List of Papers Browse by Subject Areas Author List
Abstract
Endoscopic imaging plays a crucial role in modern diagnostics and minimally invasive procedures. However, artifacts caused by specular and diffuse reflections present significant challenges, particularly in tasks such as endoscopic image segmentation. Existing methods tackling endoscopic artifacts typically address only one type of reflection, failing to fully account for the non-Lambertian reflectance of endoscopic tissue structures. Therefore, inspired by the simplified Phong model for endoscopy, we propose a two-stage artifact inpainting framework. The first stage suppresses specular artifacts, while the second stage focuses on inpainting diffuse artifacts. Additionally, we introduce a weight map to control the handling of diffuse artifacts, ensuring a more precise enhancement. To evaluate its effectiveness, we focus on its impact on endoscopic image segmentation tasks. Extensive experiments on multiple colonoscopy and dental endoscopy datasets demonstrate that our framework can robustly improve the segmentation performance of endoscopic images, offering better enhancement than existing state-of-the-art methods. Particularly, for zero-shot SAM segmentation of teeth, a significant performance boost is observed after inpainting, with mDice and mIoU increasing from 51.3%/39.3% to 96.1%/93.0%. Code is available at GitHub.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1575_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YuZha_Endoscopic_MICCAI2025,
author = { Yu, Zhangyuan and Du, Chenlin and Liang, Hongrui and Zheng, Xiuqi and Ma, Zeyao and Wu, Mingjun and Ao, Mingwu and Lao, Qicheng},
title = { { Endoscopic Artifact Inpainting for Improved Endoscopic Image Segmentation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15969},
month = {September},
page = {190 -- 200}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents a new method to replace overexposure regions due to specular reflection and diffuse reflection within endoscopic images to improve image segmentation performance. Experiments were performed on two datasets: a combined polyp dataset from several colonoscopy datasets and a tooth endoscopic dataset. Five existing methods were compared. Improvement was measured indirectly through the improved accuracy of the two image segmentation tasks.
- 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.
Strengths:
The paper is well-written with the preliminary to explain the rationale for the proposed design.
- Several ablation studies were presented (e.g., impact of specular and diffuse detection components), DICE or BCE loss, and dilate kernel size, which leads to improvement.
- Two different endoscopic datasets were tested.
- Convincing qualitative results were shown.
- Quantitative results show improvement of the proposed technique over EndoSRR.
- 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.
Results:
- The performance difference between the proposed method and the second-best-performing model (EndoSRR) is very small. Statistical tests should be done to see whether the improvement is statistically significant, especially on the tooth endoscopic dataset with only 195 images.
Experimental Design:
- Lack of clarity on whether the compared methods were trained using any part of the training dataset used in this work or not. Were the original authors’ source code used to train their model on the same training dataset used in this study? The original StableDelight, M2-Net, TSHRNet, and DHAN-SHR were not trained on endoscopic images. Only EndoSRR was trained on the CVC-Clinic-DB which is part of the dataset used in this work. It would be better to separate models that have any components trained on endoscopic images from the one that does not.
- What was the kernel size and threshold values for the performance in Table 1?
Writing is great with some minor issues.
- INP() is defined with two input parameters, but only one input was used in Eq. 2.
- Stack() used in Eq. 2 is not defined.
- For DILATE(), it is not clear what structuring element of the dilation operation was used.
- t+ in Eq. 4 not defined.
- The abstract has a typo in this sentence “Particularly, for zero-shot SAM segmentation of teeth, a significant performance boost is observed after inpainting, with mDice and mIoU increasing from 51.3%/39.3% to 96.1%/93.0%.” Table 1 does not contain 51.3%, but 51.5% for image segmentation on original images.
- Please rate the clarity and organization of this paper
Good
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
Reproduceability: Since several pre-trained models were used in this work, sharing of code that links to specific pretrained models used to generate the results would help with reproducibility.
- 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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The main issue is whether there is a statistically significant performance improvement of the proposed method over the EndoSRR method that was trained on other endoscopic datasets. Other compared models were not originally trained on endoscopic datasets. Hence, their lower performance compared to EndoSRR and the proposed method is expected.
