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

Chengjun Liu; Chengyu Yang

REVIEW 5 major objections 6 minor 14 references

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

T0 review · 5 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This ablation study of the ULW laparoscopic desmoking framework finds that each component—the learnable Wiener filter, SSIM loss, and perceptual loss—contributes, with SSIM loss being the most critical for preserving anatomical structure.

desk verdict Ablation of the authors' own ULW desmoking pipeline, but Table 1 undercuts the central claim that every component is necessary. read the letter →

arxiv 2509.09849 v1 pith:O3PK5YPI submitted 2025-09-11 cs.CV

Chengyu Yang , Chengjun Liu This is my paper

classification cs.CV

keywords laparoscopicimagedesmokingablationstudyU-NetlearnableWienerfilterSSIMlossperceptualrestorationsurgicalimaging

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 aims to show that every design choice in the ULW framework for laparoscopic image desmoking is necessary and effective. By removing one component at a time, it claims that SSIM loss is indispensable for structural fidelity, the learnable Wiener filter improves pixel-level accuracy and low-contrast clarity despite a slight metric drop, and perceptual loss restores texture at a small numerical cost. The conclusion is that the full combination balances quantitative accuracy with perceptual quality, making it suitable for clinical use. A sympathetic reader would care because this helps decide which loss terms and priors matter for real surgical image restoration.

What carries the argument

The central mechanism is the compound loss plus a differentiable learnable Wiener filter layer. The Wiener filter estimates the clean image as x̂ = s ⊙ p/(p+σ²+ε), where s is a Gaussian-initialized convolution output, p = s² is local signal power, and σ² is a learnable noise variance; this acts as a spatial soft gate that attenuates low-SNR smoke regions. The loss combines MSE, SSIM loss (1 − SSIM), and VGG-based perceptual loss, each targeting pixel accuracy, structural similarity, and high-level perceptual fidelity respectively. The ablation removes each to isolate its contribution.

What would settle it

Rerun the four variants on the same paired laparoscopic dataset with the exact same loss weights, training budget, and seed; if removing SSIM loss (leaving MSE and perceptual loss with fixed weights) does not drop SSIM to roughly 0.9675 and PSNR by about 4 dB relative to the full model, the central claim that SSIM loss is critical would be overturned.

Watch

Extended reading notes

Core claim

The paper claims, based on four ablation variants evaluated on a paired laparoscopic dataset, that the ULW framework's components are complementary and individually justified. Removing the SSIM loss causes the largest performance drop (SSIM falls from 0.9907 to 0.9675, PSNR by more than 4 dB, CIEDE-2000 nearly doubles), identifying it as the structural backbone. Removing the learnable Wiener filter leaves SSIM and PSNR slightly higher but worsens MSE and visual quality in smoke-heavy, low-contrast regions, supporting its role as a physics-inspired denoising prior. Removing perceptual loss improves some numbers but produces over-smoothed textures, so perceptual loss is credited with visual re

Load-bearing premise

The paper's ablation results assume the compound loss is a fixed, explicitly weighted sum of MSE, SSIM, and perceptual terms, but the manuscript never states the total loss formula or the weights, so removing one term may not cleanly isolate that component's contribution.

Editorial extensions

If this is right

Future laparoscopic desmoking models should include an SSIM-style loss as a primary component, since its removal produces the steepest decline in both structural and color fidelity.
A learnable Wiener filter can serve as a trainable physics prior that improves pixel-wise accuracy and low-contrast detail at a small cost in global metrics, making it a useful plug-in for restoration networks.
Perceptual loss is valuable when clinical interpretability matters more than a marginal gain in PSNR or SSIM, because it prevents over-smoothed outputs.
The full ULW combination demonstrates that accuracy and perceptual quality are not mutually exclusive when loss terms are combined with a denoising prior.
The ablation methodology itself provides a template for evaluating components in image restoration frameworks beyond desmoking.

