REVIEW 4 major objections 8 minor 15 references
A Realistic Collimated X-Ray Image Simulation Pipeline
T0 review · 4 major / 8 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read Synthetic collimator shadows produced by a physics-based pipeline can replace real collimated X-ray images when training a detector, and can even improve its performance on real data.
desk verdict A plausible physics-based collimator simulator with real promise, but the headline generalization claim rests on a flawed test split and overlapping Dice intervals; the pipeline itself is worth refereeing. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The carrying object is the randomized collimator simulation pipeline. It first samples a rectangle's centroid, width, and height from truncated normal distributions, then applies randomized rotation and distortion to the binary mask. The mask is turned into a damping map $M_d$ and convolved with a Gaussian kernel $G_b$ to reproduce the blur of a finite focal spot, giving the collimation operation $L(I_{\mathrm{input}}) = (M_d * G_b) \cdot I_{\mathrm{input}}$. Scatter is modeled with a scatter potential $S_p(I|I_0) = c\,(I/I_0)^{\alpha}\,\ln(I_0/I)^{\beta}$ convolved with a Gaussian kernel $G_s$ to obtain an estimated scatter map; the pipeline subtracts the scatter estimated from the input and re-adds a simulated scatter map after collimation. Finally, Poisson noise is approximated by adding zero-mean normal noise with standard deviation $\sigma_x = \sqrt{\lambda(1-\alpha)}$ in damped regions, so that the collimated areas show the reduced signal-to-noise ratio of real photon-limited measurements.
What would settle it
Acquire phantom images with and without a real collimator in the same geometry, run the pipeline over a grid of scatter hyperparameters ($\alpha$, $\beta$, $c$), and measure whether the simulated intensity profile across the collimator edge can be made to match the real profile within repeatability; if no parameter set achieves that, the simulated shadows are not interchangeable with real ones. A complementary falsifier is to train SimNet on synthetic collimators and evaluate it on an independent real collimated dataset from a different X-ray system, where a clear drop below RealNet's Dice would contradict the generalization claim.
Extended reading notes
Core claim
The central claim is that a simulation composed of randomized geometry, Gaussian edge blur, convolution-based scatter estimation, and Poisson noise produces collimator shadows close enough to real ones to serve as training data for a detector. The paper demonstrates this on an anthropomorphic thorax phantom by comparing a simulated collimator on an open-field image with a real collimated acquisition of the same phantom, reporting near-identical statistics in non-collimated patches (nMSE $0.0001$, SSIM $0.9998$, PSNR $33.23\,\mathrm{dB}$) and high SSIM in collimated patches. It then trains two DeepLabV3 networks, SimNet on pipeline-generated collimators and RealNet on real collimated images. On a general test set of real images, SimNet's Dice is $0.9718 \pm 0.027$ versus RealNet's $0.9641 \pm 0.048$; SimNet also leads on line-artifact images ($0.9778$ vs. $0.9652$), while RealNet leads on implant images ($0.9780$ vs. $0.9494$). The paper reads this overall pattern as evidence that the simulated data are realistic enough to train a collimator detector that generalizes to real-world data.
Load-bearing premise
The pipeline's usefulness rests on the assumption that its simplified scatter and noise models reproduce real collimator physics closely enough that a detector trained on simulated shadows behaves like one trained on real shadows; this assumption is asserted from the outcome rather than verified component by component.
Editorial extensions
If this is right
- Collimator detection training can proceed without hand-labeled real collimated images, since random labels and shadows are generated on the fly.
- X-ray datasets that have no collimation can be augmented with realistic collimator shadows, increasing the variety of shapes, rotations, and distortions seen by the network.
- The same pipeline can be reused for other image-domain tasks that require collimator masks, such as anatomy-preserving cropping or automatic quality checks, by changing only the label sampler.
- When privacy, safety, or acquisition logistics limit access to real collimated images, synthetic collimators act as a replacement that preserves detection performance on real data.
Reading between the lines
- The scatter hyperparameters $\alpha$, $\beta$, and $c$ are not reported, so a sensitivity study over these values is the natural next step; the visible discrepancy between simulated and real scatter profiles is likely concentrated in the Gaussian shape of the scatter kernel.
- A direct extension of the same recipe would simulate other structured X-ray artifacts, such as anti-scatter grid lines, dead detector elements, or detector line noise, by swapping the binary-mask sampler and the physical parameters while keeping the scatter-and-noise stages.
