List of Papers Browse by Subject Areas Author List
Abstract
Real-time, text-promptable 4D reconstruction is indispensable for autonomous surgical interaction. Severe misalignment between semantic meaning and physical anatomy still persists, largely because existing solutions integrate Vision–Language Models into deformable fields via a rigid coupling scheme that tightly binds semantic features to geometric warping. In this paper, we propose DeGenseGS, Geometrically and Semantically Decoupled Surgical Scene Understanding in 4D Gaussian Splatting, a novel framework that independently models semantic evolution and geometric deformation. Specifically, we propose a HexPlane-based spatiotemporal entanglement module that uses shared kinematic latents to synchronize semantic mutations with scene dynamics, while explicitly disentangling semantic updates from geometric deformation. To further ensure robustness against reconstruction artifacts, we devise a Rasterization-Native Semantic Extraction mechanism that infers semantics from topologically continuous feature maps. Additionally, we incorporate an angular-aligned optimization strategy that conforms to the native hyperspherical latent space, thereby preventing semantic distortion. Extensive evaluations on the CholecSeg8k and EndoVis18 datasets demonstrate that DeGenseGS achieves state-of-the-art performance. Our framework yields enhanced geometric completeness and robust semantic-anatomic alignment, enabling spatially continuous segmentation despite drastic tissue deformation and topological transitions.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1749_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/yimowwwww/DeGenseGS
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WanYim_DeGenseGS_MICCAI2026,
author = { Wang, Yimo AND Kang, Bin AND Yang, Shuojue AND Jin, Yueming},
title = { { DeGenseGS: Geometrically and Semantically Decoupled Surgical Scene Understanding in 4D Gaussian Splatting } },
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 presents DeGenseGS, a text-promptable 4D Gaussian Splatting framework that explicitly decouples geometric deformation and semantic evolution for dynamic surgical scene understanding. Specifically, it introduces a Kinematics-Conditioned Latent Disentanglement module built on HexPlane, which leverages shared kinematic latents to separate geometric and semantic optimization branches, thereby eliminating spurious geometric warping caused by dominant semantic gradients. In addition, the paper proposes a Rasterization-Native Semantic Extraction mechanism that directly extracts semantics from 2D rasterized feature maps with topological continuity and under the structural guidance of RGB images, generating robust and hole-free segmentation masks. To enhance semantic stability in highly deformable scenes, it also introduces Angular-Aligned Semantic Distillation to regularize the vision-language model latent space in the hypersphere domain, effectively mitigating feature collapse. Experiments demonstrate that DeGenseGS achieves state-of-the-art performance on the CholecSeg8k and EndoVis18 datasets, with clear advantages in handling severe tissue deformation and topological changes such as surgical cauterization.
- 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: (1) Existing 4DGS methods tightly couple semantics and geometry in a single deformation branch, leading to severe gradient interference and false geometric warping during semantic mutation. DeGenseGS explicitly disentangle these two processes in 4DGS for surgery. It uses shared kinematic latents to enable independent geometric and semantic decoding, which is a novel paradigm that fundamentally solves the misalignment problem. (2) Instead of extracting semantics from 3D Gaussians directly, this work proposes to infer semantics from topologically continuous 2D rasterized feature maps, combining superpixel aggregation and edge-preserving filtering. This design is original for 4DGS and effectively improves boundary accuracy and robustness against reconstruction artifacts, which is critical for clinical applications. (3) The paper formulates semantic distillation as angular similarity learning in the native hyperspherical latent space, rather than simple L2 loss. This is a novel optimization strategy for open-vocabulary 4D reconstruction, which effectively prevents semantic distortion and feature collapse in dynamic scenes. (4) Extensive comparisons against SOTA methods (LangSplat, OpenGaussian, SurgTPGS, etc. ). Validated on two standard surgical benchmarks: CholecSeg8k and EndoVis18.Clear ablation studies verifying each core component. Quantitative gains are substantial: +14.74% mIoU over the best baseline on CholecSeg8k. Qualitative results show consistent improvement in deformed and topological-changing regions.
- 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.
Weaknesses: (1) The abstract claims “real-time” reconstruction, but the paper does not report FPS, inference time, or rendering speed. Readers cannot verify whether the method meets the real-time requirement for intraoperative usage. (2) The paper only qualitatively explains that decoupling reduces gradient interference, but lacks quantitative gradient analysis, e. g. , gradient norm comparison between coupled and decoupled modes. The theoretical justification for why semantic gradients no longer distort geometry is not fully supported. (3) The paper mentions “Open-Vocabulary” and “Text-Promptable” capabilities. However, the evaluation (Tables 1 and 2) focuses on standard segmentation classes (Gallbladder, Instruments, etc. ). There is a lack of qualitative results showing the system’s performance on truly unseen or arbitrary text prompts. If the evaluation is restricted to pre-defined classes present in the dataset, the “Open-Vocabulary” aspect is not fully demonstrated beyond standard semantic segmentation. (4) The authors have not released the source code. For the better development of the community, the authors are encouraged to make their code publicly available, which will also facilitate the reproduction of the proposed method and fair comparisons among peers.
