List of Papers Browse by Subject Areas Author List
Abstract
Despite recent progress in 4D Gaussian Splatting (4DGS), dynamic surgical reconstruction remains challenging due to complex tool-tissue interactions, non-rigid deformations, and wet-surface specularities. Existing methods largely rely on unconstrained visual optimization, ignoring the physical reality of the scene reconstruction. We introduce \textit{PhysioSplat}, a novel framework that advances 4DGS by grounding surgical scene modeling in a unified, three-fold physical model. \textbf{First}, we incorporate \textit{kinematic physics} by explicitly disentangling the scene into two disjoint fields: a tool field governed by rigid or articulated motion, and a tissue field driven by a deformation model. Supervised by semantic masks, this separation enables selective tool removal and the temporally consistent recovery of occluded anatomy. \textbf{Second}, we enforce \textit{biomechanical physics} through a latent, differentiable regularization model. By predicting per-Gaussian stiffness parameters, it minimizes deformation energy and applies collision-aware repulsion to ensure biologically plausible tissue dynamics during tool interactions. \textbf{Third}, we integrate \textit{optical physics} via a specular-aware appearance model. Leveraging the camera-coincident light source common in endoscopy, we mathematically decompose Gaussian radiance into diffuse and view-dependent specular components, successfully stabilizing highlight reconstruction. Together, \textit{PhysioSplat} shifts surgical 4DGS from purely visual reconstruction toward anatomy-aware, physics-constrained modeling.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2002_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/2002_supp.zip
Link to the Code Repository
https://github.com/hritam-98/physiosplat
Link to the Dataset(s)
N/A
BibTex
@InProceedings{BasHri_PhysioSplat_MICCAI2026,
author = { Basak, Hritam AND Yin, Zhaozheng},
title = { { PhysioSplat: Physics-Informed Dynamic Gaussian Splatting for Surgical Scene Reconstruction } },
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 introduces PhysioSplat, a novel framework designed to address the fundamental challenges of dynamic surgical scene reconstruction, such as endoscopic video reconstruction. By integrating physical laws into 4D Gaussian Splatting (4DGS), the method shifts from traditional unconstrained visual optimization toward a physics-grounded modeling approach. The main contributions of this work are summarized as follows:
1.Physio-Semantic Gaussian Disentanglement To address the intrinsic differences in motion characteristics between surgical tools and soft tissues, PhysioSplat introduces a dual-branch architecture that explicitly partitions the scene into a Tool Field and a Tissue Field. Departing from previous methods that treat the scene as a monolithic deformation field, this approach—supervised by semantic masks—governs tools via rigid or articulated motion models while driving tissues through an elastic deformation model. This kinematic separation eliminates geometric shearing artifacts at tool-tissue interfaces and enables the temporally consistent recovery of underlying anatomy occluded by tools through the selective rendering of the tissue field.
2.Differentiable Biomechanical Regularization To constrain the deformation field to generate biologically plausible dynamics, PhysioSplat proposes a suite of differentiable biomechanical loss functions. By approximating tissue as a discrete mass-spring system, it enforces three key physical constraints: (1) Local Elastic Strain Energy (based on the ARAP formulation) to preserve tissue topology and local rigidity;(2) Tool-Tissue Collision Repulsion, which treats the tool volume as a proxy Signed Distance Field (SDF) to exert repulsive forces on tissue Gaussians penetrating the tool boundary, thereby simulating realistic physical interactions;(3) Volume Preservation to preclude non-physical compression or collapse. Together, these constraints ensure quasi-static elasticity and physical realism during dynamic tool-tissue interactions.
3.Specular-Aware Appearance Modeling Addressing the high specularities common to wet anatomical surfaces in endoscopic environments, PhysioSplat introduces a specialized specular-aware appearance model. Leveraging the physical constraint of co-located camera and light sources in endoscopy, the framework decomposes Gaussian radiance into view-independent diffuse and view-dependent specular components (modeled via Gaussian lobes). This design forces specular highlights to remain strictly consistent with the underlying geometry, effectively suppressing the “floater” artifacts common in traditional methods and enhancing geometric reconstruction precision in reflective environments.
- 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.
Kinematic Disentanglement: The shift from a monolithic deformation field to a disentangled architecture—partitioning the scene into a rigid tool field and an elastic tissue field—is highly effective. This approach successfully resolves the blurring and shearing artifacts typically found at the interface of hard instruments and soft tissue. Moreover, its ability to clearly reconstruct the underlying anatomy even when occluded by surgical tools is of significant clinical importance.