On the tooth dataset, EndoSRR provided much improvement over TSHRNet, M2-Net, DHAN-SHR, as well as the original image. The improvement of the proposed approach over EndoSRR is less than 1 point and the number of images on this dataset is only 195 images.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper presents a novel two-stage framework for endoscopic artifact inpainting, specifically addressing the challenges of non-Lambertian reflectance in endoscopic imaging. The key innovation lies in its hierarchical approach: first suppressing high-intensity specular artifacts through precise localization and inpainting of overexposed regions to recover underlying tissue textures, followed by adaptive intensity-guided blending to mitigate residual diffuse artifacts while preserving critical anatomical details. The authors rigorously evaluate their method through extensive experiments, including fully supervised learning and zero-shot adaptation across multiple endoscopic datasets. The results convincingly demonstrate the framework’s robustness, generalizability, and clinical relevance, with notable improvements in segmentation accuracy, artifact resilience, and adaptability to diverse anatomical structures. This work represents a meaningful advancement in endoscopic image enhancement, with potential implications for improving computer-assisted diagnosis and surgical guidance.
- 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.
- Clinical Relevance & Technical Contribution: Endoscopic artifacts significantly impair examination efficiency and diagnostic accuracy. This paper addresses a critical gap in current research on endoscopic artifact suppression by proposing a clinically grounded solution. The framework demonstrates robust performance across multiple tasks, highlighting its practical utility in real-world medical scenarios.
- Methodological Clarity & Innovation: The paper provides a lucid and well-structured description of the artifact removal algorithm. The two-stage approach—specular artifact suppression followed by adaptive diffuse artifact refinement—is presented with clear technical rigor, enabling reproducibility and further development in the field.
- Comprehensive Validation: The authors validate their method through extensive experiments, including supervised and zero-shot settings, showcasing strong generalizability to diverse anatomical structures and imaging conditions. The results substantiate significant improvements in segmentation accuracy and artifact resilience.
- 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.
- Limited Algorithmic Innovation: While the paper provides clear descriptions of the methodology and experimental results, the technical novelty appears limited. The proposed artifact removal framework relies heavily on existing pre-trained models (LaMa and StableDelight), with only DuckNet being fine-tuned. This raises concerns about the actual contribution of the work beyond the integration of established components.
- Lack of Justification for Model Selection: The paper does not sufficiently justify the choice of key network architectures in the framework. For instance, it remains unclear why DuckNet was selected for specular segmentation over more conventional alternatives (e.g., UNet). A thorough ablation study or comparative analysis would strengthen the rationale behind these design decisions.
- Please rate the clarity and organization of this paper
Good
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
- Methodological Rigor & Reproducibility: The paper provides a clear and well-structured description of the artifact removal algorithm, with thorough experimental design and rigorous result analysis. The methodology is presented in sufficient detail to support reproducibility, and the conclusions are convincingly substantiated by the empirical evidence.
- Clinical Relevance & Practical Impact: The study addresses a clinically significant challenge by proposing a solution tightly aligned with real-world endoscopic imaging needs. Its focus on resolving pressing issues in medical applications enhances its potential for adoption and translational impact.
- Reviewer confidence
Very confident (4)
- [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.
In their rebuttal, the authors further elaborated on the contributions of their work.
Review #3
- Please describe the contribution of the paper
A two-stage endoscopic artifact inpainting framework to address the challenges of non-Lambertian reflectance. The framework first suppresses high-intensity specular artifacts by locating and inpainting overexposed regions, restoring underlying tissue textures. It then addresses residual diffuse artifacts by adaptively blending intensity-guided refinements, harmonizing inconsistent illumination while preserving critical anatomical features.
- 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.
A novel application and a particularly strong evaluation
- 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.
- In Section 3.1 (Dataset), a table summarizing dataset details, including the data split, should be provided. The authors should acknowledge the critical role of dataset size and distribution in the proposed task.
- In Section 3.1 (Implementation), the selection criteria for hyperparameters should be clearly justified, and supporting results should be provided.