- 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.
- 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?
This paper proposes DeGenseGS for 4D surgical scene understanding, which novely decouples geometry and semantics to avoid false warping. It has solid technical designs including a HexPlane-based disentanglement module, rasterization-native extraction, and angular-aligned distillation. Experiments show significant performance gains over SOTA methods on standard surgical benchmarks with convincing ablation studies. However, it overclaims open-vocabulary ability, lacks real-time speed verification, has insufficient theoretical analysis, and does not release code. Overall, the work is innovative and effective with clear clinical value, so I give weakly accept.
- 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.
Based on the authors’ rebuttal, the concerns have been addressed. Acceptance is recommended. However, it is still suggested that the authors consider open-sourcing their work.
Review #2
- Please describe the contribution of the paper
This paper introduces DeGenseGS, a 4D Gaussian Splatting framework designed for surgical scene understanding, with the central methodology stemming from decoupling geometric deformation from semantic evolution. For this aim, the work introduces the Rasterization-Native Semantic Extraction mechanism and the Manifold-Aligned Optimization strategy, demonstrating the effectiveness of the proposed method.
- 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.
-Interesting point of view: The work identifies a meaningful issue in dynamic surgical scenes: semantic changes do not necessarily imply geometric deformation. The example illustrated in Fig. 1 clearly demonstrates this mismatch and provides an intuitive motivation. -Clear empirical improvements over the strongest baseline: On the CholecSeg8k/EndoVis18 datasets, the method consistently demonstrates meaningful improvement.
- 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 central “decoupling” claim is intuitively reasonable but insufficiently validated. Specifically, in Table 3, the authors define S-G Decoupling as the combination of Section 2.2 (the actual ‘decoupling’ idea) and Section 2.4 (RNSE post-processing, if I understand properly), and ablate both jointly — not sure how much improvement can be attributed to the core decoupling idea. -The method description is unclear and a bit incomplete at the implementation level, especially for the textural/semantic feature parts, and the total losses to train the Gaussian (the pipeline shows rendered RGB and depth, but does not refer to how they are specifically used). If i understand correctly, the semantic attribute per gaussian is dynamically predicted at each timestep? and the loss for training the gaussian attributes contains rgb rendering err and depth rendering err or? A revision of the methodology presentation can help the readers to better understand the method. -A bit overclaim(?): 1)The title identifies the work to address ‘surgical scene understanding’, while from the content, only prompt-instructed-segmentation is considered as the downstream task for the Gaussians. 2) The performance of OpenGaussian is insanely low—would be good to include some discussion or more background regarding how these baselines are set up.
-Others: obvious mistakes in Tab 1 and Tab 2 of the best bolded values. Reference 23 is wrongly placed for SurgTPGS. Pipeline figure quality should be improved(e.g., the way to plot a Gaussian in the attributes ).
- 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.
- 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?
As mentioned in the weakness: 1)The ablation combines the actual decoupling module with RNSE, making it unclear how much of the improvement truly comes from the proposed disentanglement. 2)The methodology description is somewhat incomplete—particularly regarding the semantic feature representation and the training losses—making it difficult to clearly understand the optimization pipeline.
- Reviewer confidence
Somewhat confident (2)
- [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 reviewers’ concerns have been largely addressed. While the reviewer still highly suggests revising the framing as a segmentation method rather than the much broader ‘scene understanding’ at its current state.
Review #3
- Please describe the contribution of the paper
This paper proposes DeGenseGS, a 4D Gaussian Splatting framework that introduces semantic understanding into a dynamic surgical scene reconstruction pipeline and ultimately enables semantic segmentation.
- 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.Problem to solve is meaningful in surgical scenes, where semantic change and geometric deformation are often not well aligned. 2.The reported results are promising, with clear improvements over prior baselines, and the ablation supports the effectiveness of the main design.
- 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 contribution is presented somewhat diffusely. The paper discusses dynamic reconstruction, 3D semantics, and text-promptable understanding, while the final evaluation is still mainly based on 2D segmentation results. The authors should better highlight the paper’s own core contribution and, in particular, clarify the actual difference from SurgTPGS. 2.The paper claims that the method can avoid spurious geometric displacement, but does not provide basic reconstruction comparisons to directly support this point. If improved geometric stability is part of the contribution, it should be evaluated more explicitly. 3.Important practical statistics are missing: runtime-related results such as FPS should be reported since the authors mentioned ‘real-time’ is important.
- 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
1.The current 2D datasets are not sufficient to fully evaluate 3D semantics, and they also limit the demonstration of the method’s full capability. It would be better to include dedicated 3D evaluation in the future, for example using simulated or ex vivo data. 2.The “first” claim could be worded more carefully. My concern is not necessarily overclaiming, but that the statement is qualified by too many conditions, making the novelty positioning less clear.
- 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?
