acceptodds
Withdrawn from ICLR 2027

High-Fidelity Monte Carlo Reconstruction from Finite-Photon Simulations with a Photon-Budget-Conditioned Volumetric State-Space Model

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

High-fidelity Monte Carlo (MC) simulation demands substantial sampling effort. We make the photon budget an explicit conditioning variable for learned reconstruction, allowing one volumetric model to adapt across eight sampling regimes while retaining MC as the forward solver. Photon-budget affine modulation coordinates directional spatial gating with Fourier-domain Mamba-3 processing. On a three-dimensional skin-lesion benchmark, the model improves mean PSNR by 4.82 dB over the strongest macro-average baseline (configuration-paired 95% interval: [4.57, 5.07] dB). Gains extend to inverse-transformed field summaries and tested held-out physical regimes, with capacity-matched controls supporting the design. Reconstruction from photons exceeds raw -photon MC in mean PSNR, SSIM, and RMSE against matched references, using fewer input photons. At photons, MC plus inference achieves an same-GPU speedup over direct -photon MC, with a 0.67-dB mean PSNR gap against independent -photon references; a separate CPU-MC/GPU-inference workflow gives approximately timing speedup at the same budgets. These results establish photon-budget conditioning as a practical route to efficient volumetric MC reconstruction and motivate using known sampling effort as a design variable for learned correction in scientific simulation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Monte Carlo Simulation, Finite-Photon Reconstruction, Photon-Budget Conditioning, State-Space Models, Scientific Machine Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_AwncOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a photon-budget-conditioned volumetric state-space model for reconstructing high-fidelity Monte Carlo (MC) photon transport simulations from low-photon inputs. Experimental results demonstrate improvements over several reconstruction baselines, with a reported average PSNR gain of 4.82 dB over the strongest baseline.

Critical strengths
  1. The proposed method achieves consistently strong reconstruction performance across a wide range of photon budgets.
  2. Incorporating photon-budget information into a shared reconstruction model is practically meaningful.
Critical weaknesses
  1. The proposed framework mainly combines existing techniques, including a 3D encoder-decoder, directional spatial gating, Fourier-domain state-space modeling, and affine feature modulation.
  2. All experiments are based on a single family of synthetically generated skin-lesion models, using one wavelength (633 nm), a fixed illumination configuration, and a fixed 64×64×64 resolution.
Questions for authors
  1. What is the fundamental methodological innovation beyond integrating existing techniques?
  2. How does the proposed model perform under substantially different tissue geometries, illumination configurations, wavelengths, and spatial resolutions?
  3. can the authors evaluate whether the reconstructed fields preserve important physical properties, such as energy deposition consistency, local absorption distributions, and accuracy near tissue interfaces?
Reviewer_Mi46OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies learned reconstruction of three-dimensional Monte Carlo (MC) photon-transport absorption fields while retaining MC as the forward solver. A residual encoder-decoder combines directional spatial gating, Fourier-domain Mamba-3 processing, and photon-budget affine modulation (PBAM). A categorical budget embedding allows one 4.21M-parameter model to handle eight budgets from 10² to 10⁶ photons. Training uses log-preprocessed, Anscombe-transformed fields paired with 10⁸-photon references. Evaluation uses a configuration-disjoint 243/81/81 split of 405 synthetic skin-lesion configurations, architectural and conditioning ablations, independent 10⁹-photon references, and two held-out physical-parameter splits. The main results are a 4.82 dB macro-average PSNR gain over ReconFormer and an 8.09× same-GPU simulation-plus-inference speedup over direct 10⁸-photon MC, with a 0.67 dB PSNR deficit against a common independent reference (Tables E10 and 5).