- If the printed noise equations are taken literally, the Poisson mass function appears with $e^{-z}$ rather than $e^{-\lambda}$, and the stated rearrangement for $\sigma_x$ does not follow from the preceding equation; the exact noise implementation therefore needs to be recovered from code or a corrected derivation, even though the reported Dice scores suggest the central effect survives.
- A testable prediction is that mixing simulated and real collimated images during training should push Dice above SimNet alone, because the simulation adds geometric diversity while real images keep tissue texture.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper presents a three-stage image-processing pipeline for synthesizing collimator shadows in X-ray images: random binary mask generation with rotation and distortion, scatter estimation based on the Ohnesorge convolution-kernel model, and a Poisson-noise model. The authors compare the pipeline output against a real collimated thorax phantom image using line profiles and patch metrics (nMSE, SSIM, PSNR), and train two DeepLabV3 networks—SimNet on real collimator-free images with simulated collimators, and RealNet on real collimated images—reporting Dice scores on three test sets. The central claim is that the simulated collimators are realistic enough to substitute for real collimated training data and to improve generalization on real test images.
Significance. If the claim were fully supported, the pipeline would provide a practical, physically motivated data-augmentation tool for collimator detection, avoiding GAN-based augmentation and enabling dataset expansion in mobile radiography. The paper's idea is attractive, the pipeline is simple, and the qualitative comparison in Figs. 4–5 is suggestive. However, the current evidence is not yet convincing: the DNN comparison lacks statistical support and uses a training-derived test set, and the noise model derivation contains algebraic inconsistencies. The paper also does not report the scatter-model hyperparameters, so the pipeline is not reproducible as presented. With additional experiments and corrections, the contribution could be valuable.
major comments (4)
- [Section 3.2, Table 2] The claim in the Abstract and Discussion that SimNet 'exceeds' RealNet is not supported by the reported results. The General Test set is a subset of the training images, so it speaks to memorization rather than generalization. On the Line Artifacts set the Dice intervals overlap substantially (0.9778 ± 0.025 vs 0.9652 ± 0.038), and on the Implants set RealNet is numerically better (0.9780 ± 0.015 vs 0.9494 ± 0.071). No paired significance test or confidence interval on the differences is provided. Please report per-image paired comparisons (e.g., Wilcoxon signed-rank or bootstrap CIs) and, if appropriate, a proper held-out test set.
- [Section 2.1, Eqs. (6)–(8)] The Poisson noise model is internally inconsistent. Equation (6) contains a typo (e^{-z} instead of e^{-λ}); for a Poisson variable σ = √λ, so the statement 'λ = σ = μ' after Eq. (7) is incorrect. Solving Eq. (8) with σ_x as the added noise standard deviation yields σ_x^2 = αλ(λ − α), not λ(1 − α) as stated. The noise stage should be re-derived from the intended physical process (deterministic intensity scaling versus photon-count reduction) and the implementation checked.
- [Section 2.1, Eq. (2)] The values of the scatter-potential hyperparameters α, β, and c are never reported. The surrounding text cites scatter fractions of 1.2%–2% and 2%–4% of primary intensity, but does not explain how these translate to parameter values. Without these values, the pipeline cannot be reproduced and the realism of the scatter stage cannot be assessed. Please report the parameters and include a sensitivity analysis.
- [Section 2.2] The training setup for RealNet is under-specified: the number of real collimated images, their collimator variation, and the exact training conditions are not stated. Since SimNet generates 'unlimited' random collimators on the fly, the reported comparison may conflate data quantity with physical fidelity. Please report the RealNet training set size and matching data-augmentation conditions, or explicitly discuss this confound.
minor comments (8)
- [Section 2.1, Eq. (1)] The phrase 'assigning its zeros to a damping factor' is unclear; please define M_d explicitly, for example as values 1 outside and α inside the collimated region.
- [Section 2.1, heading] The heading 'Scattered Radation Simulation' contains a typo ('Radation' should be 'Radiation').
- [Section 2.2] The text uses 'DDN' instead of 'DNN' in the sentence beginning 'This is called SimNet...'.
- [Fig. 4 caption] The caption contains 'ouput', which should be 'output'.
- [Section 2.1, Eqs. (2) and (7)–(8)] The symbol α is used both for the scatter-potential exponent in Eq. (2) and for the collimator damping factor in Eqs. (7)–(8); please use distinct symbols to avoid confusion.
- [Section 3.1] The physical validation uses patches from a single phantom acquisition; reporting statistics over multiple acquisitions would strengthen the realism claim.
- [General] The paper does not state whether code or pipeline parameters will be released; for a method paper, a reproducibility statement would be helpful.
