Multispectral CT Denoising via Simulation-Trained Deep Learning: Experimental Results at the ESRF BM18

Andreas Balles; Astrid H\"olzing; Charls Clark; Guangpu Yang; Jajnabalkya Guhathakurta; Paul Tafforeau; Peter G\"anz; Simon Zabler; Steffen Kie{\ss}; Sven Simon

REVIEW 3 major objections 5 minor 2 references

A neural network trained exclusively on simulated projections denoises real multispectral CT scans at the BM18 beamline across 22.9–400 keV.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 20:28 UTC pith:U6SCAR2Q

load-bearing objection Transparent conference-paper merge with a useful architecture, but an unflagged simulation/experiment geometry mismatch weakens the sim-to-real claim until sensitivity is shown. the 3 major comments →

arxiv 2509.08528 v1 pith:U6SCAR2Q submitted 2025-09-10 eess.IV

Multispectral CT Denoising via Simulation-Trained Deep Learning: Experimental Results at the ESRF BM18

Peter G\"anz , Steffen Kie{\ss} , Guangpu Yang , Jajnabalkya Guhathakurta , Tanja Pienkny , Charls Clark , Paul Tafforeau , Andreas Balles

show 3 more authors

Astrid H\"olzing Simon Zabler Sven Simon

This is my paper

classification eess.IV

keywords computed tomographymultispectral imagingdeep learningdenoisingsimulation-to-real transfersynchrotronenergy-resolved imagingconvolutional neural networks

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Multispectral CT splits the X-ray beam into narrow energy bins, so each bin receives far fewer photons than an energy-integrating detector would, making scans either very noisy or impractically long. This paper claims that a neural network trained exclusively on simulations of the BM18 beamline can denoise real experimental projections across the full 22.9–400 keV range, outperforming a single-frame CNN baseline and classical denoisers while preserving structural detail. The method exploits two redundancies in the data—similarity between adjacent energy bins and similarity between neighboring projection angles—through two sub-networks combined by stacked generalization. If the claim holds, multispectral CT can be run at practical exposure times and its low-flux energy bins become usable for material characterization.

Core claim

The central claim is that a neural network trained only on simulated BM18 projections denoises real experimental multispectral CT data. The network pairs a spectral-spatial subnetwork (denoising each energy bin with 64 adjacent bins) with an angular-spatial subnetwork (denoising via neighboring projections of the same bin), fusing both by stacked generalization. On two custom phantoms, it gave lower normalized RMSE than raw data and a DnCNN baseline across sampled energy bins, and higher SNR in homogeneous regions over the first 100 bins, with no added reconstruction artifacts down to roughly 0.076% of detector full scale.

What carries the argument

Two-path neural network plus a physics-based simulator. The spectral path, adapted from a hyperspectral-image denoiser, extracts spatial and spectral features across 64 adjacent energy bins with attention-guided convolutions. The angular path, adapted from a video-denoising patch-craft method, builds patch frames from neighboring projections of the same energy bin and applies an angular-post filter to keep projection-to-projection continuity. Stacked generalization—training a third convolution layer on the two frozen subnetworks' feature maps—fuses the paths. The simulator generates training data by modeling Beer-Lambert absorption, silicon prism refraction, scintillator conversion, two Pois

Load-bearing premise

The load-bearing premise is that the simulated noise and signal model faithfully represents the real BM18 detector chain; if the simulator misses a noise source or miscalibrates the energy-to-detector-line mapping, the demonstrated transfer from simulation to experiment on two phantoms could fail on other samples or energy bins.

What would settle it

Scan a third phantom with known material inserts, denoise its projections with the trained network, and compare the lowest-energy-bin reconstructions against a heavily averaged reference; the transfer claim fails if NRMSE does not improve over raw data or if new streak artifacts appear where the simulator predicted none. Alternatively, record blocked-beam flat frames at BM18 and compare measured dark-signal and readout-noise statistics with the simulated Gaussian model.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

Multispectral scans at BM18 can use short exposure times and recover image quality by denoising, avoiding scan times that would otherwise stretch from hours to days.
Denoising in projection space produces cleaned data that can be fed directly into standard filtered backprojection without changing the reconstruction pipeline.
Low-flux energy bins—previously too noisy to use—become analyzable, extending the usable spectral range down to about 22.9 keV.
The spectral and angular sub-networks together outperform either alone and beat the single-frame DnCNN baseline, suggesting both redundancies matter.