Reading between the lines

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

Because the manuscript never states the total loss formula or the weights assigned to MSE, SSIM, and perceptual terms, the ablation's clean isolation of each component is not fully verifiable; a re-tuned single-loss variant might behave differently from simply zeroing one term.
The learnable Wiener filter's soft-gating behavior is conceptually similar to attention mechanisms; an untested extension would be to replace it with a standard attention module and compare, which could reveal whether the physics-inspired form offers a unique advantage.
For video desmoking, the Wiener filter's noise suppression may reduce temporal flicker, but this paper does not test video; a temporal-consistency extension would be a natural next step and could amplify the filter's benefits.
The perceptual loss's reliance on an ImageNet-pretrained VGG network may transfer imperfectly to medical image domains; a testable extension is to use a pretrained medical imaging encoder and see whether texture fidelity improves further.

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.

Desk Editor's Note

read the letter

Dear X,

Short version: this is a straightforward ablation of the authors' own ULW framework for laparoscopic desmoking. The one new thing is the table of quantitative results from removing the Wiener filter, the SSIM loss, and the perceptual loss. The components are described clearly, and the paper is easy to follow. It also gets a small point for testing on a public dataset. That's the upside.

The soft spots are the story. The conclusion in Section 6 says the full framework "confirms the necessity and effectiveness of each component," but their own Table 1 shows the opposite for perceptual loss: removing it improves SSIM, PSNR, MSE, and CIEDE-2000. The only defense is an unquantified claim about over-smoothing and reduced texture. That might be true, but there is no perceptual metric, no blinded evaluation, and no statistical test. So the claimed necessity is unsupported.

On top of that, the paper never gives the composite loss formula or the weights for the loss terms. The ablation only makes sense if those are fixed and reported. Without them, "w/o SSIM loss" is ambiguous and the central comparative claim cannot be reproduced.

There are smaller issues: no error bars or run counts, the dataset is not named even though the abstract says it is public, no code, and no baseline as simple as U-Net plus MSE. The Wiener filter ablation is also mixed; the numbers are slightly worse on some metrics and the justification is again visual.

So this is not a paper that supports its conclusion. The ablation itself is a minor empirical result, and it could be useful if the authors redid it honestly, reported the loss weights, added statistics, and softened the conclusion. Right now it reads as an argument for a method the authors already proposed, with the numbers saying something different.

For whom is this? Someone tracking ULW specifically might glance at the table. It won't change clinical practice or open a new direction. I wouldn't cite it, and I wouldn't send it to a serious referee in its current form; it needs major revision first. If a workshop wanted a cautionary example of a conclusion running ahead of the data, it would serve.

Best, ...

Referee Report

5 major / 6 minor

Summary. This manuscript presents an ablation study of the ULW framework for laparoscopic image desmoking. The framework, proposed by the same authors in [3], couples a U-Net backbone with a compound loss of MSE, SSIM, and perceptual terms, plus a differentiable learnable Wiener filter. The paper removes each component in turn and evaluates the variants on a 'publicly available paired laparoscopic images dataset' using SSIM, PSNR, MSE, and CIEDE-2000, with qualitative visual comparisons. The conclusion claims that each component is necessary, and that the full framework achieves a strong balance between quantitative accuracy and perceptual quality.

Significance. A careful, reproducible ablation of a published desmoking method would be a useful contribution, and the paper has a sensible experimental skeleton: paired public data, standard metrics, and explicit removal of the Wiener filter and each loss term. The SSIM-loss ablation is informative, since removal causes a clear drop on all metrics. However, the manuscript lacks the experimental detail needed to make the numerical comparisons meaningful, and the main conclusion about perceptual loss is contradicted by the paper's own Table 1. With additional experiments and an honest reframing, the study could be salvageable; in its current form, the evidence is not strong enough to support the central claims.

major comments (5)