Robust Biomechanical Grounding: By introducing Differentiable Biomechanical Regularization, the framework achieves high-quality tissue reconstruction that exceeds the capabilities of pure visual optimization. Explicitly modeling collision and deformation provides reliable physical and geometric constraints, which directly result in superior and more credible reconstruction outcomes.
Realistic Optical Modeling: The strategy for optical modeling is particularly distinctive. By accounting for the dynamic nature of specular highlights and decomposing light reflectance—combined with surface normal derivation from rendered depth gradients—the method stabilizes the reconstruction of highly reflective surfaces. This approach aligns much more closely with the complexities of actual clinical environments, such as fluid-covered anatomical surfaces.
- 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.
While the methodology of this paper is quite compelling, I have several concerns regarding specific technical aspects: 1.Sensitivity to Semantic Mask Accuracy: The disentanglement process relies on input semantic tool masks to partition the scene. However, in practical surgical environments, acquiring high-precision tool masks in real-time remains a significant challenge. I am concerned about the framework’s tolerance for mask noise. If the masks contain boundary errors or misclassifications, would this lead to more severe artifacts in the reconstructed fields? 2.Residual “Ghosting” and Blurry Artifacts: While the reconstruction of occluded regions shows visual improvement, the supplementary materials still reveal noticeable residual “ghosting” or blurry shadows after tool removal. Curiously, these artifacts even appear to affect regions that were not occluded in the original input video. Could the authors clarify the root cause of this? Is this a fundamental limitation of the proposed disentanglement architecture? 3.Circular Dependency and Instability in Normal Estimation: The derivation of surface normals from the spatial gradients of the rendered depth map introduces a potential circular dependency. During the early stages of training, these gradients are often noisy and unstable, which can disrupt specular highlight rendering and lead to photometric errors. Since the optimization minimizes photometric loss by adjusting Gaussian positions, the system might inadvertently warp the underlying geometry to compensate for incorrect highlight distributions, resulting in non-physical “wrinkles” or geometric noise. I would appreciate an explanation of how the training process specifically mitigates this instability. 4.Kinematic Discontinuity in Binary Motion Switching: The hybrid motion model utilizes a binary threshold to bifurcate Gaussians into either the rigid tool field or the deformable tissue field. This hard-switching mechanism may fail to ensure kinematic continuity at the tool-tissue interface. Gaussians located at the contact boundary might experience step-wise jumps in motion, potentially leading to geometric tearing or unnatural seams. This likely contributes to the edge-blurring issues mentioned in my second point. Furthermore, because Gaussians cannot transition freely between fields once assigned, the system appears to be highly dependent on the accuracy of initialization. I believe the impact of this initialization on the final reconstruction fidelity warrants further investigation.
- 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
Please report the training speed and FPS in the inference.
- 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?
This paper successfully addresses key pain points in surgical reconstruction by incorporating physical priors into 4D Gaussian Splatting (4DGS), demonstrating superior performance. The proposed dual-branch Physio-Semantic Disentanglement architecture is highly aligned with clinical reality. By disentangling rigid tool motion from elastic tissue deformation, it not only eliminates geometric shearing artifacts at tool-tissue interfaces but also realizes a clinically valuable “Physio-Inpainting” capability for the temporally consistent recovery of occluded anatomy.
- 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 rebuttal addressed my concern. Therefore, I will maintain my rating.
Review #2
- Please describe the contribution of the paper
This paper proposes PhysioSplat, a dynamic 4D Gaussian Splatting (4DGS) framework for surgical scene reconstruction that incorporates three categories of physical constraints. The first contribution is a Physio-Semantic Gaussian Disentanglement (PSGD) module that partitions the Gaussian primitives into a rigid tool field and a deformable tissue field via a learnable semantic logit, enabling selective tool removal and occluded anatomy recovery. The second contribution is a Differentiable Biomechanical Regularization (DBR) that combines an As-Rigid-As-Possible (ARAP) strain energy term, a tool-tissue collision repulsion potential based on an approximate SDF of the tool volume, and a volume preservation constraint enforced on the local deformation gradient. The third is a Specular-Aware Appearance Model (SAAM) that decomposes Gaussian radiance into diffuse (SH degree-0) and view-dependent specular components under the co-located camera-light assumption common in endoscopy, with surface normals estimated on-the-fly from the rendered depth gradient. The method is evaluated on three benchmarks (EndoNeRF, StereoMIS, SCARED) against eight baseline methods, and achieves top PSNR/SSIM/LPIPS scores across all datasets. An ablation study and modular plug-in generalization experiments are also provided.
