Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Laparoscopic Image Desmoking Using the U-Net with New Loss Function and Integrated Differentiable Wiener Filter

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims a U-Net with a differentiable Wiener filter and a three-part loss outperforms both the base U-Net and Pix2Pix on real paired laparoscopic desmoking images.

desk verdict A plausible desmoking recipe whose reported gains may be inflated by a scene-leaking split; worth refereeing with a demand for video-level evaluation. read the letter →

arxiv 2505.21634 v1 pith:SWXGOVDR submitted 2025-05-27 eess.IV cs.CV

Chengyu Yang , Chengjun Liu This is my paper
classification eess.IVcs.CV
keywords surgicalsmokeremovallaparoscopicimagingU-NetWienerfilterperceptuallossSSIMimagedesmokingmedicalenhancement
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a standard U-Net, augmented with a learnable Wiener filter layer and a three-term loss, removes surgical smoke from laparoscopic frames better than the same U-Net alone and better than the Pix2Pix conditional GAN. On the test split of a public dataset of 961 paired smoke and smoke-free images, it reports SSIM 0.9907, PSNR 33.7061, MSE 0.0006, and a CIEDE-2000 color-difference score of 1.8159, all better than both baselines. The practical point is that high-quality desmoking can be achieved with a modest supervised network on real surgical footage, without adversarial training.

What carries the argument

The load-bearing mechanism is the differentiable Wiener filter layer placed after the input: $\hat{x} = s \odot \frac{p}{p+\sigma_n^2+\epsilon}$, where $s=\mathcal{F}(x)$ is the filtered signal, $p=s^2$ is the local signal power estimate, $\sigma_n^2$ is a learnable noise variance, and $\epsilon$ is a stability constant. Initialized as Gaussian kernels and trained with the rest of the network, it gives the U-Net an explicit degradation model for smoke. The second mechanism is the composite loss $L = \alpha L_{\mathrm{MSE}} + \beta L_{\mathrm{SSIM}} + \gamma L_{\mathrm{perceptual}}$ with $\alpha=\beta=\gamma=1/3$, where the perceptual term is computed from a pre-trained VGG network, a deep image-classification network. This loss is what the authors credit for balancing pixel accuracy, structural fidelity, and visual realism.

What would settle it

Retrain the same U-Net and the two baselines with the same 961 image pairs but split by video sequence or surgical procedure, so no frames from one surgery appear in both training and test; if ULW no longer beats Pix2Pix by the reported margins, the original random split was responsible.

Watch

Extended reading notes

Core claim

The authors' central claim is that combining a differentiable Wiener filter with a U-Net and training with an equal-weight mix of MSE loss, SSIM loss, and perceptual loss yields the best desmoking quality among the three compared methods on the paired laparoscopic benchmark. They attribute the PSNR and noise-suppression gains to the Wiener filter layer, which learns its noise variance end-to-end, and the structural gains to the explicit SSIM training objective. They conclude that the method offers a practical route to real-time enhancement of laparoscopic imagery.

Load-bearing premise

The load-bearing assumption is that the test images come from surgeries the model has never seen; if the random split mixes frames from the same surgical video into both training and test sets, the reported quality gains could be exaggerated.

Editorial extensions

If this is right

  • The reported results imply that adversarial training is not necessary for supervised laparoscopic desmoking; a U-Net with a physics-inspired filter layer can beat a conditional GAN on this dataset.
  • The learnable Wiener layer can be reused as a front-end in other image restoration tasks where noise or haze follows a signal-dependent degradation.
  • An equal-weight combination of MSE, SSIM, and perceptual loss is a transferable recipe for training medical image enhancement networks.
  • If the inference cost is dominated by a small U-Net plus one filter layer, the method is plausible for real-time use within a surgical video pipeline.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not test whether a procedure-level split preserves the margin; if frames from the same surgery appear in both training and test, the reported numbers could be optimistic.
  • A natural follow-up the paper does not explore is temporal desmoking: feeding consecutive frames through a recurrent or 3D version of the Wiener filter might smooth smoke removal across time.
  • The paper fixes the loss weights at equal thirds and does not report sensitivity; weighting the perceptual term differently could trade PSNR for texture realism.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes ULW, a U-Net backbone augmented with a differentiable Wiener-filter layer and a composite loss combining MSE, SSIM, and VGG-based perceptual loss, for laparoscopic image desmoking. The method is evaluated on the paired in vivo laparoscopic smoke dataset of Xia et al. (961 image pairs from 21 video sequences), with an 80/10/10 train/validation/test split. The authors report that ULW outperforms both a U-Net baseline trained with MSE only and a Pix2Pix baseline on SSIM, PSNR, MSE, and CIEDE-2000, with best values of 0.9907, 33.7061, 0.0006, and 1.8159, respectively.