Where Pith is reading between the lines

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

The paper validates on two phantoms; I would not extrapolate the sim-to-real transfer to arbitrary samples until a third, unseen object is tested.
The same two-path design could likely be retrained for other spectral CT geometries, but the prism-dispersion and detector-noise models would need to be re-derived for each setup.
The reported low-signal limit (roughly 50 expected gray values) suggests that emphasizing low-flux bins in training, or modeling the noise floor more explicitly, could extend the usable spectral range beyond the first 100 detector lines.
Because the experimental comparison uses a 50-times-averaged reconstruction as reference, the NRMSE numbers measure agreement with an average; a noise-removal test on raw projection residuals would more directly confirm that fine structure is preserved.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Transparent conference-paper merge with a useful architecture, but an unflagged simulation/experiment geometry mismatch weakens the sim-to-real claim until sensitivity is shown.

read the letter

The short version: this is a transparently assembled merge of two conference papers, and the method is a plausible engineering contribution, but the paper contains an internal inconsistency that weakens the sim-to-real claim until it is addressed. The simulation (Section III.C) uses a 60° prism tip angle and 50 prisms, while the experimental setup (Section I.A) is described with a 70.52° tip angle and 100 prisms. The distance also disagrees: 10 m in the introduction versus 28 m in the setup section. Because the energy-to-detector-row mapping in Eq. (6) is proportional to the prism refraction angle, these discrepancies shift every simulated energy boundary. The paper admits 'notable discrepancies' between simulation and experiment but never checks sensitivity to these specific parameters. This is not a minor detail; it's a load-bearing inconsistency.

What the paper does well: the architecture is sensible — an angular-spatial branch inspired by PaCNet and a spectral branch inspired by AODN, combined with stacked generalization. The simulation pipeline is detailed, including double Poisson noise, PRNU, dark signal and readout noise, which gives the training data a degree of realism. Validating on two custom phantoms with DnCNN as a baseline is a reasonable first step, and the authors are honest that the manuscript reuses their own conference papers.

The soft spots are real but mostly addressable. The synthetic test set is generated by the same simulator used for training, so the excellent PSNR/SSIM values partly reflect the network inverting its own training distribution. The real-data quantitative evaluation is thin: two phantoms, four energy bins, and no error bars. No code or trained model is provided, so the results aren't reproducible. And the geometry mismatch above is the main barrier: before the sim-to-real claim can stand, the authors need to either retrain with corrected geometry or show that the network's output is insensitive to a factor-of-two change in prism count and a 17% change in tip angle.

Who this is for: researchers working on spectral CT denoising or synchrotron imaging will find the architecture and noise model useful, even though the novelty is modest. The paper deserves a serious referee, but not in its current form. The right call is to send it to review and ask for major revision: fix the geometry inconsistency or add a sensitivity analysis, provide error bars on the real-data metrics, and ideally release the trained model.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes a deep-learning denoiser for multispectral CT projections acquired at the ESRF BM18 beamline, where a prism array disperses the polychromatic beam onto detector rows corresponding to energy bins. The architecture combines a spectral-spatial subnetwork (based on AODN) and an angular-spatial subnetwork (based on PaCNet) via stacked generalization and attention modules. Training is done exclusively on simulated projections produced by a physics-based model incorporating Beer-Lambert absorption, prism refraction, scintillator conversion, double Poisson noise, PRNU, dark signal, and readout noise. The authors validate on a synthetic test set from the same simulator and on real BM18 scans of two phantoms, reporting improved PSNR/SSIM/LPIPS over NLM/TV/DnCNN on synthetic data and lower NRMSE relative to a 50x averaged reference plus improved ROI SNR on real data.

Significance. The application is timely: multispectral CT at synchrotron facilities is photon-starved, and a denoiser that transfers from simulation to real data would enable faster acquisitions. The paper's strengths are the explicit physics-based noise model, the use of external phantom datasets, and a head-to-head comparison with DnCNN. However, the central sim-to-real claim is only as strong as the simulator's fidelity, and the geometry inconsistency identified below is currently unaddressed. If properly resolved, the method is a useful engineering contribution, although the architecture itself is largely assembled from existing building blocks and the synthetic results are in-distribution.

major comments (3)

[Section III.C vs. Section I.A, Eqs. (2)-(6)] The simulator in Section III.C uses n_prism=50 and alpha=60 deg in Eqs. (2)-(6), while Section I.A describes the actual BM18 prism array as 100 prisms with a 70.52 deg tip angle. Since Eq. (6) maps the total refraction angle to detector row, doubling n_prism and increasing tan(alpha/2) by ~22% materially changes every simulated energy-bin boundary. The paper does not state that the K-edge calibration of Section I.A was folded into the training simulator, nor does it provide a sensitivity analysis. This directly affects the spectral subnetwork, which is trained on correlations among k=64 adjacent bins; a mismatched energy axis could cause the network to learn incorrect spectral structure. Please clarify which geometry was used in training and either retrain with the correct parameters or demonstrate that the real-data results are robust to this discrepancy.
[Section IV.A, Table 3] The synthetic test set is generated by the same simulator used for training (different objects, same noise model and geometry). The high PSNR/SSIM/LPIPS values therefore partly measure the network's ability to invert the training simulator and should not be used to support sim-to-real generalization. Please relabel Table 3 as an in-distribution benchmark and add an out-of-distribution test (e.g., different noise levels, different prism parameters, or real-data-only evaluation) to substantiate the generalization claim.
[Section IV.B, Table 4 and Fig. 12] The real-data evaluation uses a 50x averaged reconstruction as reference, which is itself noisy and not ground truth, and the 'mean normalized NRMSE' is not defined. Only four energy bins (0, 5, 50, 99) are reported numerically, and the SNR analysis covers only the low-Z phantom. To support the claim of performance over the full 22.9-400 keV range, please define the metric precisely, report bin-wise curves or error bars for both phantoms, and discuss how the 50x averaging affects the comparison.

