arxiv: 2605.01543 · v1 · submitted 2026-05-02 · 📡 eess.SP · physics.comp-ph

Recognition: unknown

Physics-Guided Deep Learning For High Resolution X-ray Imaging

Aashwin Ananda Mishra, Andrea Zabala, Anna Grassi, Ariel Arnott, Bob Nagler, Caterina Riconda, Dimitri Khaghani, Eric Cunningham, Eric Galtier, Gilliss Dyer, Gourab Chatterjee, Hae Ja Lee, Jonathan Ehni, Marc Welch, Meriame Berboucha, Mickael Grech, Nick Czapla, Nina Boiadjieva, Paul Neumayer, Peregrine McGehee, Philip Hart, Quynh L. Nguyen, Robert Ettelbrick, Shao Xian Lee

Authors on Pith no claims yet

Pith reviewed 2026-05-09 18:02 UTC · model grok-4.3

classification 📡 eess.SP physics.comp-ph

keywords x-ray imagingartifact suppressionu-netdeep learninghigh energy densityinertial fusionstructural similarityepistemic uncertainty

0 comments

The pith

Modeling X-ray artifacts as a separable feature layer lets a U-Net estimate and remove them directly from single-shot data.

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

The paper models structured, non-stationary artifacts in X-ray imaging systems as an independent layer that overlaps with the signal of interest. A U-Net is trained on synthetic injections of these artifacts to infer and subtract them, replacing the residual patterns left by ordinary flat-field normalization. This targets high-energy-density and inertial-fusion experiments that rely on single shots where drift between exposures cannot be averaged away. In controlled tests the network raises mean structural similarity from 0.345 to 0.906 and from 0.0679 to 0.945 while retaining filament profiles more faithfully than Fourier filtering or dynamic flat-field normalization. Deep ensembles supply epistemic uncertainty maps to indicate when the reconstruction may be unreliable on unseen data.

Core claim

By treating artifacts as a separable feature layer and training a U-Net to estimate that layer from experimental radiographs, the method reconstructs transmission maps that preserve filament signals and yield higher structural similarity scores than conventional normalization techniques.

What carries the argument

U-Net architecture trained to estimate the artifact feature layer from raw measurements, using physics-guided synthetic injections for supervision.

If this is right

Transmission maps with higher structural fidelity improve downstream calculations of electron density, velocity, and feature size.
Filament profiles remain intact during artifact removal, reducing systematic bias in plasma diagnostics.
Epistemic uncertainty estimates from deep ensembles flag shots where the reconstruction may be unreliable.
The approach reduces reliance on repeated reference exposures that are often unavailable in single-shot experiments.

Where Pith is reading between the lines

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

If the method holds on real data it could shorten the time needed to extract quantitative results from each expensive shot.
The separable-layer premise may transfer to other single-exposure imaging modalities that suffer structured detector artifacts.
Uncertainty maps could be used to weight or discard individual pixels in downstream physics analysis.

Load-bearing premise

X-ray artifacts can be treated as a separable feature layer independent of the true signal and that a U-Net trained only on synthetic examples will generalize to real experimental images without introducing new biases or signal loss.

What would settle it

Direct application of the trained U-Net to real HED X-ray radiographs that shows either lower SSIM than dynamic flat-field normalization or measurable attenuation of known filament features.

Figures

Figures reproduced from arXiv: 2605.01543 by Aashwin Ananda Mishra, Andrea Zabala, Anna Grassi, Ariel Arnott, Bob Nagler, Caterina Riconda, Dimitri Khaghani, Eric Cunningham, Eric Galtier, Gilliss Dyer, Gourab Chatterjee, Hae Ja Lee, Jonathan Ehni, Marc Welch, Meriame Berboucha, Mickael Grech, Nick Czapla, Nina Boiadjieva, Paul Neumayer, Peregrine McGehee, Philip Hart, Quynh L. Nguyen, Robert Ettelbrick, Shao Xian Lee.