- 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.Clinically motivated problem decomposition. The explicit separation of rigid tool kinematics from deformable tissue dynamics is well-motivated by the actual physical structure of surgical scenes. This disentanglement is more principled than treating the whole scene as a monolithic deformation field, and the resulting tool-removal capability has clinical relevance. 2.Modular design with demonstrated generalizability. Figure 3 shows that the DBR and SAAM modules can be integrated as drop-in regularizers into four existing SOTA methods (EndoGaussian, SurgicalGaussian, Deform3DGS, D⁴Recon), all yielding consistent improvements. This architecture-agnostic property is a practical strength. 3.Comprehensive experimental coverage. The paper evaluates across three distinct benchmarks with different imaging conditions and surgical procedures, covering both photometric (PSNR, SSIM) and perceptual (LPIPS) metrics. The ablation table systematically builds up the full model from a shared baseline.
- 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 paper positions PhysioSplat as a framework that grounds surgical reconstruction in a “unified, three-fold physical model,” covering kinematic, biomechanical, and optical physics. However, upon closer examination, the three components correspond respectively to a semantically supervised dual-branch motion split, a combination of ARAP strain energy, collision penalty, and Jacobian determinant regularization, and a co-located specular decomposition — all of which are established physics-inspired regularization techniques rather than physics-consistent simulation in the strict sense (e.g., differentiable FEM/MPM solvers or PINN-style PDE residuals). The authors are encouraged to clarify: in what sense does the combination of these three components constitute a “physics-constrained model” rather than a “physics-regularized” one?
2.Section 2.2 adopts a fixed threshold τ = 0.5 to partition Gaussian primitives into the tool field and the tissue field. It is unclear how this value was determined and whether the method is sensitive to its choice. More importantly, at tool-tissue boundaries involving blood, smoke, or specular highlights, would a hard step-function partition of a continuous semantic logit space risk introducing boundary instabilities or artifacts — the very shearing effects the method aims to resolve? Is the choice of a hard binary split, as opposed to a soft probability-weighted compositing strategy, sufficiently motivated? Does the sensitivity of τ warrant further discussion?
3.In Section 2.4, the surface normal nᵢ is derived on-the-fly from the spatial gradient of the rendered depth map (nᵢ ≈ ∇D_rend), which is itself produced by the Gaussian geometry currently being optimized. This raises a concern: during early training stages when the rendered depth map is noisy, would differentiation through this estimate inject unreliable gradients into the specular loss and consequently perturb the geometric optimization? How does the method ensure that this feedback loop does not amplify early-stage errors rather than stabilize them? The claim that this design “heavily stabilizes the optimization in fluid-filled regions” appears to be made without direct empirical verification — can the authors provide evidence supporting this?
4.Section 2.5 states that the photometric loss is computed against “the ground-truth frame Iₜ (without the tool).” In real surgical video, obtaining a tool-free frame at the same timestep and viewpoint is generally not feasible. Is the actual implementation using a masked photometric loss that excludes tool-occluded pixels, rather than a truly tool-free ground-truth image?
5.The central thesis of PhysioSplat concerns geometric and physical plausibility — eliminating shearing at tool-tissue interfaces, suppressing volume collapse, and enforcing biomechanically realistic deformations. Yet the evaluation relies entirely on PSNR, SSIM, and LPIPS, with LPIPS described in the paper as a measure of “geometric consistency.” Is LPIPS an appropriate proxy for geometric accuracy? The SCARED dataset provides structured-light ground-truth geometry — why are no surface-level geometric metrics reported on this benchmark? How do the authors propose to substantiate claims about biomechanical plausibility (e.g., volume preservation, inter-penetration reduction) through the current evaluation protocol?
6.The paper raises a number of questions regarding reproducibility. First, Eq. (1) introduces rigid tool motion parameters R_tool(t) and T_tool(t) as the foundation of the PSGD module, but their estimation procedure is never described — are these parameters obtained through joint visual optimization, an external tracker, or per-frame fitting? For articulated surgical instruments with multiple degrees of freedom, is a single global rigid transform sufficient? Second, the “gradient reweighting based on temporal consistency” applied to occluded tissue Gaussians is described only in words without any formal specification — how is temporal consistency defined and measured in this context? Third, the ablation in Table 2 evaluates only cumulative combinations; is it possible to assess the independent contribution of DBR and SAAM in isolation? Would the omission of these conditions affect the conclusions drawn about each module?
- 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 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?