Significance. If the reported results are robust, the contribution is a simple, practical recipe—adding a learnable Wiener-filter layer and a multi-term loss to a standard U-Net—that improves laparoscopic desmoking over two reasonable baselines on real paired data. Strengths include the use of a public in vivo dataset, the inclusion of an external perceptual color metric (CIEDE-2000), and a public code repository. However, the significance is currently limited by the evaluation protocol: a single random split with no video-level separation, no multiple runs or error bars, and no statistical testing. The reported SSIM/MSE gains are also partly expected because these terms appear in the training loss, although the comparison against baselines trained with different losses and the CIEDE-2000 result provide independent grounding.

major comments (4)
  1. [IV-A] The split protocol is not described at the video or procedure level. Section IV-A says only that eighty percent of the 961 paired images are used for training and ten percent for testing. Because the dataset contains 961 frames drawn from 21 video sequences across 63 procedures, a random image-level split will almost certainly place frames from the same surgical video into both training and test sets. Adjacent laparoscopic frames are highly correlated in anatomy, illumination, and smoke pattern, so the network can memorize scene-specific features and report inflated SSIM/PSNR/MSE values. The paper should either state explicitly that the split was performed at the video/procedure level or re-run the evaluation with a video-level split; otherwise the claimed generalization advantage over Pix2Pix and the base U-Net is not established.
  2. [IV-C, Table I] All quantitative results in Table I come from a single train/validation/test split and a single training run. No error bars, multiple seeds, or statistical significance tests are reported. The SSIM gap between ULW and Pix2Pix is 0.0115 and the PSNR gap is 3.09 dB, but without a variance estimate or a leakage-free split it is impossible to know whether these differences are meaningful. The authors should report mean and standard deviation over multiple splits or seeds, and ideally a paired test across test images, to support the claim of consistent superiority.
  3. [III-B, Eq. (1)] Equation (1) defines the Wiener-filter layer as a pointwise gain x_hat = s ⊙ p/(p + σ_n^2 + ε), where s = F(x) and p = s^2. This is a local denoising operation, not a Wiener deconvolution in the usual sense of estimating an inverse filter from a known degradation kernel H. The manuscript claims the layer 'models the degradation process caused by surgical smoke' and cites Deep Wiener Deconvolution, but the relation between smoke degradation and this filter form is not formalized. The authors should specify how F is parameterized and learned, whether it operates on spatial patches or in the frequency domain, and why this form constitutes a Wiener filter for smoke rather than a generic learned denoising layer.
  4. [III-C, Eqs. (2)-(6)] Because the training loss directly minimizes MSE and 1-SSIM, reporting that ULW achieves lower MSE and higher SSIM than the base U-Net (which is trained with MSE only) is partly a consequence of optimizing those exact objectives. The authors acknowledge this in the discussion of the SSIM score, but the abstract and conclusion present SSIM/MSE superiority as evidence of general effectiveness. The CIEDE-2000 result and the comparison against Pix2Pix are more independent and should be foregrounded; the authors should also report at least one metric not present in any training loss (beyond CIEDE-2000) to strengthen the evaluation.
minor comments (5)
  1. [II] There are multiple typos and grammar issues, including 'funciton' (Section III and IV), 'downsamping' (Section III-A), 'eletrocautery' and 'obsures' (Section I), and 'surgeons's' (Section I). The manuscript would benefit from a careful proofread.
  2. [IV-A] The ablation and training details are incomplete: the authors do not specify the optimizer, learning rate, batch size, number of epochs, input resolution, or hardware. Without these, the experiments are difficult to reproduce. The code repository may supply these, but the paper should state them.
  3. [IV-B] The terms 'base model' and 'foundational model' are used inconsistently; Table I lists 'Pix2Pix' while the text sometimes calls it 'foundational model.' Please use a single consistent name for each method.
  4. [III-C, Eq. (6)] The perceptual loss uses features from a VGG network, but the specific layer index l is never given. Since perceptual loss values depend strongly on the layer chosen, the authors should specify which layer is used.
  5. [V] The conclusion states the method offers a promising solution for 'real-time enhancement,' but no runtime or inference speed measurements are reported. Either add timing results or remove the real-time claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is an empirical combination of a standard U-Net, a learnable Wiener-filter layer, and a composite loss; the reported metrics are test-set measurements against held-out paired data, not quantities forced by construction.

