pith. sign in

arxiv: 2604.11437 · v1 · submitted 2026-04-13 · 🧮 math.NA · cs.NA· physics.comp-ph

Data-efficient extraction of optical properties from 3D Monte Carlo TPSFs using Bi-LSTM transfer learning

Joubine Aghili , R\'emi Imbach , Anne Pallar\`es , Philippe Schmitt , Wilfried Uhring This is my paper

Pith reviewed 2026-05-10 15:23 UTC · model grok-4.3

classification 🧮 math.NA cs.NAphysics.comp-ph
keywords Bi-LSTMtransfer learningMonte Carlo simulationoptical propertiestime-resolved spectroscopyinverse problemturbid mediaphysics-informed learning
0
0 comments X

The pith

Transfer learning with a Bi-LSTM pre-trained on deterministic solvers and fine-tuned on few 3D Monte Carlo TPSFs extracts absorption and reduced scattering coefficients without analytical bias and at near-instant speed.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that pre-training a Bi-LSTM network on a fast deterministic solver creates a physical prior that allows effective fine-tuning on only a small number of 3D Monte Carlo time-point spread functions. This bridges the gap between biased but quick analytical models and accurate but expensive stochastic simulations for recovering optical properties in turbid media. A reader would care because time-resolved spectroscopy could then support real-time non-invasive measurements instead of remaining limited by computation time. The approach keeps error rates competitive while delivering instantaneous inference once trained.

Core claim

Pre-training a Bidirectional Long Short-Term Memory network on outputs from a deterministic solver establishes a physical prior; subsequent fine-tuning on a restricted set of 3D Monte Carlo simulations then eliminates the systematic bias of purely analytical models and recovers absorption μ_a and reduced scattering μ_s' coefficients from stochastic measurements with competitive accuracy and near-instantaneous inference time.

What carries the argument

Bi-LSTM transfer learning pipeline that first trains on deterministic solutions to learn physical structure before fine-tuning on limited 3D Monte Carlo TPSFs to close the analytical-to-stochastic domain gap.

If this is right

  • Real-time extraction of optical properties from 3D measurements becomes practical for time-resolved spectroscopy applications.
  • Only a small number of expensive stochastic simulations are required after deterministic pre-training.
  • The method applies directly to turbid media characterization without prohibitive computational overhead.
  • Inference remains near-instantaneous after training while accuracy stays competitive with full stochastic approaches.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same pre-train-then-fine-tune pattern could reduce data requirements in other inverse problems that mix deterministic approximations with stochastic simulations.
  • Extending the fine-tuning stage to include real experimental measurements might further close the simulation-to-experiment gap.
  • The approach suggests a general route for making stochastic forward models usable in real-time settings across imaging and sensing domains.

Load-bearing premise

Fine-tuning the Bi-LSTM on a small number of 3D Monte Carlo simulations removes analytical bias without introducing new errors or overfitting to the limited stochastic data.

What would settle it

A test set of independent 3D Monte Carlo TPSFs on which the fine-tuned model produces larger errors than the original analytical model would falsify the central claim.

Figures

Figures reproduced from arXiv: 2604.11437 by Anne Pallar\`es, Joubine Aghili, Philippe Schmitt, R\'emi Imbach, Wilfried Uhring.

Figure 1
Figure 1. Figure 1: (Left) Schematic representation of the TROT measurement device; (Right) a photog [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of the temporal decay profiles between the deterministic ( [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cumulative explained variance by Principal Components for both the Finite Difference [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Temporal Pearson correlation between TPSF time series and the optical parameters [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Architecture of the proposed Dual-Head Bi-LSTM. The normalized TPSF is processed [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Baseline model predictions evaluated on the deterministic [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: SHAP summary plot illustrating the impact of hyperparameters on the model’s success [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Scatter plots of the predicted optical properties (in blue dots) versus the ground truth [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
read the original abstract

Time-Resolved Spectroscopy (TRS) is a powerful modality for non-invasive characterization of turbid media. However, extracting optical properties, absorption $\mu_a$ and reduced scattering $\mu_s'$, from 3D stochastic measurements remains computationally expensive for real-time applications. In this paper, we propose a data-efficient, physics-informed transfer learning strategy using a Bidirectional Long Short-Term Memory (Bi-LSTM) network. By leveraging a fast deterministic solver to establish a physical prior before fine-tuning on a restricted set of 3D Monte Carlo simulations, our model successfully bridges the analytical-to-stochastic domain gap. The proposed method eliminates the systematic bias of analytical models while maintaining a competitive error with near-instantaneous inference time.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes a data-efficient physics-informed transfer learning strategy using a Bidirectional LSTM (Bi-LSTM) network to extract absorption coefficient μ_a and reduced scattering coefficient μ_s' from 3D Monte Carlo time point spread functions (TPSFs) in time-resolved spectroscopy. The approach pre-trains the network on a fast deterministic solver to establish a physical prior, then fine-tunes it on a restricted set of 3D Monte Carlo simulations to bridge the analytical-to-stochastic domain gap, claiming to eliminate the systematic bias of analytical models while achieving competitive error and near-instantaneous inference.

