List of Papers Browse by Subject Areas Author List
Abstract
Accurate reconstruction of soft tissue is crucial for advancing automation in image-guided robotic surgery. The recent 3D Gaussian Splatting (3DGS) techniques and their variants, 4DGS, achieve high-quality renderings of dynamic surgical scenes in real-time. However, 3D-GS-based methods still struggle in scenarios with varying illumination, such as low light and over-exposure. Training 3D-GS in such extreme light conditions leads to severe optimization problems and devastating rendering quality. To address these challenges, we present Endo-4DGX, a novel reconstruction method with illumination-adaptive Gaussian Splatting designed specifically for endoscopic scenes with uneven lighting. By incorporating illumination embeddings, our method effectively models view-dependent brightness variations. We introduce a region-aware enhancement module to model the sub-area lightness at the Gaussian level and a spatial-aware adjustment module to learn the view-consistent brightness adjustment. With the illumination adaptive design, Endo-4DGX achieves superior rendering performance under both low-light and over-exposure conditions while maintaining geometric accuracy. Additionally, we employ an exposure control loss to restore the appearance from adverse exposure to the normal level for illumination-adaptive optimization. Experimental results demonstrate that Endo-4DGX significantly outperforms combinations of state-of-the-art reconstruction and restoration methods in challenging lighting environments, underscoring its potential to advance robot-assisted surgical applications. Our code is available at https://github.com/lastbasket/Endo-4DGX.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1320_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2025/supp/1320_supp.zip
Link to the Code Repository
https://github.com/lastbasket/Endo-4DGX
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HuaYim_Endo4DGX_MICCAI2025,
author = { Huang, Yiming and Bai, Long and Cui, Beilei and Li, Yanheng and Chen, Tong and Wang, Jie and Wu, Jinlin and Lei, Zhen and Liu, Hongbin and Ren, Hongliang},
title = { { Endo-4DGX: Robust Endoscopic Scene Reconstruction and Illumination Correction with Gaussian Splatting } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15968},
month = {September},
page = {181 -- 191}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper improves endoscopic reconstruction using Gaussian Splatting by adapting exposure and varying illumination. To this end, (i) an illumination embedding is computed (no idea, which one), which is then used to adapt colors of Gaussians, plus another similar network, which is applied after deformation. As a result, reconstruction quality is significantly improved.
- 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.
- important topic
- good results
- 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.
- description partially unclear
- often verbose description, in other places missing details
- not self-contained
- Please rate the clarity and organization of this paper
Poor
- 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
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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The overall idea of the paper is quite simple - allow Gaussians in endoscopic reconstruction to adapt their luminance in order to adapt to varying exposure and lighting conditions. I found it a surprising that the approach does not try to adapt a simple lighting model, but just adapts colors, independent from normals. It is surprising that the Gaussian’s normal is nowhere included in the process. But also several other details remained unclear to me:
- The deformation network in Fig. 2 is not at all described, how does it look like? Is it a hash grid? How is it regularized, trained, parameterized, …?
- According to Fig. 2, the Gaussians and the illumination correction are trained jointly, is this correct? My guess is that it would make sense to turn on certain losses (e.g. deformation or color adjustment) over time, i.e. to first optimize Gaussian positions and sizes, before deformation is optimized, and then the colors are adjusted? How is this done exactly?
- Which “illumination embedding” is used? Something pre-trained, or is this network trained together with the Gaussians (hopefully, not). How large is $k$?
- It is not at all clear to me, why f_region are two different networks, and why this is important. Shouldn’t the training make the adaptation on its own? Why does IC (Equ. 2) require an illumination estimation algorithm, couldn’t one use just the average brightness?
- f_region gets as input the color of a Gaussian and $e$, but not the position of the Gaussian. So it can only adapt the Gaussian’s color globally, but not locally. Is this correct? This would mean, the best f_region can do is to make a global exposure correction, but nothing local. What does “region” mean then, does this not refer to the spatial position?
- The difference between “region-aware” and “spatial-aware” is not clear to me. According to Fig. 2 the first one is applied before deformation, the latter one after deformation, but what is the idea behind this? Also f_spatial does not receive the Gaussian’s position, so how can it work?
- In general, while lots of technical detail is missing, other parts are often overly verbose with fancy terminology, but unclear motivation and implementation.
In summary, although the results look interesting, I found the description too confusing and lacking to be able to judge the magnitude of the contribution and the value of the submission.
- 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.
Even after reading the rebuttal and the other reviews, I see too many things that are unclear to me, so I cannot recommend acceptance.
Review #2
- Please describe the contribution of the paper
This paper presents a novel endoscopic reconstruction method with illumination adaptive Gaussian Splatting. It achieves illumination correction and reconstruction in challenging uneven illumination. Experiments have been conducted on real surgical datasets to demonstrate the effectiveness and robustness.
- 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 proposed illumination embeddings, region-aware enhancement, and spatial-aware adjustment modules enhance Gaussian splatting by enabling effective illumination correction, resulting in high-quality scene reconstruction.
