List of Papers Browse by Subject Areas Author List
Abstract
In minimally invasive surgery, surgical smoke generated by electrocautery devices degrades the laparoscopic view, obstructing the surgeon’s visual field and increasing operative risks. Existing smoke removal methods often overlook the inherent spectral discrepancies between global smoke artifacts and local scene details. Consequently, these approaches struggle to strike a balance between removing heavy smoke and preserving fine surgical textures. Based on wavelet decomposition analysis, we observe a critical phenomenon: replacing the low-frequency component of a smoke-degraded image with that of a clear reference effectively eliminates visual smoke, while residual errors are predominantly concentrated in high-frequency components. This indicates that surgical smoke degradation is structurally decoupled: global smoke dominates low frequencies, while texture degradation resides in high frequencies. Motivated by this observation, we propose LoHi-Net, a Low-to-High Hierarchical Wavelet Network. Unlike conventional holistic or independent processing schemes, LoHi-Net adopts a hierarchical guided restoration mechanism. Specifically, it first restores low-frequency structural components to eliminate global smoke effects, which then serve as a structural prior to guide the precise refinement of high-frequency textures. To maintain model compactness while capturing the distinct characteristics of each frequency band,we introduce specialized Mamba variants into the architecture for hierarchical modeling. Extensive experiments on real-world datasets demonstrate that LoHi-Net achieves state-of-the-art performance, effectively recovering high-fidelity surgical scenes with compact model capacity.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0650_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/ShiweiWu98/LoHi-Net
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WuShi_LoHiNet_MICCAI2026,
author = { Wu, Shiwei AND Hu, Xiyuan AND Zhu, Xiaobo AND Gao, Dongyang AND Zhang, Song AND Yan, Hui},
title = { { LoHi-Net: A Low-to-High Hierarchical Wavelet Network for Laparoscopic Surgical Smoke Removal } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16892},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a novel LoHi-Net for laparoscopic surgical smoke removal. Based on wavelet analysis, it explicitly decouples smoke degradation into distinct problems: low-frequency and high-frequency. A hierarchical guided restoration mechanism (low-frequency first, then high-frequency) and a low-to-high structural guidance module are introduced, closely integrated with the wavelet transform. Customized Mamba variants are developed for different frequency components, which is key to achieving a lightweight model.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper is built upon a clear, data-driven observation: surgical smoke degradation is structurally decoupled between low-frequency and high-frequency components, as demonstrated through a simple yet effective wavelet substitution experiment. This frequency-decoupling perspective provides a strong, intuitive theoretical foundation for the entire method, making the subsequent architectural choices logical and compelling.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.While the paper reports parameters and MACs, which are relevant for practical deployment, it is suggested to also include a comparison of inference speed. 2.Among the compared methods, only one is specifically designed for endoscopic desmoking. To better evaluate the performance within the targeted application domain, it is recommended to increase the proportion of comparison methods that are specifically developed for surgical or endoscopic smoke removal.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed LoHi-Net translates the core insight directly into a novel network architecture. The “Low-to-High Hierarchical” restoration strategy (restoring structure first, then details) and the dedicated L2H structural guidance block are original and elegant implementations of the frequency-decoupling principle.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have provided a compelling response that effectively addressed the most critical concerns raised during my review process. The core strengths of the paper—a clear frequency-decoupling insight translated into a novel low-to-high hierarchical architecture that delivers both strong performance and efficiency—remain compelling.
Review #2
- Please describe the contribution of the paper
The main contribution of this paper is the proposal of LoHi-Net, a lightweight framework for laparoscopic surgical smoke removal. Based on wavelet decomposition, it introduces a Hierarchical Guided Restoration (HGR) strategy that explicitly decouples global smoke degradation (low-frequency) from surgical details (high-frequency). Additionally, it integrates sub-band-specific Mamba variants for frequency-aware modeling , delivering state-of-the-art performance with a highly compact model capacity of only 1.38M parameters.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Formulating surgical smoke removal through wavelet decomposition to decouple global smoke (low-frequency) from fine textures (high-frequency) is physically intuitive and insightful. Additionally, introducing sub-band-specific Mamba variants is an effective architectural design to balance sequence modeling and structural alignment.