full rationale

The derivation chain in this paper is empirical rather than circular. The proposed ULW method combines (i) a U-Net backbone (Section III-A), (ii) a differentiable Wiener-filter layer defined by Eq. (1) with learnable noise variance and Gaussian-initialized filters, and (iii) a composite loss in Eqs. (2)-(6) that mixes MSE, SSIM, and perceptual losses. The experimental section then measures SSIM, PSNR, MSE, and CIEDE-2000 on a held-out 10% test split. Although SSIM and MSE appear both in the training objective and in the evaluation table, the claim is that the trained model generalizes to unseen test pairs and outperforms two baselines trained under different objectives; that is a contingent empirical result, not an identity. The baseline comparisons and the CIEDE-2000 metric, which is not part of the loss function, provide independent evidence. The self-citations [4] and [6] appear only as motivating examples in the introduction and are not load-bearing for the desmoking method or its evaluation. No fitted parameter is renamed as a prediction, and no uniqueness claim is imported from the authors' prior work. The potential concern that the 80/10/10 split may not be video-level and could leak temporally correlated frames into the test set is a data-validity issue, not a circularity of the derivation chain, and does not affect the circularity score.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim relies on the unverified modeling assumption of the Wiener layer, the evaluation split assumption, and the representativeness of the single dataset. The loss weights and learnable noise variance are free parameters. No new physical entities are introduced.

free parameters (3)
  • Loss weights alpha, beta, gamma = alpha=beta=gamma=1/3
    Chosen by hand, not optimized; any weights satisfying alpha+beta+gamma=1 are used, and the choice is not justified or ablated.
  • Learnable noise variance sigma_n^2 = learned during training
    The noise variance in Eq. (1) is a learnable parameter that is fit to the training data.
  • Wiener filter convolution kernels F = initialized as Gaussian kernels, then learned
    The filter is learnable, but its size, number of channels, and exact placement are unspecified.
assumptions (3)
  • ad hoc to paper Wiener filter layer models smoke degradation
    Eq. (1) is asserted to model the degradation process caused by surgical smoke, but no physical or statistical derivation connects surgical smoke to this element-wise gain formula. The claim is an assumption.
  • domain assumption Random 80/10/10 split of image pairs provides an unbiased evaluation
    Section IV-A states the split percentages but does not confirm that images from the same video sequence are kept in the same split; if not, temporal correlation biases the test set.
  • domain assumption The in vivo paired dataset is representative of laparoscopic smoke conditions
    The method is evaluated solely on this dataset, with no cross-dataset validation, so generalization to other surgical conditions is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Laparoscopic Image Desmoking Using the U-Net with New Loss Function and Integrated Differentiable Wiener Filter." pith.science (2026). https://pith.science/paper/SWXGOVDR

@misc{pith2026250521634,
  author       = {Pith},
  title        = {Pith review of: Laparoscopic Image Desmoking Using the U-Net with New Loss Function and Integrated Differentiable Wiener Filter},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SWXGOVDR}},
  note         = {Machine review of arXiv:2505.21634}
}
read the original abstract