minor comments (5)

[Section III.A] Duplicate phrase: 'First, spatial features of First, spatial features of one band and projection angle are extracted.'
[Section III.A] 'The denoising block from [6] was modified' likely refers to reference [26] (AODN), not [6]. Please correct the citation.
[Throughout] Typos and inconsistent notation: 'LeRU' should be 'Leaky ReLU'; Table 4 caption 'Desnoing' should be 'Denoising'; 'unfilted', 'horizontaly', 'refrac' should be corrected; and the dark current/readout noise notation 'sigma=darkcurrent!' and 'sigma=0.8!' needs units and explanation.
[Section III.D] The normalization of detector lines uses the maximal possible gray value from simulation. Please state explicitly how the same normalization is applied to real experimental data, since a mismatch here could affect the network's input distribution.
[General] The paper's self-description as literal quotes from two conference papers is unusual for a journal submission. Please rewrite into a single coherent document with consistent nomenclature and avoid duplicated or truncated sentences.

Circularity Check

0 steps flagged

No circularity: the real-data evaluation is an external benchmark, and the self-citations are source attributions, not load-bearing justifications.

full rationale

The central claim is that a network trained exclusively on simulated BM18-like data denoises real BM18 multispectral projections. This is not circular by construction: the network is trained on synthetic inputs generated from a physics-based noise model, and its generalization is evaluated against experimentally acquired phantom scans using a 50x-averaged reconstruction as an independent reference (Section IV.B, Table 4, Figures 11-12). The synthetic test-set results (Section IV.A, Table 3) use held-out objects from the same simulator, so they test in-distribution denoising rather than sim-to-real transfer; the real-data results are the actual external validation, and no fitted parameter from the real data is renamed as a prediction. The paper's self-citations [1], [2] merely identify the source of the quoted simulation and results sections; they are not used to justify the generalization claim. The architecture is adapted from external works (PaCNet [23], AODN [26]), and no uniqueness theorem or prior self-derived constraint is invoked. Limitations are present but are correctness risks, not circularity: the manuscript admits 'Although there are notable discrepancies between the simulation setup and the actual experimental conditions', and there is an internal inconsistency in the refraction parameters (Section III.C uses n_prism=50 and alpha=60 degrees in Eqs. (2)-(6), while Section I.A states 'This array, composed of 100 silicon prisms integrated onto a single wafer' with a '70.52 degrees' tip angle). A shift in these parameters would change the simulated energy-to-detector-line mapping, but this is a sensitivity/calibration concern about the simulator, not a case of the network's output being equivalent to its input by definition. Therefore no circular step can be exhibited, and the circularity score is 0.

Axiom & Free-Parameter Ledger

12 free parameters · 5 axioms · 0 invented entities

The central claim rests on a simulation-based training distribution and a small real-data evaluation. The simulation is built from standard physics (Beer-Lambert, refraction, Poisson statistics) plus several hand-chosen detector parameters. There are no invented physical entities. The main unproved premise is that the simulator captures all relevant noise sources and that the 50x averaged reference is a valid surrogate for ground truth.

free parameters (12)

VideoNet learning rate = 1.5e-4 (reported)
Tuned via Bayesian optimization (Table 2); affects training convergence and final denoising quality.
HSINet learning rate = 2.33e-3 (reported)
Tuned via Bayesian optimization; affects training.
Stacked generalization learning rate = factor 10 smaller than HSINet LR
Chosen by rule of thumb; affects the combination network.
Number of nearest neighbors = 5
Hyperparameter for angular patch matching in the PaCNet-style branch.
Number of adjacent energy bands k = 64
Spectral context window; hyperparameter from Bayesian search.
Number of denoising blocks = 6
Architecture depth.
Number of octave convolution blocks = 6
Architecture depth.
Percentage of low frequencies alpha = 0.1
Octave convolution split ratio; chosen by hyperparameter search.
Optical loss factor = 4000
Assumed attenuation between scintillator and sensor; not measured in the paper.
PRNU standard deviation = 0.001
Gaussian model for photo-response non-uniformity; assumed value.
Readout noise sigma = 0.8 electrons
From PCO Edge 4.2 datasheet; used in the simulation.
Dark signal non-uniformity = sigma = darkcurrent^2 (not quantified)
Noise model parameter not fully specified.