Significance. If the transfer-learning pipeline demonstrably removes analytical bias without introducing overfitting or new generalization errors on unseen 3D geometries and noise levels, the method would provide a practical route to real-time optical-property recovery that combines the speed of deterministic models with the accuracy of stochastic simulations, reducing the computational burden of pure Monte Carlo approaches for turbid-media characterization.

major comments (2)
  1. [Abstract] Abstract: the claim that the model 'eliminates the systematic bias of analytical models while maintaining a competitive error' is presented without any quantitative error values, bias metrics, validation splits, baseline comparisons (e.g., pure MC-trained Bi-LSTM or analytical-only), or statistical significance tests. This absence prevents verification of the central claim that fine-tuning successfully bridges the domain gap.
  2. [Methods (transfer learning pipeline)] Transfer-learning pipeline description: the assumption that pre-training on deterministic data followed by fine-tuning on a small set of 3D MC TPSFs yields unbiased recovery without residual bias or overfitting is load-bearing for the data-efficiency claim, yet no ablation on fine-tuning set size, no comparison to non-transfer baselines, and no generalization tests on held-out 3D geometries or noise levels are referenced.
minor comments (1)
  1. [Notation] Notation for optical properties (μ_a, μ_s') should be introduced once and used consistently; any subsequent re-definition of symbols should be avoided.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the presentation of our transfer-learning results. We address each major comment below and have revised the manuscript to strengthen the quantitative support for our claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that the model 'eliminates the systematic bias of analytical models while maintaining a competitive error' is presented without any quantitative error values, bias metrics, validation splits, baseline comparisons (e.g., pure MC-trained Bi-LSTM or analytical-only), or statistical significance tests. This absence prevents verification of the central claim that fine-tuning successfully bridges the domain gap.

    Authors: We agree that the abstract should be supported by explicit quantitative metrics. In the revised version we have added the mean absolute percentage errors for μ_a and μ_s' (with standard deviations), bias values relative to ground truth, details on the 80/20 train/validation split used for fine-tuning, and direct numerical comparisons against both a pure Monte-Carlo-trained Bi-LSTM baseline and the analytical model. Statistical significance (paired t-tests) is now reported in the results section and referenced from the abstract. revision: yes

  2. Referee: [Methods (transfer learning pipeline)] Transfer-learning pipeline description: the assumption that pre-training on deterministic data followed by fine-tuning on a small set of 3D MC TPSFs yields unbiased recovery without residual bias or overfitting is load-bearing for the data-efficiency claim, yet no ablation on fine-tuning set size, no comparison to non-transfer baselines, and no generalization tests on held-out 3D geometries or noise levels are referenced.

    Authors: We have expanded the methods and results sections with the requested analyses. New figures and tables now show: (i) ablation curves for fine-tuning set sizes from 50 to 500 TPSFs, (ii) side-by-side error metrics for the transfer-learned model versus a non-transfer Bi-LSTM trained only on Monte Carlo data, and (iii) generalization performance on held-out 3D slab and cylindrical geometries as well as across SNR levels from 20 dB to 40 dB. These additions directly substantiate the data-efficiency and bias-removal claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity; transfer-learning pipeline is self-contained

full rationale

The paper presents a standard transfer-learning pipeline: pre-train Bi-LSTM on outputs from a fast deterministic solver, then fine-tune on a small set of 3D Monte Carlo TPSFs. No equations, procedures, or self-citations in the provided text reduce the claimed bias elimination or error performance to a quantity defined by the same data or by construction. The central claim rests on the empirical effectiveness of the pre-training/fine-tuning sequence using independent external sources, which does not collapse into a tautology or fitted-input prediction. This is the normal non-circular case for a data-driven method whose performance is evaluated against held-out stochastic data.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the untested premise that a deterministic solver supplies a sufficiently rich physical prior for subsequent fine-tuning on sparse stochastic data; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)
  • domain assumption A fast deterministic light-transport solver provides a useful physical prior that can be transferred to correct biases in 3D Monte Carlo TPSFs.
    Invoked to justify the two-stage training procedure described in the abstract.

pith-pipeline@v0.9.0 · 5442 in / 1272 out tokens · 64387 ms · 2026-05-10T15:23:42.407294+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages

  1. [1]

    Ontheradiativeequilibriumofastellaratmosphere.II.The Astrophysical Journal, 100:76, July 1944

    S.Chandrasekhar. Ontheradiativeequilibriumofastellaratmosphere.II.The Astrophysical Journal, 100:76, July 1944

    work page 1944

  2. [2]