This paper addresses a meaningful problem in dynamic surgical scene understanding and presents a reasonably method that gains over strong baselines. Although I still have some concerns about the novelty positioning and the lack of more direct reconstruction/runtime evaluation, I think the overall contribution is solid and relevant for acceptance.
- 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.
Thank you for the response. The response addresses my main concerns. The clarification of the difference from SurgTPGS and the reported runtime numbers make the contribution and practicality clearer. I still think direct 3D/4D geometric evaluation would better support the claim of improved geometric stability, but this does not change my overall assessment. I am satisfied with the rebuttal and maintain my recommendation.
Author Feedback
We thank reviewers and are encouraged by your positive comments on novel paradigm, meaningful motivation, and substantial gains. Below we address main concerns. Code will be released upon acceptance. 1.Running time. (R1, R3) DeGenseGS renders at 70 FPS; full text-prompted querying with VLM matching and RNSE takes ~0.56s/frame, on par with SurgTPGS[15] at 67 FPS and 0.6s/frame. Will be reported. 2.Task definition. (R1) We agree that Open-Vocabulary is inappropriate and will rephrase it to Text-Promptable throughout. Following SurgTPGS, all baselines use the same CLIP-style prompts on standard surgical classes for fair comparison. Unseen prompts will be evaluated in the extension. 3.Theoretical justification of decoupling. (R1) Eq. 2 decodes ΔG and Δf_sem via independent branches. By the chain rule, the two gradients do not interfere, resolving the gradient-interference issue of PCGrad[21] architecturally. It preserves only motion-consistent components as a shared inductive bias. Formal proof will be added in Sec. 2.2.4.Ablation of KCLD and RNSE. (R2) Following the reviewer, we report the individual effect of each module. KCLD alone gives 64.19/41.02 on CholecSeg8k/EndoVis18 and RNSE alone gives 56.84/40.54, both over the baseline in Tab. 3 row 1.KCLD contributes a larger gain via gradient-level decoupling, while RNSE corrects 2D readout artifacts. They are complementary and the full model (row 4) is best. This will be added to Tab. 3.5.Method description. (R2) The semantic attribute is dynamically predicted per timestep: each Gaussian stores a base semantic feature, and the semantic branch predicts a per-timestamp residual that is alpha-blended into the rendered semantic map. For training, the rendered RGB and stereo depth errors supervise the geometry branch, while semantic alignment and smoothness losses supervise the semantic branch only, acting on disjoint parameters and realizing decoupling at the loss level. We will clarify the methodology accordingly. 6.Scene understanding framing. (R2) Our method is essentially an interactive segmentation framework: beyond text prompts, it can also segment the scene without any prompt and output masks with corresponding instances, which is itself a form of scene understanding. We will adjust the claim accordingly. 7.OpenGaussian baseline. (R2) Designed for static 3D scenes, its instance-level semantic association is fragile under deformation, occlusion, and topology change, causing unstable labels and near-zero masks. We followed official settings and tuned key hyperparameters; similar low results are reported in SurgTPGS on dynamic data. Setup details will be added. 8.Minor fixes. (R2) All noted issues, including bolded values in Tab. 1/2, the misplaced Ref. 23, and pipeline figure quality, will be fixed. 9.Difference from SurgTPGS. (R3) Our core contribution is geometry-semantic decoupling in dynamic 4DGS. SurgTPGS tracks semantic features with deformation, so semantic change may still be coupled with geometric motion. We instead share kinematic conditioning but use separate geometry and semantic branches, so appearance changes like cauterization do not cause false deformation. Theoretical analysis: see #3.10.Reconstruction comparison. (R3) DeGenseGS gains +1.06 PSNR, +0.030 SSIM, −0.0010 LPIPS over SurgTPGS, while segmentation improves much more, +14.74 mIoU. With ablation and qualitative results, the gain mainly comes from geometry-semantic decoupling rather than RGB reconstruction. 11.Segmentation evaluation. (R3) We follow the task setting of SurgTPGS and related work for fair comparison. The 2D segmentation is performed on views rendered from 4DGS, and is the prerequisite for downstream tasks such as instrument tracking and VQA. Currently, semantics for dynamic 3D reconstruction are mainly evaluated on 2D datasets, as no 3D semantic ground truth is available. Dedicated 3D evaluation on simulated or ex vivo data will be added in the extension.
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.
While reviewers agree on the overall contribution, they question on ablation performed, lack of direct reconstruction comparison in 4D/3D sense and runtime evaluation.
- 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.
Authors addressed the reviewers concerns and all reviewers suggested for acceptance. It is suggested that authors maintain reproducibility through open source code and datasets. Also, R2 suggestion on revising the framing as a segmentation method rather than scene understanding.
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.
All reviewers suggest the acceptance of this paper as the rebuttal addressed their main concerns.
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 reviewing the rebuttal letter and reviewers’ post-rebuttal comments, I recommend the paper for acceptance. The authors are suggested to open-sourcing thier work and perform the 3D/4D geometric evaluation before the camera-ready submission.