Laparoscopic surgeries often suffer from reduced visual clarity due to the presence of surgical smoke originated by surgical instruments, which poses significant challenges for both surgeons and vision based computer-assisted technologies. In order to remove the surgical smoke, a novel U-Net deep learning with new loss function and integrated differentiable Wiener filter (ULW) method is presented. Specifically, the new loss function integrates the pixel, structural, and perceptual properties. Thus, the new loss function, which combines the structural similarity index measure loss, the perceptual loss, as well as the mean squared error loss, is able to enhance the quality and realism of the reconstructed images. Furthermore, the learnable Wiener filter is capable of effectively modelling the degradation process caused by the surgical smoke. The effectiveness of the proposed ULW method is evaluated using the publicly available paired laparoscopic smoke and smoke-free image dataset, which provides reliable benchmarking and quantitative comparisons. Experimental results show that the proposed ULW method excels in both visual clarity and metric-based evaluation. As a result, the proposed ULW method offers a promising solution for real-time enhancement of laparoscopic imagery. The code is available at https://github.com/chengyuyang-njit/ImageDesmoke.

Figures

Figures reproduced from arXiv: 2505.21634 by the authors.

Figure 1
Figure 1. The system architecture of the proposed ULW method. The U-Net is used as the backbone while a differentiable Wiener filter layer is integrated after [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The visual presentation of the desmoking results produced by the base model, the foundational model, and the proposed ULW method, respectively. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Investigating the Impact of Various Loss Functions and Learnable Wiener Filter for Laparoscopic Image Desmoking

    cs.CV 2025-09 reject novelty 2.0 of 10

    Ablation of the ULW desmoking framework finds SSIM loss is essential, while dropping perceptual loss improves all reported metrics, contradicting the paper's necessity claim.

Reference graph

Works this paper leans on

21 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [1]

    Removal of smoke effects in laparoscopic surgery via adversarial neural network and the dark channel prior,

    S. Salazar-Colores, H. A. Moreno, U. Moya, C. J. Ortiz-Echeverri, L. A. Tavares de la Paz, and G. Flores, “Removal of smoke effects in laparoscopic surgery via adversarial neural network and the dark channel prior,” Cirugia y Cirujanos (English Edition) , vol. 90, no. 1, pp. 74–83, 2022

    work page 2022

  2. [2]

    A new benchmark in vivo paired dataset for laparoscopic image de-smoking,

    W. Xia, V . Fan, T. Peters, and E. C. Chen, “A new benchmark in vivo paired dataset for laparoscopic image de-smoking,” in International Conference on Medical Image Computing and Computer-Assisted Inter- vention. Springer, 2024, pp. 3–13

    work page 2024

  3. [3]

    Desmoke-lap: improved unpaired image-to-image translation for desmoking in laparoscopic surgery,

    Y . Pan, S. Bano, F. Vasconcelos, H. Park, T. T. Jeong, and D. Stoy- anov, “Desmoke-lap: improved unpaired image-to-image translation for desmoking in laparoscopic surgery,” International Journal of Computer Assisted Radiology and Surgery , vol. 17, no. 5, pp. 885–893, 2022

    work page 2022

  4. [4]

    Interpretable automatic rosacea detection with whitened cosine similarity,

    C. Yang and C. Liu, “Interpretable automatic rosacea detection with whitened cosine similarity,” in 2025 IEEE 17th International Conference on Computer Research and Development (ICCRD) . IEEE, 2025, pp. 42–46

    work page 2025

  5. [5]

    Skin disease detection using deep learning,

    S. Inthiyaz, B. R. Altahan, S. H. Ahammad, V . Rajesh, R. R. Kalangi, L. K. Smirani, M. A. Hossain, and A. N. Z. Rashed, “Skin disease detection using deep learning,” Advances in Engineering Software , vol. 175, p. 103361, 2023

    work page 2023

  6. [6]