Critical strengths
  1. Large reconstruction gains supported by complementary checks. The 4.82 dB macro-average gain over ReconFormer has a configuration-paired 95% interval of [4.57, 5.07] dB, with positive configuration-level macro gains for all 81 test cases. Comparisons against the strongest observed architectural baseline support superiority at seven of eight budgets (Table E10). Configuration-disjoint splitting and evaluation against independent 10⁹-photon references without retraining or checkpoint reselection reduce important leakage and reference-dependence concerns. This makes the empirical result more persuasive than a comparison of aggregate image-quality scores alone.
  2. A useful unified model with meaningful internal controls. PBAM improves average PSNR by 1.14 dB over the same unconditioned backbone, and one 4.21M-parameter model replaces eight photon-specific models with approximately 7.8× lower aggregate parameter storage (Table 3). The unified model has better mean PSNR and SSIM, although slightly worse mean RMSE. The matched slice-wise comparison and capacity-controlled convolutional and spatial-Mamba alternatives further support the value of volumetric context and the proposed representation (Tables 2, 4, and E12). These are substantive controls, not merely comparisons against smaller networks.
Critical weaknesses
  1. The physical interpretation remains uncertain because the raw-MC discrepancy is unresolved. Table E11 reports integrated absorption errors of 45.87% at 10⁴ photons and 45.88% at 10⁶; the corresponding depth-profile errors are 67.73% and 67.88%. Appendix E.7 explicitly leaves the cause unresolved. Sensitivity of inverse-log metrics to high-valued voxels explains how good log-domain PSNR can coexist with large field errors, but does not explain why these aggregate errors barely change over a 100-fold budget increase. Without a native-output and normalization audit, I cannot distinguish correction of finite-sampling noise from correction of a systematic simulation/preprocessing/reference discrepancy. This is a possibility to investigate, not evidence of an established implementation bug. Since physical reconstruction fidelity is central to the paper, the distinction matters. Moreover, the reported field metrics are summaries of inverse-transformed log-resampled fields, not native-grid local accuracy or an energy-conservation check (Appendix C.4).

  2. The advantage over simpler budget-aware alternatives is not sufficiently established. Tables D2–D3 show that most architectural baselines do not receive the photon budget as a network input, while objectives and optimization settings also differ. VEMamba is an important exception, but uses a different conditioning pathway. The PBAM ablation establishes a benefit within the proposed backbone; it does not establish how much advantage remains over a 3D U-Net, MC-Fourier, or the capacity-matched convolutional control given the same conditioning and training objective. The spatial-Mamba control is useful, but its 0.4195 dB macro-average gap comes from a single training run and also changes projections and local mixing, so it does not isolate the Fourier transform (Appendix E.8). These gaps matter because the principal claim is a budget-conditioned reconstruction strategy built from established components. Evidence is also confined to one skin-lesion generator and fixed illumination; the held-out absorption/scale tests are valuable but do not test new shape families or unseen photon budgets (Section 5 and Appendix F).

Questions for authors
  1. Clarification and validation of the existing pipeline: What causes the approximately budget-independent raw-MC integrated absorption error in Table E11? Please report native-grid integrated absorption against independent references across photon budgets and repeated MC realizations, and specify normalization at each stage. Since Appendices D.2 and B.1 both describe 64³ grids, what spatial resampling is actually performed?
  2. Targeted additional comparison: Under the same preprocessing, objective, training configurations, and budget input, how does the proposed model compare with a 3D U-Net or the existing capacity-matched convolutional control equipped with PBAM? Can training-seed variability be reported for the central conditioned/unconditioned and FACA/spatial-Mamba comparisons? These experiments would directly affect the assessment of the claimed methodological advantage.
Reviewer_WKqUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors introduce a photon budget an explicit condition for learned MAMBA-based volumetric reconstruction, allowing one volumetric model to adapt across eight sampling regimes while retaining Monte-Carlo as the forward solver.

Critical strengths
  • The application domain is interesting and broadly applicable (although this is not really discussed in the paper)
  • The architecture itself appears to be interesting and potentially novel, including the spatial-frequency state-space block (although there is no mention of a DFT or FFT in the main text). The use of MAMBA itself could be useful as a non-standard architecture, but even more so since its compared to other architectures.
  • The method-internal ablations strengthen the argument significantly
Critical weaknesses
  • The problem and reconstruction model are not well explained. There is a lot of jargon used that makes it difficult to understand the paper and its machine learning or representation learning contributions. The text is unnecessarily complicated. The notation used in Equation 1 seems strange -- don't you mean to write \hat{x} = A^{-1}(y) ? The use of tilde and hat is unclear.
  • In fact, I believe the framing of the problem and reconstruction prevents adequate ablation and comparison to other methods. Essentially none are discussed or compared against. It is simply untrue there are no reconstruction algorithms (including non Monte-Carlo-based) that apply to volumetric imaging. This must be discussed in the text at minimum. More importantly, alternative reconstruction algorithms will highlight my main issue (next comment).
  • The entire method seems to come down to a photon budget, which I interpret as a noise level that the learned reconstruction algorithm must accommodate. I do not see this as new. This is the biggest gap.
Questions for authors
  • What is a general model for reconstruction? Why Monte Carlo? What other methods could you compare against?
  • What should readers learn from Figure 2 and Figure 3 and Figure 4? Are the differences significant?
  • Figure 5 is entirely unclear and essentially undiscussed besides saying "Figure 5 tests whether these learned represen- tations are functionally used." Please explain.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.