- [Section 3.2, Table 2] The 'SimNet Simulated' column is reported but not formally compared with 'SimNet Real'; please clarify whether this column is intended as a sanity check and interpret it explicitly.
Circularity Check
Partial circularity in the generalization claim: Table 2's 'General Test' is a subset of the DNN training images, so SimNet's advantage there is a training-set measurement, not an independent prediction; the pipeline physics itself is externally validated and not circular.
-
fitted input called prediction
[Section 2.2 (Application Case DNN) and Table 2]
"Evaluation is performed on three datasets calculating the Dice score: a subset of 80 randomly extracted training images, 30 challenging cases with dark attenuating line-shaped implants, and 20 images showing detector line artifacts."
SimNet was trained on the same 1500 'real in-house X-ray images' (with collimators cropped out) and RealNet on the uncropped versions. The 'General Test' is described as 'a subset of 80 randomly extracted training images,' i.e., it is drawn from that same training pool. Presenting SimNet's Dice on this set (0.9718) as evidence that it 'even exceeds the performance of RealNet' (0.9641) and 'enhances the generalization performance' is not an independent prediction: both networks were fitted to those base images, so the result can reflect memorization or label leakage. The independent Implants set actually favors RealNet (0.9780 vs 0.9494), so the headline claim is not supported by a clean external comparison.
full rationale
The simulation pipeline itself is not circular: the scatter model (Eqs. 2-3) and Gaussian blur (Eq. 1) are applied to random masks and the output is compared against physically acquired collimator images (Figs. 4-5, Table 1) using nMSE, SSIM, and PSNR. The Poisson noise correction (Eqs. 6-8) is attributed in part to prior work by the same group ([3],[4]), a minor self-citation, but the real-vs-simulated image comparison provides external grounding, so that is not load-bearing. The material circularity is in the DNN evaluation: the paper's claim that simulated data 'enhances the generalization performance' relies on Table 2, whose first row is measured on training images. Because the General Test is a subset of the training images for both networks, that particular prediction reduces to training-set performance by construction. The Line Artifacts and Implants sets are independent, but Implants favors RealNet and the confidence intervals overlap, so the abstract's blanket 'enhances generalization' claim is only partly redeemed by external evidence. This is a partial circularity in the evaluation of the headline claim, not in the physical derivation.
Assumptions & free parameters
free parameters (4)
- Scatter potential hyperparameters (alpha, beta, c) =
not reported
- Gaussian blur kernel widths (Gb, Gs) =
not reported
- Collimator mask sampling parameters =
not reported
- Primary intensity I0 and photon count lambda =
not reported
assumptions (4)
- domain assumption Poisson photon statistics describe X-ray detection at each pixel
- domain assumption Scattered radiation is a smooth low-frequency field estimable by convolution of a scatter potential
- domain assumption Scatter intensity is 1.2-2% of primary in thorax c-arm images, and 2-4% inside collimated regions
- ad hoc to paper Randomized rectangular masks with rotation and distortion cover clinically relevant collimator variations
Cite this review
Pith. "Pith review of A Realistic Collimated X-Ray Image Simulation Pipeline." pith.science (2026). https://pith.science/paper/RKQU2ELM
@misc{pith2026241110308,
author = {Pith},
title = {Pith review of: A Realistic Collimated X-Ray Image Simulation Pipeline},
year = {2026},
howpublished = {\url{https://pith.science/paper/RKQU2ELM}},
note = {Machine review of arXiv:2411.10308}
}
read the original abstract
Collimator detection remains a challenging task in X-ray systems with unreliable or non-available information about the detectors position relative to the source. This paper presents a physically motivated image processing pipeline for simulating the characteristics of collimator shadows in X-ray images. By generating randomized labels for collimator shapes and locations, incorporating scattered radiation simulation, and including Poisson noise, the pipeline enables the expansion of limited datasets for training deep neural networks. We validate the proposed pipeline by a qualitative and quantitative comparison against real collimator shadows. Furthermore, it is demonstrated that utilizing simulated data within our deep learning framework not only serves as a suitable substitute for actual collimators but also enhances the generalization performance when applied to real-world data.