PhysioSplat tackles a well-motivated problem in surgical scene reconstruction, and its consistent quantitative improvements across three benchmarks are encouraging. The modular design — particularly the semantic disentanglement and the specular decomposition — addresses real challenges in endoscopic imaging, and the plug-in generalization results in Figure 3 suggest the components have broader applicability beyond the proposed framework. However, several concerns collectively justify a weak reject at this stage. The framing of the method as a “physics-constrained model” is not sufficiently justified relative to the actual components employed, which are better characterized as physics-inspired regularizers; the authors have not explained what theoretical grounds support the stronger claim. The hard semantic threshold τ = 0.5 is adopted without justification or sensitivity analysis, leaving open the question of whether it introduces the very boundary artifacts the method aims to eliminate. The online normal estimation in the specular module creates a potential feedback loop whose stability is asserted but not demonstrated. The photometric supervision in Section 2.5 is described in a way that implies tool-free ground-truth frames, which raises questions about the experimental setup that need to be clarified. Finally, the paper’s core claims are about geometric and biomechanical improvement, yet no geometric metrics are reported even where ground-truth geometry is available (SCARED), and the ablation does not allow independent assessment of each module. None of these are insurmountable. A rebuttal that clarifies the supervision protocol, addresses the normal estimation stability concern with empirical evidence, provides geometric evaluation on SCARED, reconsiders the framing of “physics-informed,” and resolves the reproducibility gaps in Section 2.2 would significantly strengthen the submission and could shift this recommendation toward 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.
The authors have addressed all my concerns.
Review #3
- Please describe the contribution of the paper
This paper proposes PhysioSplat, a Gaussian Splatting based method for surgical scene reconstruction. Specifically, they come up with a dual-branch architecture to separate rigid (tool) and deformable (tissue) components. They also propose a biomechanical regularization to constrain tissue motion. Finally, they account for specular-aware appearance of scenes. Across surgical scene datasets, the proposed method shows state of the are reconstruction results.
- 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 well written and easy to follow. 2.The problem formulation is well motivated. Physical models and regularizers have been shown to help with unconstrained 4D reconstruction problems, and applying it to a surgical setting makes sense. 3.The proposed use case (tool removal) is clinically and algorithmically meaningful. 4.Results and comparisons shown in the paper shows superior performance across evaluated benchmarks and datasets.
- 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.It is not clear what the monocular depth is exactly used for. It seems it is used to estimate normals for Gaussians, but how is it used in the training pipeline, formulation? is the depth used for anything else? 2.It is not clear exactly how exactly the depth and segmentation maps are obtained in practice. which models are used, etc. 3.The biomechanical model seems very heuristic, with a lot of parameters defining exactly how the biomechanical regularizer works. Is there any guidance on the choice of parameters? 4.Is the method truly applicable to articulated tools, or only globally rigid tool motion? Equation (1) appears to use one rigid transform per frame for the full tool field. If articulated tools are supported, please describe the kinematic model and how part-level motion is inferred. 5.How sensitive is the method to mask quality?
- 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.
(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 an innovative method for surgical scene reconstruction using 4DGS, by incorporating relevant physical constraints and regularizers. The method is reasonable, and the results show competitive performance. There remain several open questions (mentioned above) as well as some necessary information that seems missing in the draft. However, assuming the authors are able to respond to these questions satisfactorily, I lean towards accepting this work.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 reviewers(R) for comments(C), finding our work novel & innovative(R1,R3), principled(R2), well-written(R3), clinically motivated(R1,R2,R3), with superior performance(R1,R2,R3):
R1C1 Mask sensitivity:Logit s_i is jointly optimized by photometric & BCE losses(Eq8), yielding soft, correctable gradients instead of hard partition errors. Collision repulsion(Eq5) further preserves tissue geometry. New experiments with ±10px boundary perturbation on EndoNeRF-Cutting show PSNR drops only 0.2dB(42.8→42.6), confirming robustness
