium. The authors anticipate that localized perturbations in real-world applications will allow for a larger range of reliable scattering perturbations.
  • Further Optimization of RR: While RR significantly improved results, exploring different weight thresholds and survival probabilities could further refine the process.

Summary updated

Abstract

These findings indicate that reference simulations utilizing continuous absorption weighting (CAW) with the Russian Roulette method and executed using optical properties with a low <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:mo>(</mml:mo><mml:msubsup><mml:mrow><mml:mi>μ</mml:mi></mml:mrow><mml:mrow><mml:mi>s</mml:mi></mml:mrow><mml:mrow><mml:mo>'</mml:mo></mml:mrow></mml:msubsup><mml:mo>/</mml:mo><mml:msub><mml:mrow><mml:mi>μ</mml:mi></mml:mrow><mml:mrow><mml:mi>a</mml:mi></mml:mrow></mml:msub><mml:mo>)</mml:mo></mml:mrow></mml:math> ratio spanning the desired range of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:msub><mml:mrow><mml:mi>μ</mml:mi></mml:mrow><mml:mrow><mml:mi>s</mml:mi></mml:mrow></mml:msub></mml:mrow></mml:math> values, are highly advantageous for the deployment of pMC to obtain radiative transport estimates over a wide range of optical properties.
BibTeX
@article{parsanasab2023analysis,
  author = {Mahsa Parsanasab and Carole K. Hayakawa and Jerome Spanier and Yanning Shen and Vasan Venugopalan},
  title = {{Analysis of relative error in perturbation Monte Carlo simulations of radiative transport}},
  journal = {Journal of Biomedical Optics},
  year = {2023},
  doi = {10.1117/1.jbo.28.6.065001},
}

Loading related content...

Figure