2.The experiments demonstrate state-of-the-art performance on two real-world datasets (VASST-desmoke and LSVD). Notably, the proposed method achieves superior quantitative metrics (SSIM, PSNR, CIEDE) while maintaining an extremely lightweight model capacity of only 1.38M parameters.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The authors state they “randomly divide” the VASST-desmoke dataset into 70/30 train/test sets. However, the original dataset is explicitly composed of continuous video sequences.A simple image-level random split on these highly temporally correlated frames guarantees severe data leakage and over-optimistic results. The authors must clarify if their split was strictly performed at the independent sequence level. 2.There is a logical gap between Section 2.2 and Section 2.3 regarding the high-frequency components. The authors define the enhanced high frequency component as $\hat{W}$ in Eq.(3),but switch to $\hat{H}$,$\hat{V}$,$\hat{D}$ in Eq.(6) without any definition. 3.In Equation (7), the authors apply a global FFT loss with an empirically set weight. However, they fail to provide a clear justification for applying a global FFT constraint on top of the already localized wavelet-domain representations. Furthermore, there is a lack of ablation studies to demonstrate whether this specific frequency loss term quantitatively contributes to the final performance. 4.Exploring frequency cues for restoration has been valided in many tasks, sush as [1-3], the unique insight of this method from these methods are less dicussed.
[1]AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation. ICLR25 [2]Seeing the Unseen: A Frequency Prompt Guided Transformer for Image Restoration. ECCV 2024 [3]When Fast Fourier Transform Meets Transformer for Image Restoration. ECCV 2024
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While this paper proposes an intuitive frequency-domain decoupling strategy for surgical smoke removal, its ambiguous dataset splitting strategy (“randomly divide”) on temporally correlated video sequences raises a severe concern for potential image-level data leakage. Furthermore, the manuscript lacks mathematical rigor with unexplained notation switches (Eq.3 and Eq.6) and fails to provide theoretical justification or ablation studies for the empirically weighted global FFT loss.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After carefully reading the rebuttal from the authors and reviews from reviewers, most of my concerns are addressed. The promised revised version could be ready to be published.
Review #3
- Please describe the contribution of the paper
This paper presents LoHi-Net, a wavelet-based hierarchical restoration network for laparoscopic surgical smoke removal. The key idea is to first restore the low-frequency component to suppress global smoke degradation, and then use the recovered low-frequency structure as guidance for refining high-frequency details. The method further introduces frequency-specific Mamba variants for different wavelet sub-bands.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper is built around an intuitive observation from wavelet decomposition: smoke degradation mainly affects global low-frequency appearance, while fine texture loss is more reflected in high-frequency components. Based on this, the proposed low-to-high hierarchical restoration pipeline is coherent and easy to follow. 2.The reported results suggest that the method achieves competitive or best performance on the tested datasets while maintaining a compact model size, which is attractive for practical deployment.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.While the overall framework is well organized, the technical novelty appears moderate. The method combines several existing ingredients, including wavelet decomposition, structural guidance, and Mamba-based modeling. The novelty may need to be more clearly distinguished from prior frequency-aware restoration frameworks. 2.The paper’s design is motivated by the claim that smoke degradation is mainly low-frequency, whereas detail restoration is mainly high-frequency. However, the evidence is largely based on a qualitative example in Fig. 1.This observation is interesting, but it would be much more convincing if supported by dataset-level quantitative analysis. 3.There are a few writing issues and typos (e.g., “Experitments”), suggesting the manuscript could benefit from further polishing. 4.The design of the directional kernels in Eq. (4), especially the diagonal kernel, would benefit from more explanation, as the current formulation is not fully intuitive. 5.What is the inference speed of the proposed method?
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper addresses an important problem and proposes a technically reasonable and efficient solution. The framework is coherent, the empirical results are promising, and the parameter efficiency is attractive. These are positive aspects. However, in my view, the current submission falls short of the acceptance bar for MICCAI for two main reasons: (1) the methodological novelty is moderate; (2) the central frequency-domain motivation is not validated rigorously enough.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The rebuttal clarifies the dataset split, but the contribution remains incremental. Clinical validation, temporal consistency, and detailed ablations are still insufficient for MICCAI acceptance.