Figures
Figures from the paper (3 more)
Reference graph
Works this paper leans on
-
[1]
arXiv preprint arXiv:1810.10863 (2018)
Bowles, C., Chen, L., Guerrero, R., Bentley, P., Gunn, R., Hammers, A., Dickie, D.A., Hern´ andez, M.V., Wardlaw, J., Rueckert, D.: Gan augmentation: Aug- menting training data using generative adversarial networks. arXiv preprint arXiv:1810.10863 (2018)
arXiv 2018
-
[2]
Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. ArXiv abs/1706.05587 (2017)
arXiv 2017
-
[3]
In: Medical Imaging 2022: Physics of Medical Imaging
Eckert, D., Ritschl, L., Herbst, M., Wicklein, J., Vesal, S., Kappler, S., Maier, A., Stober, S.: Deep learning based denoising of mammographic x-ray images: an investigation of loss functions and their detail-preserving properties. In: Medical Imaging 2022: Physics of Medical Imaging. vol. 12031, pp. 455–462. SPIE (2022)
work page 2022
-
[4]
In: Bildverarbeitung f¨ ur die Medizin 2020: Algorithmen–Systeme–Anwendungen
Eckert, D., Vesal, S., Ritschl, L., Kappler, S., Maier, A.: Deep learning-based denoising of mammographic images using physics-driven data augmentation. In: Bildverarbeitung f¨ ur die Medizin 2020: Algorithmen–Systeme–Anwendungen. Pro- ceedings des Workshops vom 15. bis 17. M¨ arz 2020 in Berlin. pp. 94–100. Springer (2020)
work page 2020
-
[5]
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Commun. ACM 63(11), 139–144 (oct 2020). https://doi.org/10.1145/3422622, https://doi.org/ 10.1145/3422622
doi:10.1145/3422622 2020
-
[6]
Kawashita, I., Aoyama, M., Kajiyama, T., Asada, N.: Collimation detection in digital radiographs using plane detection hough transform. In: Knowledge-Based Intelligent Information and Engineering Systems: 7th International Conference, KES 2003, Oxford, UK, September 2003. Proceedings, Part II 7. pp. 394–401. Springer Berlin Heidelberg (2003)
work page 2003
-
[7]
Future Internet 13(1), 8 (2020)
Kora Venu, S., Ravula, S.: Evaluation of deep convolutional generative adversarial networks for data augmentation of chest x-ray images. Future Internet 13(1), 8 (2020)
work page 2020
-
[8]
Krieger, H.: Grundlagen der Strahlungsphysik und des Strahlenschutzes, vol. 2. Springer (2007)
work page 2007
Show all 15 references
-
[9]
In: CT meeting (2018)
Luckner, C., Mertelmeier, T., Maier, A., Ritschl, L.: Estimation of the source- detector alignment of cone-beam x-ray systems using collimator edge tracking. In: CT meeting (2018)
2018
-
[10]
In: Medical imaging 2018: Image processing
Madani, A., Moradi, M., Karargyris, A., Syeda-Mahmood, T.: Chest x-ray gen- eration and data augmentation for cardiovascular abnormality classification. In: Medical imaging 2018: Image processing. vol. 10574, pp. 415–420. SPIE (2018)
2018
-
[11]
In: Medical Imaging 2014: Image Processing
Mao, H., Peng, Z., Dennerlein, F., Shinagawa, Y., Zhan, Y., Zhou, X.S.: Multi- view learning based robust collimation detection in digital radiographs. In: Medical Imaging 2014: Image Processing. vol. 9034, pp. 525–530. SPIE (2014)
2014
-
[12]
Engineering Proceedings 31(1), 84 (2023)
Ng, M.F., Hargreaves, C.A.: Generative adversarial networks for the synthesis of chest x-ray images. Engineering Proceedings 31(1), 84 (2023)
2023
-
[13]
European radiology 9(3), 563–569 (1999) 10 B
Ohnesorge, B., Flohr, T., Klingenbeck-Regn, K.: Efficient object scatter correction algorithm for third and fourth generation ct scanners. European radiology 9(3), 563–569 (1999) 10 B. El-Zein et al
1999
-
[14]
Medical physics 40(5), 051915 (2013)
Sisniega, A., Zbijewski, W., Badal, A., Kyprianou, I., Stayman, J., Vaquero, J.J., Siewerdsen, J.: Monte carlo study of the effects of system geometry and antiscatter grids on cone-beam ct scatter distributions. Medical physics 40(5), 051915 (2013)
2013
-
[15]
In: International Conference on Nuclear Engineering
Xu, S., Chen, G., Li, W., Xiang, X.: A physics-driven x-ray image data augmenta- tion method for automated threat detection in nuclear facility entrancement. In: International Conference on Nuclear Engineering. vol. 86397, p. V005T05A041. American Society of Mechanical Enginee...
2022
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.