    Increasing rosacea awareness among population using deep learning and statistical approaches,

    C. Yang and C. Liu, “Increasing rosacea awareness among population using deep learning and statistical approaches,” in International Confer- ence on Medical Imaging and Computer-Aided Diagnosis . Springer, 2024, pp. 110–119

    work page 2024

  7. [7]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 . Springer, 2015, pp. 234–241

    2015

  8. [8]

    Deep wiener deconvolution: Wiener meets deep learning for image deblurring,

    J. Dong, S. Roth, and B. Schiele, “Deep wiener deconvolution: Wiener meets deep learning for image deblurring,” Advances in Neural Infor- mation Processing Systems , vol. 33, pp. 1048–1059, 2020

    work page 2020

Show all 21 references
  1. [9]

    Loss functions for image restoration with neural networks,

    H. Zhao, O. Gallo, I. Frosio, and J. Kautz, “Loss functions for image restoration with neural networks,” IEEE Transactions on computational imaging, vol. 3, no. 1, pp. 47–57, 2016

    2016

  2. [10]

    Perceptual losses for real-time style transfer and super-resolution,

    J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losses for real-time style transfer and super-resolution,” in Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11- 14, 2016, Proceedings, Part II 14 . Springer, 2016, pp. 694–711

    2016

  3. [11]

    Single image haze removal using dark channel prior,

    K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel prior,” IEEE transactions on pattern analysis and machine intelligence, vol. 33, no. 12, pp. 2341–2353, 2010

    2010

  4. [12]

    Generative smoke removal,

    O. Sidorov, C. Wang, and F. A. Cheikh, “Generative smoke removal,” in Machine Learning for Health Workshop . PMLR, 2020, pp. 81–92

    2020

  5. [13]

    Multi-stages de-smoking model based on cyclegan for surgical de-smoking,

    X. Su and Q. Wu, “Multi-stages de-smoking model based on cyclegan for surgical de-smoking,” International Journal of Machine Learning and Cybernetics , vol. 14, no. 11, pp. 3965–3978, 2023

    2023

  6. [14]

    Desmoking laparoscopy surgery images using an image-to-image translation guided by an embedded dark channel,

    S. Salazar-Colores, H. M. Jim ´enez, C. J. Ortiz-Echeverri, and G. Flo- res, “Desmoking laparoscopy surgery images using an image-to-image translation guided by an embedded dark channel,” IEEE Access , vol. 8, pp. 208 898–208 909, 2020

    2020

  7. [15]

    Vision transformers for single image dehazing,

    Y . Song, Z. He, H. Qian, and X. Du, “Vision transformers for single image dehazing,” IEEE Transactions on Image Processing , vol. 32, pp. 1927–1941, 2023

    1927

  8. [16]

    Medical image segmentation review: The success of u-net,

    R. Azad, E. K. Aghdam, A. Rauland, Y . Jia, A. H. Avval, A. Bozorgpour, S. Karimijafarbigloo, J. P. Cohen, E. Adeli, and D. Merhof, “Medical image segmentation review: The success of u-net,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2024

    2024

  9. [17]

    R. C. Gonzales and P. Wintz, Digital image processing . Addison- Wesley Longman Publishing Co., Inc., 1987

    1987

  10. [18]

    Evaluation of ssim loss function in rir generator gans,

    M. Pekmezci and Y . Genc, “Evaluation of ssim loss function in rir generator gans,” Digital Signal Processing , vol. 154, p. 104685, 2024

    2024

  11. [19]

    Comparison of the cielab and ciede2000 color difference formulas,

    C. G ´omez-Polo, M. P. Mu ˜noz, M. C. L. Luengo, P. Vicente, P. Galindo, and A. M. M. Casado, “Comparison of the cielab and ciede2000 color difference formulas,” The Journal of prosthetic dentistry , vol. 115, no. 1, pp. 65–70, 2016

    2016

  12. [20]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556 , 2014

    2014 arXiv

  13. [21]

    Image-to-image translation with conditional adversarial networks,

    P. Isola, J.-Y . Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 1125– 1134

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.