**Figure 1.** Figure 1: Comparison of defect-suppression methods for the X-ray transmission map. (a) Fourier-filtered image, (b) DFFN image, and (c) U-Net reconstructed image. The dashed line indicates the lineout location at y = 43 µm from the sample surface. (d) Corresponding lineouts at y = 43 µm from the raw (unprocessed), Fourier-filtered, DFFN, and U-Net reconstructed images. Filaments 1 to 6 are marked in (d), and their po… view at source ↗

**Figure 2.** Figure 2: (a) Synthetic ground truth 1 (contrast-enhanced filament map with stronger filament contrast). The dashed line marks the lineout location at y = 45.2 µm from the sample surface. (b) Lineout comparison at y = 45.2 µm across the filament region, showing the ground truth profile and the profiles obtained from Fourier Filtering, DFFN, and U-Net reconstruction. (c) Synthetic ground truth 2 (contrast-enhanced fi… view at source ↗

**Figure 3.** Figure 3: (a), (c) Synthetic ground truths with the labeled filaments used for length measurement. (b), (d) Comparison of filament lengths measured from the ground truth, U-Net reconstruction, Fourier filtering and DFFN. We then processed each synthetic image using Fourier filtering, DFFN, and the trained U-Net model. We quantified the filament signal preservation by computing the mean structural similarity index (M… view at source ↗

**Figure 4.** Figure 4: (a) Raw X-ray Transmission map of a shock wave. (b) Reconstructed X-ray transmission map using the filament model. (c) Reconstructed X-ray transmission map using the shock wave model. filtering and 11.3% for DFFN. This result shows U-Net reconstruction suppresses the imaging artifacts and also provides low-error filament length measurement, which is important for subsequent measurements such as electron ve… view at source ↗

**Figure 5.** Figure 5: Pixel-wise entropy map computed from an ensemble of 10 models for (a) a filament image and (b) a shock wave. Brighter color (higher entropy) represents higher uncertainty, highlighting regions where the ensemble shows the largest disagreement (typically near the prominent physical structures). To assess the reliability of the artifact suppression and detect potential OOD features, we implemented a deep ens… view at source ↗

**Figure 6.** Figure 6: (a) Laser-driven shot image with filaments formed on the bottom of the image. Laser hit the Si bulk sample from the bottom. (b) Flat-field image. (c) X-ray transmission map computed using the images from (a) and (b). Dynamic Flat-Field Normalization DFFN was used to correct the dynamics in flat field before normalization7 . Let If lat,m denote the mth flat-field image and ¯If lat the mean flat field. Each … view at source ↗

**Figure 7.** Figure 7: U-Net architecture diagram of the model used in this study. The Convolution layers are shown in yellow, the pooling layers in deep red and the upsampling layers in blue. The input is denoted in green and the final output in red. The Encoder consists of three blocks, each blocks is composed from convolution-convolution-pooling operations. The Decoder mirrors this by using three blocks, composed of upsamplin… view at source ↗

read the original abstract

Imperfections in X-ray imaging systems can limit their performance, especially in High Energy Density (HED) or Inertial Fusion Energy (IFE)-relevant experiments that are typically single shot, by introducing structured, non-stationary features that overlap with the signal of interest. When the X-ray transmission is reconstructed by typical flat-field normalization, even small shot-to-shot drift of structured features imprints residual patterns onto transmission maps, degrading signal visibility and biasing measurements such as electron density, velocity and feature sizes. We investigate this limitation by modeling the artifacts as a separable feature layer and training a U-Net architecture to estimate and infer them directly from the experimental data. We compare our method against Fourier filtering and more advanced procedures like Dynamic Flat-Field Normalization (DFFN) to evaluate artifact suppression capability and signal preservation in the reconstructed transmission maps. In multiple synthetic injection tests, our Physics-Guided Deep Learning approach is able to obtain an improvement in mean Structural Similarity Index (SSIM) from 0.345 to 0.906 and from 0.0679 to 0.945, while better preserving filament profiles and reducing degradation of the filament signal during artifact suppression. Additionally, we utilize deep ensembles to obtain predictive epistemic uncertainty estimates for the U-Net based reconstruction, to ensure Out Of Distribution (OOD) robustness for this procedure.

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.