[§4, Table 1; §4.3] The claim that perceptual loss is necessary is directly contradicted by Table 1. Removing perceptual loss improves SSIM (0.9907 → 0.9933), PSNR (33.7061 → 35.8490), MSE (0.0006 → 0.0005), and CIEDE-2000 (1.8159 → 1.3511). Section 4.3 responds by asserting that outputs are 'over-smoothed' and lack texture, but no quantified perceptual measure, blinded visual assessment, or statistical test is presented. Without such evidence, the 'necessity' of perceptual loss is unsupported; at best it is a hypothesis to be tested.
[§3 and §4] The total objective is never written. Equations (3) and (4) define L_SSIM and L_perceptual, but the paper never gives L = λ_MSE L_MSE + λ_SSIM L_SSIM + λ_perc L_perc, nor the weights λ. Consequently, 'w/o SSIM loss' is ambiguous: did the authors simply set that term to zero, or re-tune the remaining weights? Without the full loss formulation and fixed weights, the ablation does not isolate the component's contribution. This is a load-bearing reproducibility gap.
[§4.1, Table 1] The Wiener-filter ablation is described as improving pixel-level reconstruction and visual clarity 'despite slightly lower SSIM and PSNR.' Table 1 shows differences that are tiny and inconsistent: MSE 0.0007 → 0.0006, PSNR 33.9055 → 33.7061, CIEDE-2000 1.8136 → 1.8159. There are no error bars or paired significance tests, so these differences cannot be distinguished from noise. The qualitative claim about low-contrast regions is not quantified.
[§4 (experimental setup)] The manuscript omits essential experimental details: the name and split of the dataset, preprocessing, U-Net configuration, optimizer, learning rate, batch size, epochs, and number of independent runs. All metrics are point estimates. Since the differences in Table 1 are often small (e.g., SSIM 0.9907 vs 0.9909), the absence of reproducibility information is a serious limitation and prevents the reader from verifying the ablation.
[§6] The conclusion that 'the full ULW framework achieves a strong balance... confirming the necessity and effectiveness of each component' overstates the evidence. Only SSIM loss removal causes a decisive drop on all metrics (0.9675 SSIM, 29.1446 PSNR, 0.0016 MSE, 3.1500 CIEDE-2000). For the Wiener filter and perceptual loss, the numerical results are mixed or contrary. The conclusion needs to be limited to what the data can support, or supplemented with additional perceptual or clinical evaluation.

minor comments (6)

[§4, Table 1] The header 'CIEDI-2000' should be 'CIEDE-2000'.
[§3.1] The phrase 'mean squared error MSE)' is missing an opening parenthesis. Also, the definitions of s and p in Eq. (1) could be clarified; p = s^2 makes the filtering operation unusual (the output is a cubic function of s).
[§3.3] Specify which VGG variant and which layer l is used in Eq. (4). Currently the perceptual loss is not reproducible without this information.
[§4] The sentence 'PSNR is the peak signal-to-noise ration' contains a typo; it should read 'ratio'.
[Fig. 2] Provide zoomed-in crops or overlays to support qualitative claims about texture, edge sharpness, and low-contrast regions. The current visual comparison is too small to be assessed by the reader.
[§4 (general)] Please state whether the code and trained models will be released. This is especially important because the paper evaluates the authors' own prior framework and gives no training details.

Circularity Check

0 steps flagged · score 2.0 of 10

No equation-level circularity; the only self-citation is to the authors' own ULW paper and is not load-bearing.

full rationale

This paper is an empirical ablation study rather than a derivation, so there is no chain in which a prediction is definitionally equal to its inputs. The ULW framework is attributed to the authors' prior work [3], but the present manuscript re-specifies the component equations (Wiener filter Eq. 1, SSIM loss Eq. 3, perceptual loss Eq. 4) and reports its own measurements on a public dataset, making the self-citation a matter of provenance rather than load-bearing evidence. The closest potential circularity—training with SSIM loss and then observing that the SSIM metric drops when that loss is removed—is an expected empirical sanity check, not a reduction of the conclusion to the input by construction. The perceptual-loss conclusion is quantitatively weak: Table 1 shows that removing perceptual loss improves every reported metric, and the paper's rebuttal relies on unquantified visual inspection. That is an evidentiary and correctness problem, not a circularity under the enumerated patterns. Similarly, the missing composite-loss formula and the internal PSNR/MSE inconsistency in the Wiener-filter row are reporting weaknesses, not circular reductions. No step was found in which a claimed result is forced by definition, by fitted inputs, or by a self-citation chain.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claims rest on two unstated free parameters: the relative loss weights and the trained Wiener noise variance. Additionally, three domain assumptions are made without validation: the additive-noise model for smoke, the transferability of ImageNet VGG features to surgical imagery, and the existence of a correctly paired public dataset, which is never named.