Author Feedback
We sincerely thank all reviewers for their valuable feedback. We respond to their comments below. 1.Dataset split (R2). We sincerely apologize that “randomly divide” in Sec. 3.1 may have caused confusion. The VASST-desmoke split in the submitted experiments was at the sequence/scene level, not image/frame level. Frames from the same continuous surgical sequence/scene were never shared between training and testing; thus, the reported evaluation avoids temporal leakage from correlated frames, supporting its reliability. We will revise Sec. 3.1 accordingly. 2.Motivation and Novelty (R3&R2): LoHi-Net’s main innovation is a task-specific low-to-high hierarchical wavelet restoration framework: LL is restored first to suppress global smoke-related appearance degradation, then guides H/V/D high-frequency recovery via direction-specific priors and sub-band-specific Mamba scanning. Unlike prior frequency-aware methods mainly using modulation, prompting, or global mixing, LoHi-Net treats sub-bands as explicit restoration targets with an LL-to-H/V/D dependency. Fig. 1 shows our wavelet-substitution analysis: replacing smoky LL with clean LL while keeping H/V/D unchanged largely removes visible smoke, and residual errors remain around high-frequency structures. The averaged normalized MAE ratios of DWT sub-bands on paired images from the VASST-desmoke dataset indicate that the smoke-related discrepancy is dominated by the LL sub-band (79.46%) rather than the high-frequency H/V/D sub-bands. Together, these observations support LoHi-Net’s low-to-high design. 3.Ablation (R2): Table 2 ablates the hierarchical strategy, sub-band-specific SSM, branch depth design, and L2H guidance, supporting our core contribution of low-to-high hierarchical wavelet restoration. In Eq. (7), L1+FFT is effective in our setting: both are applied to image and wavelet-coefficient domains but penalize complementary errors. L1 penalizes point-wise reconstruction errors, while FFT penalizes Fourier-domain spectral discrepancies with the ground-truth, helping reduce smoke-related low-frequency energy bias, contrast attenuation, illumination inconsistency, and color shift. Removing FFT loss, or applying it only to the final output without multi-scale wavelet-domain constraints, leads to performance degradation. 4.Kernel design (R3): The kernels in Eq. (4) extract structural priors from restored L’, rather than exactly reproduce DWT filters. K_H/K_V are first-order directional differences along two spatial axes. K_D is a Laplacian-like second-order residual operator capturing mixed-axis high-frequency variations, providing complementary structural cues for the diagonal sub-band D. These priors are fused with H/V/D to guide texture and boundary recovery without generic detail amplification. 5.Comparisons and latency (R1&R3): We thank the reviewers for these comments. Our comparison already includes two surgical/endoscopic desmoking methods, SVP-Net and LGUTransformer, covering physical-prior/CNN-based and Transformer-based designs. We also include advanced dehazing/restoration methods as transferable baselines. LoHi-Net achieves the best overall performance with the lowest Params/MACs in Table 1 and runs at ~23 FPS with single-image input under the evaluation setting described in Sec. 3.1.6.Notation and writing (R2&R3): We sincerely apologize for the confusion caused by unclear notation and writing mistakes. In Eq. (3), $\hat{W}$ denotes the L2H-enhanced version of each high-frequency sub-band $W$ in {H, V, D}, namely $\hat{H}$, $\hat{V}$, and $\hat{D}$, which are then restored by sub-band-specific Mamba branches in Eq. (6). We will clarify this transition and proofread the manuscript to correct writing issues and typos, including “Experitments”. 7.Reproducibility (R1&R2&R3): We will release code upon acceptance to support reproducibility.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper overall has a clear idea with strong empirical performance. R2 raises a potential fatal data leakage concern.
Key points to address:
- dataset split
- ablation studies
- method clarification
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The rebuttal well addresses the possibility of image-level leakage from temporally correlated video frames. The remaining weaknesses are mostly about scope and claim calibration. The novelty is moderate, and the paper should better distinguish itself from prior frequency-aware restoration methods. Overall, it’s above the threshold for acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After rebuttal there are two Accept and one Reject. R1 and R2 are satisfied with the rebuttal. However, R3 points out incremental contributions, plus inadequate clinical validation, temporal consistency tests and ablations. Considering pros outweighs cons, I incline to Accept.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After the rebuttal, the paper received two positive recommendations and one negative recommendation. I carefully reviewed the rebuttal and the reviews. Overall, I believe that the authors have addressed most of the concerns raised during the review process, and the remaining issues do not substantially undermine the contribution of the work. Therefore, I recommend accepting this paper for publication.