Desk Editor's Note private letter to a colleague

The paper shows clear SSIM gains on synthetic X-ray artifact removal with a separable-layer U-Net plus uncertainty estimates, but real experimental data is missing so generalization stays untested.

read the letter

The main point is that the authors train a U-Net to pull out structured artifacts from single-shot X-ray radiographs by treating them as an additive separable layer. On their synthetic injection tests the method lifts mean SSIM from 0.345 to 0.906 and from 0.0679 to 0.945, beats Fourier filtering and dynamic flat-field normalization, and keeps filament profiles sharper than the baselines. They also run deep ensembles to give epistemic uncertainty maps that flag out-of-distribution cases. That combination is a reasonable domain-specific extension for high-energy-density and inertial-fusion imaging where shot-to-shot drifts ruin flat-field corrections. The synthetic benchmarks are quantitative and the comparisons are straightforward, so the work gives a practical starting point for anyone already dealing with these artifacts. The soft spot is exactly what the stress-test note flags: everything rests on synthetic data generated under the separability assumption. If real artifacts correlate with the transmission signal or contain frequencies outside the training set, the network can suppress real features while still scoring high SSIM on held-out synthetics. The abstract gives no details on how the synthetic artifacts were created or matched to experiment, and no real radiographs with independent ground truth appear to have been tested. That leaves the central claim about signal preservation and bias reduction only moderately supported. The uncertainty estimates help but do not close the gap. This paper is for experimental groups in plasma physics or fusion diagnostics who already have X-ray imaging pipelines and want a drop-in tool to try on their own data. A reader in that niche would get immediate value from the architecture choice and the reported numbers. It is coherent enough and grounded enough in a real problem to deserve peer review; the referees can press on the synthetic-to-real step and ask for at least one real-data case or a clearer description of the injection process. I would send it out rather than desk-reject.

Referee Report

3 major / 0 minor

Summary. The paper proposes modeling X-ray imaging artifacts in HED/IFE experiments as a separable additive feature layer and training a U-Net to estimate and subtract them directly from experimental radiographs. The method is compared to Fourier filtering and Dynamic Flat-Field Normalization (DFFN) on synthetic artifact-injection tests, reporting mean SSIM improvements from 0.345 to 0.906 and from 0.0679 to 0.945 while claiming better filament-profile preservation; deep ensembles are used to provide epistemic uncertainty estimates for OOD robustness.

Significance. If the separability assumption and synthetic-to-real generalization hold, the approach could meaningfully improve single-shot transmission measurements by reducing residual artifact bias without excessive signal loss, with the deep-ensemble uncertainty providing a practical safeguard. The independent SSIM evaluation on held-out synthetics and the explicit comparison to DFFN are strengths that support the quantitative claims within the synthetic regime.

major comments (3)

[Abstract] Abstract: The headline SSIM gains (0.345→0.906 and 0.0679→0.945) and the claim of “better preserving filament profiles and reducing degradation of the filament signal” are demonstrated exclusively on synthetic injections that presuppose artifacts form an additive, separable layer independent of the transmission signal. No real experimental radiographs are shown, leaving the central generalization claim untested against the structured, shot-to-shot, non-stationary artifacts actually present in HED/IFE data.
[Abstract] Abstract: The manuscript provides no description of how the synthetic artifacts were generated (e.g., spatial-frequency content, amplitude distribution, or correlation with the underlying filament signal). Without these details it is impossible to judge whether the training distribution adequately covers the real artifact statistics that the U-Net is intended to handle.
[Abstract] Abstract: No error bars, statistical significance tests, or cross-validation results accompany the reported mean SSIM values, making it difficult to assess the robustness of the claimed improvement over DFFN and Fourier filtering.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed comments on our manuscript. We have reviewed each major comment carefully and provide point-by-point responses below, indicating where revisions will be made to strengthen the paper.

read point-by-point responses

Referee: [Abstract] Abstract: The headline SSIM gains (0.345→0.906 and 0.0679→0.945) and the claim of “better preserving filament profiles and reducing degradation of the filament signal” are demonstrated exclusively on synthetic injections that presuppose artifacts form an additive, separable layer independent of the transmission signal. No real experimental radiographs are shown, leaving the central generalization claim untested against the structured, shot-to-shot, non-stationary artifacts actually present in HED/IFE data.