free parameters (2)

Relative weights of MSE, SSIM, and perceptual losses in the compound loss
The total loss is never written down. The ablation results depend on how the removed term was weighted relative to the others; without this, 'w/o SSIM loss' cannot be interpreted.
Wiener filter learnable noise variance sigma^2
Equation (1) treats sigma^2 as learnable, but the trained value is never reported, so the filter's behavior and sensitivity cannot be assessed.

assumptions (3)

domain assumption Surgical smoke can be represented as additive signal-independent noise suitable for a Wiener filter.
Eq. (1) assumes a clean signal plus noise with local power p and noise variance sigma^2; smoke is a spatially varying scattering/absorption process, and the paper provides no validation of this model.
domain assumption VGG features trained on ImageNet are a valid perceptual fidelity signal for laparoscopic tissue images.
Section 3.3 applies a generic ImageNet-pretrained VGG to medical images without domain adaptation or evidence that these features correlate with surgical readability.
domain assumption A paired dataset of smoke and smoke-free laparoscopic images is available with correct alignment.
The paper never names or describes the dataset, so the assumption of valid pixel-level pairing underlies every metric in Table 1.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Investigating the Impact of Various Loss Functions and Learnable Wiener Filter for Laparoscopic Image Desmoking." pith.science (2026). https://pith.science/paper/O3PK5YPI

@misc{pith2026250909849,
  author       = {Pith},
  title        = {Pith review of: Investigating the Impact of Various Loss Functions and Learnable Wiener Filter for Laparoscopic Image Desmoking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O3PK5YPI}},
  note         = {Machine review of arXiv:2509.09849}
}

read the original abstract

To rigorously assess the effectiveness and necessity of individual components within the recently proposed ULW framework for laparoscopic image desmoking, this paper presents a comprehensive ablation study. The ULW approach combines a U-Net based backbone with a compound loss function that comprises mean squared error (MSE), structural similarity index (SSIM) loss, and perceptual loss. The framework also incorporates a differentiable, learnable Wiener filter module. In this study, each component is systematically ablated to evaluate its specific contribution to the overall performance of the whole framework. The analysis includes: (1) removal of the learnable Wiener filter, (2) selective use of individual loss terms from the composite loss function. All variants are benchmarked on a publicly available paired laparoscopic images dataset using quantitative metrics (SSIM, PSNR, MSE and CIEDE-2000) alongside qualitative visual comparisons.

Figures

Figures reproduced from arXiv: 2509.09849 by the authors.

**Figure 1.** The framework of the ULW method 3 Components in details In this section, we introduce the details about separate components of the ULW method. 3.1 Differentiable Wiener Filter Layer The Wiener filter is a classical signal processing technique used for image denoising and deblurring. It operates by minimizing the mean squared error MSE) between the estimated and true signals, assuming a linear degradation model and … view at source ↗

**Figure 2.** The visual presentation of the desmoking results produced by the ablated models. The first two rows show the paired laparoscopic images with and without smoke, respectively. The third row displays the results of the ULW method without the learnable Wiener filter layer, the fourth row reveals the results of the ULW method without the SSIM loss, the last but one row shows the results of the ULW method without the pe… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 2 linked inside Pith

[3]

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

Yang, Chengyu, and Chengjun Liu. "Laparoscopic Image Desmoking Using the U- Net with New Loss Function and Integrated Differentiable Wiener Filter." arXiv preprint arXiv:2505.21634 (2025)

arXiv 2025
[1]

Analysis of surgical smoke produced by various energy- based instruments and effect on laparoscopic visibility