- Cross-validation is required. Additionally, accuracy and loss plots should be provided to enhance the reliability of the results.
- At the end of Section 3, the limitations of the proposed approach should be discussed. If possible, a comparative analysis with existing literature or state-of-the-art methods should be included.
- Please rate the clarity and organization of this paper
Good
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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/2025/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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
80/100
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We sincerely appreciate the reviewers’ constructive feedback and recognition of the clinical relevance, methodological clarity, and strong evaluation of our proposed framework. This rebuttal mainly addresses concerns regarding the significance of performance improvements and the innovation of our algorithmic design. Upon acceptance, code and datasets will be released. ➤ R1 1.Dataset Details and Split Clarity We thank the reviewer for the constructive feedback and apologize for the initial confusion. Additional details regarding the datasets will be provided in the revision to enhance clarity, including the data split (train:val:test=8:1:1). 2.Hyperparameter Justification The Dice-to-BCE loss ratio was set to 1:1 based on prior work, as well as its stable performance during experiments. For all other hyperparameters related to DUCKNet and LaMa, we follow the official implementations for consistency and reproducibility. These will be clarified in the updated version. 3.Cross-Validation Our method achieved 92.1% mIoU on CVC-ClinicDB with 3-fold cross-validation, a 1.4% gain over EndoSRR (90.7%), consistent with prior results. Regarding training curves, they show consistently decreasing loss and increasing validation mDice without significant overfitting. 4.Discussion on Limitations We will revise the end of Section 3 to include a more detailed discussion of current limitations, including the adaptability of the model to multi-task scenarios. ➤ R2 1.Statistical Significance of Improvements We thank the reviewer for the insightful comments. Our method outperforms on colonoscopy and zero-shot dental endoscopy tasks, with statistically significant improvements over the second-best EndoSRR (p < 0.05, Wilcoxon test). For example, on Teeth (Text), p-values for mDice and mIoU are 0.0007 and 0.0005. On Kvasir: 0.0031 (mDice), 0.003 (mIoU); on ETIS: both p < 0.0001, highlighting our clear performance advantage. 2.Training Protocol of Baselines We clarify that the core inpainting module in our method is not finetuned, consistent with prior works. Only the specular reflection segmentation module is trained on CVC-ClinicDB to generate masks as localization cues. No additional endoscopic data is used beyond what is disclosed. We will revise the comparison table to clearly indicate this. 3.Kernel Size and Threshold Values We apologize for the confusion. Our method uses a kernel size of 15 and a threshold of 150, selected based on ablation results in Table 4 and 5. We will clarify this in the revision.
- Minor issues and Typos We greatly appreciate the reviewer’s careful attention to detail. In the revision, we will provide clear definitions and correct all inconsistencies to improve clarity. 5.Small Dental Dataset Despite our small dental dataset (195 images), it serves solely as a zero-shot evaluation set to assess generalization across diverse lighting and tooth types. Our method still achieves statistically significant improvements (e.g., p = 0.0007 for mDice on Teeth (Text)) over competing methods. Additionally, it yields clear gains on polyp segmentation tasks, achieving a 1.8% mIoU improvement over the second-best EndoSRR on ETIS. ➤ R3 1.Algorithmic Innovation We thank the reviewer for noting the clinical relevance. Our main contribution is not merely a combination of existing methods, but a principled, Phong-inspired two-stage framework that explicitly addresses both specular and diffuse artifacts—an area largely overlooked by prior work. Our method systematically addresses both specular and diffuse artifacts: it first restores overexposed regions, then refines diffuse artifacts via weight map–guided blending. This targeted design ensures more accurate artifact removal aligned with endoscopic imaging physics. 2.Choice of DuckNet DUCKNet was selected for its superior segmentation performance, achieving mDice gains of 3.8%, 4.0%, and 2.5% over UACANet-L, CaraNet, and LDNet—stronger U-Net variants. We will clarify this in the revision.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This work has two positive reviewers and a negative reviewer. By checking the comments and the rebuttals, I think this work has more advantages. Hence, I recommend to accept this work.
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A