Authors: We agree that all quantitative results, including the reported SSIM improvements and filament profile preservation, are obtained exclusively from controlled synthetic artifact-injection experiments that assume an additive, separable artifact layer. This modeling choice reflects the physical characteristics of X-ray system imperfections in single-shot HED/IFE setups, where shot-to-shot drifts produce structured features largely independent of the underlying transmission. Because real single-shot experiments lack pixel-wise ground truth, synthetic tests enable rigorous quantitative evaluation that would otherwise be impossible. We will revise the abstract to explicitly note that the SSIM gains and profile comparisons are demonstrated on synthetic data, and we will expand the discussion section to elaborate on the separability assumption, its physical basis, and the challenges of direct validation on real radiographs. This will clarify the scope of the current claims while highlighting the method's potential for real-world application. revision: partial
Referee: [Abstract] Abstract: The manuscript provides no description of how the synthetic artifacts were generated (e.g., spatial-frequency content, amplitude distribution, or correlation with the underlying filament signal). Without these details it is impossible to judge whether the training distribution adequately covers the real artifact statistics that the U-Net is intended to handle.

Authors: We thank the referee for highlighting this omission. Although the full manuscript describes the artifact synthesis procedure in the Methods section (including generation of non-stationary features via spatially varying low-frequency modulations with controlled amplitudes and minimal correlation to the filament signal), these details were not summarized in the abstract. In the revised version we will add a concise description of the synthetic artifact generation process to the abstract, specifying the spatial-frequency content, amplitude ranges, and correlation properties used. This will allow readers to assess how well the training distribution matches expected real-world artifact statistics. revision: yes
Referee: [Abstract] Abstract: No error bars, statistical significance tests, or cross-validation results accompany the reported mean SSIM values, making it difficult to assess the robustness of the claimed improvement over DFFN and Fourier filtering.

Authors: We acknowledge that the abstract reports only mean SSIM values without accompanying variability measures or statistical analysis. The means were computed across multiple independent synthetic injection trials, and the underlying experiments incorporated cross-validation folds, yet these details were not included in the abstract. In the revised manuscript we will add error bars (standard deviation across trials) to the SSIM results, report the outcomes of cross-validation, and include statistical significance tests (e.g., paired Wilcoxon signed-rank tests) comparing our method to Fourier filtering and DFFN. These additions will be placed both in the abstract and in the results section to demonstrate the robustness of the observed improvements. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation or claims

full rationale

The paper presents an empirical ML method: artifacts are modeled as a separable layer, a U-Net is trained on synthetic injections, and performance is measured with independent SSIM on held-out synthetics. No equations, derivations, or self-citations are shown that reduce the reported SSIM gains (0.345→0.906, 0.0679→0.945) or filament preservation claims to a fitted parameter, self-definition, or tautological input by construction. The evaluation metrics are external benchmarks, not internal to the training or modeling assumptions.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the modeling assumption that artifacts form a separable layer and on standard supervised learning assumptions for the U-Net; no new physical entities are introduced and no free parameters beyond typical network hyperparameters are described.

axioms (1)

domain assumption X-ray imaging artifacts can be modeled as a separable feature layer independent of the underlying transmission signal.
This separability is the explicit modeling choice used to train the U-Net directly from experimental data.

pith-pipeline@v0.9.0 · 5640 in / 1330 out tokens · 27461 ms · 2026-05-09T18:02:44.303812+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

28 extracted references · 19 canonical work pages

[1]

Reports15, 7588, DOI: 10.1038/s41598-025-91989-8 (2025)

Galtier, E.et al.X-ray microscopy and talbot imaging with the matter in extreme conditions x-ray imager at lcls.Sci. Reports15, 7588, DOI: 10.1038/s41598-025-91989-8 (2025)

work page doi:10.1038/s41598-025-91989-8 2025
[2]