R1C2 Ghosting: Cause:occluded ghosting stems from ill-posed inpainting; non-occluded one is from α-blending bias, where overlapping tool Gaussians distort nearby tissue transmittance(1-α_j), baking mixed colors. Mitigation: temporal gradient re-weighting (Sec2.2), densification suppression near tools(Ψ_tool<ε, Sec3), & collision repulsion(Eq5). Limitation: fully decoupling tool shadow bias at tight boundaries remains difficult; future work will explore boundary-aware gradient masking
R1C3 Circular dependency & Normal: Our 5k-iteration warm-up disables specular loss, letting D_rend stabilize before normals n_i~∇D_rend are computed, decoupling geometry from specular supervision in early noisy stages. LPIPS gain in Tab2(Exp.c→d, all dataset) confirm stable convergence
R1C4 Threshold & initialization: Hard split(τ=0.5) is preferred over soft composite as discrete motion models(rigid tools vs. elastic tissue) are necessary to eliminate geometric shearing. The switch occurs only at render time; during training, semantic logits evolve via gradient flow. Boundary tearing is prevented by: collision repulsion(Eq5), enforcing a 2mm gap between fields; & ARAP(Eq4), ensuring smooth deformation gradients. The model is robust to initialization(PSNR varies <0.3dB for τ∈{.3,.5,.7}, as final assignment is convergence-driven. +2.5dB PSNR gain from PSGD(Tab2) confirms interface artifacts are resolved
R1C5 Speed: ~145s/scene, 193 FPS(Tab1), competitive to SoTA, consistently superior reconstruction
R2C1 Physics-constrained: We’ll revise to “physics-regularized”. Our 3 components impose geometry-aware, physically-motivated inductive biases(quasi-static elasticity by ARAP, non-penetration by SDF repulsion, incompressibility by Jacobian determinant) that are stronger than generic smoothness priors. Gains in Tab2 validate these biases yield measurably more physical reconstructions
R2C2 τ: See R1C4
R2C3 Normal: See R1C3
R2C4 Photometric supervision: “Without the tool”=pixel-level masked L1, excluding tool-occluded pixels, not tool-free frames. Will clarify in revision.
R2C5 Geometric accuracy:We agree LPIPS is perceptual. Surface-level geometric metrics on SCARED: D4Recon: Chamfer Distance(CD)=2.41mm, F-score@1mm=0.58; Ours: CD=1.83mm, F-score@1mm=0.71.Our 24%CD decrease shows direct improvement in geometric fidelity
R2C6 Reproducibility: 1.R_tool&T_tool are by per-frame differentiable rigid registration, jointly optimized, no external tracker. 2.Single global rigid transform is empirically sufficient for rigid laparoscopic instruments. 3.Temporal consistency: photometric similarity between an occluded tissue Gaussian & its nearest unoccluded-frame projection, re-weighting densification gradients to prevent pruning. 4.Isolated ablations(will be added) do not change our conclusions, as DBR & SAAM target orthogonal properties(geometric vs. specular), validating distinct roles
R3C1 Depth: for Initializing Gaussian positions & as a geometric prior to anchor the rendered depth from which surface normals are derived
R3C2 Data: EndoNeRF & stereoMIS: depth by stereo-matching; SCARED: depth by structured-light. Masks by SAM, following [2,7]
R3C3 Biomechanical parameters: λ&ε by grid search(Sec3); ζ(mean dist to K=8 neighbors) is dynamic per-primitive
R3C4 Articulated tools: Eq1 applies one rigid transform per tool field per frame, sufficient for instruments. Per-segment kinematic trees are future work
R3C5 Mask: See R1C1
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.
Please justify the physics-constrained model with better insight into fundamentals and clarity on experiments. Please provide more clarity to the experiments conducted by making it more relevant to the claim. Also, the ablation should be done for each module independently.
- 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.
This is a clear accept and reviewers (R1 and R2) thought the authors addressed all their concerns in the rebuttal.
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.
This submission presents PhysioSplat, a physics-informed extension of 4D Gaussian Splatting for surgical scene reconstruction. The method combines rigid tool and deformable tissue decomposition with physics-inspired regularization and a specular-aware appearance model.
The paper received mixed-to-positive initial reviews, reflecting some uncertainty prior to clarification. After rebuttal, all reviewers converged to acceptance, though with varying confidence levels.
Reviewers consistently highlight the well-motivated separation of tool and tissue dynamics, specular-aware appearance modeling, and strong empirical performance across multiple datasets, supported by ablations and modular evaluations.
Key concerns relate to implementation clarity, including supervision signals, depth/normal estimation, and segmentation dependency. Additional concerns include robustness of hard semantic partitioning, sensitivity to design choices, and whether the method is truly physics-constrained versus physics-inspired. Several reviewers also question whether photometric metrics sufficiently support claims about geometric and biomechanical accuracy.
Following rebuttal, all reviewers accepted the paper. Overall, this is a solid contribution to surgical scene reconstruction.
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.
The reviewers highlight the clean and well-motivated disentanglement approach, the clinically meaningful problem addressed by this work, and the consistent performance across various settings. The paper is therefore recommended for acceptance.