- 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) Since the method relies on a prior correction and illumination classification step, evaluating the performance of this step in isolation is crucial. Without this, it’s unclear whether the strong results come from the proposed framework or largely from the preprocessing module. 2) Table 2 is underexplained in the main text. The paragraph introducing Table 2 is brief and doesn’t fully clarify its purpose. 3) The paper would benefit from an analysis of failure cases to better understand the limitations, particularly under extreme or unmodeled illumination conditions. Additionally, reporting statistical significance would strengthen the performance claims and ensure the reported improvements are robust.
- 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.
- 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?
The paper is in a satisfactory clarity. The main objective is supported with reasonable experiments. There are still issues need to be resolved listed in major weaknesses.
- 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 issues has been addressed.
Review #3
- Please describe the contribution of the paper
This paper presents a method for 4D Gaussian splatting under variable illumination conditions. Given three image sets categorized into dark, bright, and normal illumination conditions, the authors embed illuminations with a focus on region and pixel domains for stable 4DGS optimization. The proposed method is evaluated on an original dataset, which is an extended version of the EndoNeRF dataset, with professional exposure modifications over real-world images. The evaluations are thorough, including comparisons against state-of-the-art approaches and an ablation study.
- 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 well-written and -structured. The authors addressed one of the hardest conditions for 4DGS (i.e., dynamic geometry and illumination). The proposed method is well demonstrated on an artificially modified dataset and two datasets from real-world scenarios with varying illuminations.
- 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.
The major weakness, if I needed to point out, is the fixed three-illumination conditions, which lack flexibility and continuity. Adding discussions in this regard would help follow-up research.
- 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 mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure 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
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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The proposed approach and design choices for one of the hardest vision problems (i.e., dynamic scene reconstruction) are well-validated. Adding some discussions on limitations would further strengthen the paper.
- 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 thank the reviewers for the valuable feedback. We appreciate that the reviewers find our work to be “important topic & interesting” (R1), “well-written & well demonstrated & thorough” (R2), and “novel” (R3). We address the major concerns as follows:
Lightning model with Gaussian’s normal. (R1) While lighting model provide a physically reasonable illumination adjustment, we aim to resolve issues for tone-mapping illumination adjustment. Additionally, we do not adopt the rendering-based methods due to high computation cost, which involves more elements (e.g., normal, albedo, roughness).
Details for deformation network. (R1) We follow the baseline 4DGS [21], which is based on the hex plane representation. We will clarify it in Sec 2.1. The details of losses are presented in the optimization part of Sec. 2.2.
Training of illumination correction network. (R1) We train Gaussians and illumination correction jointly. The first six lines in Table 1 show that methods with separated training of Gaussians and illumination correction require high memory and FPS and have poor performance. Therefore, we aim to build an end-to-end joint training pipeline for Gaussians and illumination correction with less memory and training time.
Details of illumination embedding. (R1) We use the illumination embedding trained together with Gaussians (same with Q3 for reasons of end-to-end), $k$ is set to 32.
Reasons for using dual-branch design and prior illumination estimation. (R1) Based on the Retinex decomposition, we adopt a dual-branch network for f_region. VECNet (ACMMM 2024) uses a similar design, showing the importance of the two-branch design separating over/under exposure. Using average brightness on a dataset with all low-light images will mistakenly classify some (with low brightness but higher than average) as bright images. To prevent this, we adopt prior from the illumination estimation [28].
Explanation of f_region and the difference between f_spatial. (R1) Lowercase c: Color for each Gaussian; Uppercase C: Color in the rendered image. f_region can focus on small sub-3D regions (Gaussian-level) without the position input since it takes per-gaussian color (lowercase c) as input, where each c is related to a position (µ) as in preliminary. The difference between the f_region and f_spatial is defined by output target, where f_region targets at per-gaussian color (lowercase c), and f_spatial targets at pixel-level color (uppercase C). f_spatial corrects the illumination after the deformation (finish rendering), where the uppercase C is from a bunch of Gaussians on the ray direction spatially according to the rendering equation in preliminary.
Fix three illumination conditions. (R2) We use three types of conditions, while inside each degradation type, the exposure values are random (i.e., varying degrees of overexposure). We will clarify it in Section 3.1.
Evaluation of prior correction (R3) The prior estimation is part of the illumination embedding (define which branch of illumination embedding). Without the prior classification step, our model cannot define overexposure and low-light. In Table 3, our ablation study shows that the classified embedding (including the prior classification) is significant, while other modules also contribute to the performance.
Detailed explanation of Table 2 (R3) In Table 2, we compare the rendering results of varying illumination (reconstruct bright and dark without correction), demonstrating the robustness of our method.
Analysis of limitations (R3) We observe the failure cases (large deformation, illumination changes) and find that other baselines perform even poorer than ours. Regarding the statistical significance, given the page limitation, we report the performance based on the setup of previous work in 4DGS as our priority.
We will also fix the minor problems for better clarity. All suggested changes will be added to the paper. The code and dataset will be public upon acceptance.
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’
Despite one skeptical review, this appears to be the level of (at least) miccai poster session
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’
Indeed the paper is somewhat abbreviated as suggested by R1. However, this is reasonable due to the space constraints of the MICCAI submission, and the authors did rebut it successfully with clarification (i.e., details were available from other papers).
If accepted, the clarity of the manuscript needs to be improved further in the camera-ready version.