    Transfer learning for time series classification.2018 IEEE International Conference on Big Data (Big Data), pages 1367–1376, 2018

    Hassan Ismail Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, and Pierre- Alain Muller. Transfer learning for time series classification.2018 IEEE International Conference on Big Data (Big Data), pages 1367–1376, 2018

    work page 2018

  3. [3]

    Fussen, N

    D. Fussen, N. Baker, A. Berthelot, E. Dekemper, P. Gramme, N. Mateshvili, K. Rose, and S. Sotiriadis. Direct inversion method for the retrieval of ozone number density profiles from observations of solar radiation scattering by the atmospheric limb.Journal of Quantitative Spectroscopy and Radiative Transfer, 339:109426, July 2025

    work page 2025

  4. [4]

    MachineLearning Approaches for the Inversion of the Radiative Transfer Equation

    EstebanGarcia-Cuesta, FernandoDeLaTorre, andAntonioJ.DeCastro. MachineLearning Approaches for the Inversion of the Radiative Transfer Equation. In Sio-Iong Ao, Burghard Rieger, and Su-Shing Chen, editors,Advances in Computational Algorithms and Data Anal- ysis, volume 14, pages 319–331. Springer Netherlands, Dordrecht, 2009

    work page 2009

  5. [5]

    Bidirectional lstm networks for improved phoneme classification and recognition

    Alex Graves, Santiago Fernández, and Jürgen Schmidhuber. Bidirectional lstm networks for improved phoneme classification and recognition. InProceedings of the 15th International Conference on Artificial Neural Networks: Formal Models and Their Applications - Volume Part II, ICANN’05, page 799–804, Berlin, Heidelberg, 2005. Springer-Verlag

    work page 2005

  6. [6]

    Optical properties of biological tissues: A review.Physics in Medicine and Biology, 58(11):R37–R61, June 2013

    Steven L Jacques. Optical properties of biological tissues: A review.Physics in Medicine and Biology, 58(11):R37–R61, June 2013

    work page 2013

  7. [7]

    Korkin, A.M

    S. Korkin, A.M. Sayer, A. Ibrahim, and A. Lyapustin. A practical guide to writing a radiative transfer code.Computer Physics Communications, 271:108198, February 2022

    work page 2022

  8. [8]

    Artificial neural network method for determining optical properties from double-integrating-spheres measurements.Chinese Optics Letters, 8(2):173, 2010

    Chenxi Li, Huijuan Zhao, Qiuyin Wang, and Kexin Xu. Artificial neural network method for determining optical properties from double-integrating-spheres measurements.Chinese Optics Letters, 8(2):173, 2010. 13

    work page 2010

  9. [9]

    Lundberg and Su-In Lee

    Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, pages 4768–4777, Long Beach, California, USA, 2017. Curran Associates Inc

    work page 2017

  10. [10]

    Physics informed neural networks for simulating radiative transfer.Journal of Quantitative Spectroscopy and Radiative Transfer, 270:107705, August 2021

    Siddhartha Mishra and Roberto Molinaro. Physics informed neural networks for simulating radiative transfer.Journal of Quantitative Spectroscopy and Radiative Transfer, 270:107705, August 2021

    work page 2021

  11. [11]

    Comparison of time resolved optical turbidity measurements for water monitoring to standard real-time techniques.Sensors, 21(9):3136, April 2021

    Anne Pallarès, Philippe Schmitt, and Wilfried Uhring. Comparison of time resolved optical turbidity measurements for water monitoring to standard real-time techniques.Sensors, 21(9):3136, April 2021

    work page 2021

  12. [12]

    Richling, E

    S. Richling, E. Meinköhn, N. Kryzhevoi, and G. Kanschat. Radiative transfer with finite elements: I. Basic method and tests.Astronomy & Astrophysics, 380(2):776–788, December 2001

    work page 2001

  13. [13]

    Savitzky and M

    Abraham. Savitzky and M. J. E. Golay. Smoothing and differentiation of data by simplified least squares procedures.Analytical Chemistry, 36(8):1627–1639, July 1964

    work page 1964

  14. [14]

    AnEfficientSolutionTechniquefortheRadiativeTransferEquation.IMPACT of Computing in Science and Engineering, 5(3):201–214, September 1993

    StefanTurek. AnEfficientSolutionTechniquefortheRadiativeTransferEquation.IMPACT of Computing in Science and Engineering, 5(3):201–214, September 1993

    work page 1993

  15. [15]

    Yaru Zhang, Wenxing Bai, Yihan Dong, Mai Dan, Dongyuan Liu, and Feng Gao. Deep- learning approach to stratified reconstructions of tissue absorption and scattering in time- domain spatial frequency domain imaging.Journal of Biomedical Optics, 29(03), March 2024. 14

    work page 2024