K., Landen, O

Bradley, D. K., Landen, O. L., Bullock, A. B., Glendinning, S. G. & Turner, R. E. Efficient, 1–100-kev x-ray radiography with high spatial and temporal resolution.Opt. Lett.27, 134–136, DOI: 10.1364/OL.27.000134 (2002)

work page doi:10.1364/ol.27.000134 2002
[3]

Plasmas15, 072705, DOI: 10.1063/1.2957918 (2008)

Park, H.-S.et al.High-resolution 17–75 kev backlighters for high energy density experiments.Phys. Plasmas15, 072705, DOI: 10.1063/1.2957918 (2008). 11/13 4.Momose, A.et al.Demonstration of x-ray talbot interferometry.Jpn. J. Appl. Phys.42, L866–L868, DOI: 10.1143/JJAP. 42.L866 (2003)

work page doi:10.1063/1.2957918 2008
[4]

Clarke, and Federico Capasso

Buakor, K.et al.Shot-to-shot flat-field correction at x-ray free-electron lasers.Opt. Express30, 10633, DOI: 10.1364/OE. 451914 (2022)

work page doi:10.1364/oe 2022
[5]

Synchrotron Radiat.28, 52–63, DOI: 10.1107/S160057752001557X (2021)

Hagemann, J.et al.Single-pulse phase-contrast imaging at free-electron lasers in the hard x-ray regime.J. Synchrotron Radiat.28, 52–63, DOI: 10.1107/S160057752001557X (2021)

work page doi:10.1107/s160057752001557x 2021
[6]

Express23, 27975–27989, DOI: 10.1364/OE.23.027975 (2015)

Van Nieuwenhove, V .et al.Dynamic intensity normalization using eigen flat fields in x-ray imaging.Opt. Express23, 27975–27989, DOI: 10.1364/OE.23.027975 (2015)

work page doi:10.1364/oe.23.027975 2015
[7]

P.et al.Optical imaging of laser-driven fast electron weibel-like filamentation in overcritical density plasma

Dover, N. P.et al.Optical imaging of laser-driven fast electron weibel-like filamentation in overcritical density plasma. Phys. Rev. Lett.134, 025102, DOI: 10.1103/PhysRevLett.134.025102 (2025)

work page doi:10.1103/physrevlett.134.025102 2025
[8]

Commun.15, 7528, DOI: 10.1038/s41467-024-51084-4 (2024)

Sawada, H.et al.Spatiotemporal dynamics of fast electron heating in solid-density matter via xfel.Nat. Commun.15, 7528, DOI: 10.1038/s41467-024-51084-4 (2024)

work page doi:10.1038/s41467-024-51084-4 2024
[9]

S.et al.Observations of the filamentation of high-intensity laser-produced electron beams.Phys

Wei, M. S.et al.Observations of the filamentation of high-intensity laser-produced electron beams.Phys. Rev. E70, 056412, DOI: 10.1103/PhysRevE.70.056412 (2004)

work page doi:10.1103/physreve.70.056412 2004
[10]

Romagnani, L.et al.Dynamics of the electromagnetic fields induced by fast electron propagation in near-solid-density media.Phys. Rev. Lett.122, 025001, DOI: 10.1103/PhysRevLett.122.025001 (2019)

work page doi:10.1103/physrevlett.122.025001 2019
[11]

Kodama, R.et al.Fast heating of ultrahigh-density plasma as a step towards laser fusion ignition.Nature412, 798–802, DOI: 10.1038/35090525 (2001)

work page doi:10.1038/35090525 2001
[12]

Robinson, A. P. L.et al.Theory of fast electron transport for fast ignition.Nucl. Fusion54, 054003, DOI: 10.1088/ 0029-5515/54/5/054003 (2014)

2014
[13]

Plasmas1, 1626–1634, DOI: 10.1063/1.870664 (1994)

Tabak, M.et al.Ignition and high gain with ultrapowerful lasers.Phys. Plasmas1, 1626–1634, DOI: 10.1063/1.870664 (1994)

work page doi:10.1063/1.870664 1994
[14]

Commun.17, 467, DOI: 10.1038/s41467-025-67160-2 (2026)