Weld, Kyle J., et al. "Analysis of surgical smoke produced by various energy- based instruments and effect on laparoscopic visibility." Journal of endourology 21.3 (2007): 347-351

2007
[2]

Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising

Zhang, Kai, et al. "Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising." IEEE transactions on image processing 26.7 (2017): 3142-3155

2017
[4]

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 conference, Munich, Ger- many, October 5-9, 2015, proceedings, part III 18. Springer, 2015, pp. 234–241. 8 Chengyu et al

2015
[5]

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
[6]

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 Confer- ence, Amsterdam, The Netherlands, October 11- 14, 2016, Proceedings, Part II 14. Springer, 2016, pp. 694–711

2016
[7]

Pushing the limits of the wiener filter in image denoising

Bled, Clément, and François Pitié. "Pushing the limits of the wiener filter in image denoising." 2023 IEEE International Conference on Image Processing (ICIP). IEEE, 2023

2023
[8]

The development of the CIE 2000 colour-difference formula: CIEDE2000

Luo, M. Ronnier, Guihua Cui, and Bryan Rigg. "The development of the CIE 2000 colour-difference formula: CIEDE2000." Color Research & Application 26.5 (2001): 340-350

2000

Show all 14 references

[9]

Image quality assessment: from error visibility to structural similarity

Wang, Zhou, et al. "Image quality assessment: from error visibility to structural similarity." IEEE transactions on image processing 13.4 (2004): 600

2004
[10]

Understanding surgical smoke in laparoscopy through Lagrangian Coherent Structures

Kumar, Sandeep, et al. "Understanding surgical smoke in laparoscopy through Lagrangian Coherent Structures." Plos one 18.11 (2023): e0293287

2023
[11]

Very deep convolutional networks for large-scale image recognition

Simonyan, Karen, and Andrew Zisserman. "Very deep convolutional networks for large-scale image recognition." arXiv preprint arXiv:1409.1556 (2014)

2014 arXiv
[12]

Imagenet: A large-scale hierarchical image database

Deng, Jia, et al. "Imagenet: A large-scale hierarchical image database." 2009 IEEE conference on computer vision and pattern recognition. Ieee, 2009

2009
[13]

Imagequality metrics: PSNR vs. SSIM

Hore, Alain, and Djemel Ziou. "Imagequality metrics: PSNR vs. SSIM." 2010 20th international conference on pattern recognition. IEEE, 2010

2010
[14]

Comparison of the cielab and ciede2000 color difference formulas,

C. Gomez-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

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [3]

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

Yang, Chengyu, and Chengjun Liu. "Laparoscopic Image Desmoking Using the U- Net with New Loss Function and Integrated Differentiable Wiener Filter." arXiv preprint arXiv:2505.21634 (2025)

arXiv 2025

[2] [1]

Analysis of surgical smoke produced by various energy- based instruments and effect on laparoscopic visibility

Weld, Kyle J., et al. "Analysis of surgical smoke produced by various energy- based instruments and effect on laparoscopic visibility." Journal of endourology 21.3 (2007): 347-351

2007

[3] [2]

Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising

Zhang, Kai, et al. "Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising." IEEE transactions on image processing 26.7 (2017): 3142-3155

2017

[4] [4]

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 conference, Munich, Ger- many, October 5-9, 2015, proceedings, part III 18. Springer, 2015, pp. 234–241. 8 Chengyu et al

2015

[5] [5]

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

[6] [6]

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 Confer- ence, Amsterdam, The Netherlands, October 11- 14, 2016, Proceedings, Part II 14. Springer, 2016, pp. 694–711

2016

[7] [7]

Pushing the limits of the wiener filter in image denoising

Bled, Clément, and François Pitié. "Pushing the limits of the wiener filter in image denoising." 2023 IEEE International Conference on Image Processing (ICIP). IEEE, 2023

2023

[8] [8]

The development of the CIE 2000 colour-difference formula: CIEDE2000

Luo, M. Ronnier, Guihua Cui, and Bryan Rigg. "The development of the CIE 2000 colour-difference formula: CIEDE2000." Color Research & Application 26.5 (2001): 340-350

2000