axioms (5)

standard math Beer-Lambert attenuation and Snell's-law refraction describe photon transmission and spectral dispersion through the prism array.
Used in Eqs. (1)-(6) to generate simulated training projections.
standard math Photon detection follows Poisson statistics with the stated expected intensities.
Used to add shot noise in Section III.C.
domain assumption The 50x averaged experimental reconstruction is a valid reference for computing NRMSE.
Used as pseudo-ground truth in Section IV.B; no ground truth exists for real scans.
domain assumption The simulated training distribution (300 angles, 100 energy bins, Al/SiO2 volumes) is representative enough for the real high-Z/low-Z phantoms.
Required for sim-to-real transfer claims in Sections III.C and IV.B.
domain assumption All significant noise sources of the BM18 detector are captured by the simulation model.
The paper acknowledges discrepancies but does not quantify missing noise sources; this is the load-bearing premise for training purely on simulated data.

pith-pipeline@v1.3.0-alltime-deepseek · 15476 in / 14264 out tokens · 141551 ms · 2026-08-04T20:28:19.684754+00:00 · methodology

0 comments

read the original abstract

Multispectral computed tomography (CT) enables advanced material characterization by acquiring energy-resolved projection data. However, since the incoming X-ray flux is be distributed across multiple narrow energy bins, the photon count per bin is greatly reduced compared to standard energy-integrated imaging. This inevitably introduces substantial noise, which can either prolong acquisition times and make scan durations infeasible or degrade image quality with strong noise artifacts. To address this challenge, we present a dedicated neural network-based denoising approach tailored for multispectral CT projections acquired at the BM18 beamline of the ESRF. The method exploits redundancies across angular, spatial, and spectral domains through specialized sub-networks combined via stacked generalization and an attention mechanism. Non-local similarities in the angular-spatial domain are leveraged alongside correlations between adjacent energy bands in the spectral domain, enabling robust noise suppression while preserving fine structural details. Training was performed exclusively on simulated data replicating the physical and noise characteristics of the BM18 setup, with validation conducted on CT scans of custom-designed phantoms containing both high-Z and low-Z materials. The denoised projections and reconstructions demonstrate substantial improvements in image quality compared to classical denoising methods and baseline CNN models. Quantitative evaluations confirm that the proposed method achieves superior performance across a broad spectral range, generalizing effectively to real-world experimental data while significantly reducing noise without compromising structural fidelity.

discussion (0)

Reference graph

Works this paper leans on

2 extracted references

[19]

Plug-and-Play Image Restoration With Deep Denoiser Prior,

K. Zhang, Y. Li, W. Zuo, L. Zhang, L. Van Gool, and R. Timofte, “Plug-and-Play Image Restoration With Deep Denoiser Prior,” IEEE Trans Pattern Anal Mach Intell, vol. 44, no. 10, pp. 6360–6376, Oct. 2022, doi: 10.1109/TPAMI.2021.3088914. [20] C. Rota, M. Buzzelli, S. Bianco, and R. Schettini, “Video restoration based on deep learning: a comprehensive surve...

arXiv 2022
[40]

FastDVDNet: Towards real-time deep video denoising without flow estimation,

M. Tassano, J. Delon, and T. Veit, “FastDVDNet: Towards real-time deep video denoising without flow estimation,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2020, pp. 1351–1360. doi: 10.1109/CVPR42600.2020.00143. [41] Q. Zhou and A. Jacobson, “Thingi10K: A Dataset of 10,000 3D-Printing Models,” May 20...

arXiv 2020

[1] [19]

Plug-and-Play Image Restoration With Deep Denoiser Prior,

K. Zhang, Y. Li, W. Zuo, L. Zhang, L. Van Gool, and R. Timofte, “Plug-and-Play Image Restoration With Deep Denoiser Prior,” IEEE Trans Pattern Anal Mach Intell, vol. 44, no. 10, pp. 6360–6376, Oct. 2022, doi: 10.1109/TPAMI.2021.3088914. [20] C. Rota, M. Buzzelli, S. Bianco, and R. Schettini, “Video restoration based on deep learning: a comprehensive surve...

arXiv 2022

[2] [40]

FastDVDNet: Towards real-time deep video denoising without flow estimation,

M. Tassano, J. Delon, and T. Veit, “FastDVDNet: Towards real-time deep video denoising without flow estimation,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2020, pp. 1351–1360. doi: 10.1109/CVPR42600.2020.00143. [41] Q. Zhou and A. Jacobson, “Thingi10K: A Dataset of 10,000 3D-Printing Models,” May 20...

arXiv 2020