Schoenwaelder, C.et al.Time-resolved x-ray imaging of the current filamentation instability in solid-density plasmas.Nat. Commun.17, 467, DOI: 10.1038/s41467-025-67160-2 (2026)

work page doi:10.1038/s41467-025-67160-2 2026
[15]

V ., Schafer, R

Oppenheim, A. V ., Schafer, R. W. & Stockham, T. G. Nonlinear filtering of multiplied and convolved signals.IEEE Transactions on Audio Electroacoustics16, 437–466 (1968)

1968
[16]

& Brox, T

Ronneberger, O., Fischer, P. & Brox, T. U-net: Convolutional networks for biomedical image segmentation. InMedical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, 234–241 (Springer, 2015)

2015
[17]

Siddique, N., Paheding, S., Elkin, C. P. & Devabhaktuni, V . U-net and its variants for medical image segmentation: A review of theory and applications.IEEE Access9, 82031, DOI: 10.1109/ACCESS.2021.3086020 (2021)

work page doi:10.1109/access.2021.3086020 2021
[18]

Intell.46, 10076, DOI: 10.1109/TPAMI.2024.3435571 (2024)

Azad, R.et al.Medical image segmentation review: The success of u-net.IEEE Transactions on Pattern Analysis Mach. Intell.46, 10076, DOI: 10.1109/TPAMI.2024.3435571 (2024)

work page doi:10.1109/tpami.2024.3435571 2024
[19]

Fornek, T. E. Advanced photon source upgrade project final design report. Tech. Rep., Argonne National Laboratory (ANL), Argonne, IL (United States) (2019)

2019
[20]

W.et al.LCLS-II High Energy (LCLS-II-HE): A transformative x-ray laser for science

Schoenlein, R. W.et al.LCLS-II High Energy (LCLS-II-HE): A transformative x-ray laser for science. Tech. Rep., SLAC National Accelerator Laboratory (SLAC), Menlo Park, CA (United States) (2016)

2016
[21]

Synchrotron Radiat

Mishra, A.et al.A start to end machine learning approach to maximize scientific throughput from the LCLS-II-HE. Synchrotron Radiat. News38, 10–17 (2025)

2025
[22]

& Ratner, D

Edelen, A., Neveu, N., Mayes, C., Emma, C. & Ratner, D. Machine learning models for optimization and control of x-ray free electron lasers. InNeurIPS Machine Learning for the Physical Sciences Workshop(2019)

2019
[23]

& Stampanoni, M

Münch, B., Trtik, P., Marone, F. & Stampanoni, M. Stripe and ring artifact removal with combined wavelet - fourier filtering.Opt. Express17, 8567–8591, DOI: 10.1364/OE.17.008567 (2009)

work page doi:10.1364/oe.17.008567 2009
[24]

Bovik, H.R

Wang, Z., Bovik, A. C., Sheikh, H. R. & Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Process.13, 600–612, DOI: 10.1109/TIP.2003.819861 (2004)

work page doi:10.1109/tip.2003.819861 2003
[25]

& Blundell, C

Lakshminarayanan, B., Pritzel, A. & Blundell, C. Simple and scalable predictive uncertainty estimation using deep ensembles. InAdvances in Neural Information Processing Systems, vol. 30 (2017). 12/13

2017
[26]

Sepúlveda-Fontaine, S. A. & Amigó, J. M. Applications of entropy in data analysis and machine learning: A review. Entropy26, 1126, DOI: 10.3390/e26121126 (2024)

work page doi:10.3390/e26121126 2024
[27]

L., Gullikson, E

Henke, B. L., Gullikson, E. M. & Davis, J. C. X-ray interactions: photoabsorption, scattering, transmission, and reflection atE=50−30,000 eV ,Z=1−92.At. Data Nucl. Data Tables54, 181–342 (1993)

1993
[28]

2021 , url =

Ghiasi, G.et al.Simple copy-paste is a strong data augmentation method for instance segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), DOI: 10.1109/CVPR46437.2021.00294 (2021). 30.Hendrycks, D. & Gimpel, K. Gaussian error linear units (gelus).arXiv preprint arXiv:1606.08415(2016). 31.Kingma, D. P. & B...

work page doi:10.1109/cvpr46437.2021.00294 2021