List of Papers Browse by Subject Areas Author List
Abstract
Optical Coherence Tomography (OCT) is central to retinal disease diagnosis and longitudinal monitoring, yet learning-based OCT analysis is often constrained by limited expert annotations. Conditional generative models can alleviate data scarcity, but prevailing lesionor segmentation-conditioned methods typically lack physical constraints governing the biomechanical coupling between pathological fluid evolution and retinal-layer deformation. As a result, synthesized scans may look plausible while exhibiting anatomically inconsistent layer configurations near lesions. We propose PhysOCT, a physics-guided plug-in framework that augments existing conditional generators with a deformation consistent structural prior. Given consecutive lesion masks, PhysOCT (i) predicts lesion-induced external forces using a force prediction network, then (ii) propagates retinal geometry via a differentiable meshfree solver built on enhanced Generalized Moving Least Squares and a stabilized Lagrangian update, producing a predicted next-step layer prior. This physics-derived prior, together with a deformation-aligned warped sketch of the previous scan, is used as additional conditioning for diffusion-based synthesis. Experiments on multiple longitudinal retinal OCT cohorts demonstrate improved one-step deformation prediction and consistent gains in generation fidelity (PSNR/SSIM) and distributional quality (FID) across DDPM and LDM backbones. The code is publicly available at: https://github.com/Terrific0723/PhysOCT.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1750_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Terrific0723/PhysOCT
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WanYut_PhysOCT_MICCAI2026,
author = { Wang, Yutong AND Li, Xiaohui AND Zhang, Wenbo AND Zhang, Yizhe AND Chen, Qiang},
title = { { PhysOCT: Physics-Guided Retinal Modeling for OCT Image Synthesis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
OCT imaging is fundamental for the diagnosis of retinal diseases, but the availability of labeled training data is limited. Generative models such as diffusion models (e. g. , DDPM) have been used to alleviate this scarcity by synthesizing realistic samples. However, they typically do not enforce physical or anatomical constraints in the generated images.
The paper proposes PhysOCT, a physics-guided conditioning framework for OCT image synthesis. Instead of generating images only from learned distributions, the method conditions the generative model on an anatomically consistent prior obtained by physically deforming a real OCT B-scan. This deformation is driven by lesion evolution, provided as a pair of fluid masks at consecutive time steps, and is computed using a differentiable biomechanical solver. A force prediction network estimates lesion-induced forces from the fluid evolution, which are then used to guide the deformation process.
The resulting deformed B-scan, together with additional conditioning inputs, is used to guide a diffusion model to generate the final image. In this formulation, the physical module enforces structural and anatomical consistency, while the generative module focuses on synthesizing realistic image appearance.
- 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.
S1.The paper addresses an important limitation of OCT synthesis methods: the lack of explicit anatomical and physical constraints. While purely data-driven generative models may implicitly capture these constraints, there is no guarantee that anatomical consistency will be imposed in practice. The proposed method addresses this issue by explicitly enforcing deformation consistency through a physics-guided prior.
S2.The paper proposes an interesting combination of a physics-based deformation model with a deep generative framework. The use of a physically consistent warped OCT B-scan as a conditioning signal is conceptually simple but effective, according to the experiments.
S3.The method leverages the differentiability of the physical solver to learn a mapping from lesion evolution (represented as fluid masks) to the solver inputs in an end-to-end manner. This enables the integration of high-level, image-based signals with a physics-based deformation model. While similar ideas have been explored in differentiable simulation frameworks (e. g. , DiffTaichi), their application to retinal OCT deformation from fluid lesions, together with integration into generative modeling, is interesting and potentially extensible.
S4.The proposed method shows consistent improvements in deformation prediction accuracy and image synthesis quality across different datasets and generative backbones. This suggests that the approach may be broadly applicable.
- 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.
W1.My main concern with this paper is conceptual and relates to the practical usefulness of the proposed approach. While the motivation of improving synthetic OCT generation under limited data is clear, the proposed method relies on existing OCT B-scans and their corresponding fluid annotations as inputs to synthesize future B-scans. This already limits the diversity of the generated data, as it remains inherently tied to the available labeled dataset, a limitation that does not arise in purely data-driven generative models.
This limitation becomes more pronounced as the method additionally requires the fluid mask not only at time t, but also at time t+1 in order to synthesize the future image. In practice, fluid masks are obtained as annotations (or segmentation maps) of OCT images, and it is therefore unclear in which scenarios one would have access to a realistic mask at time t+1 without already having the corresponding OCT scan. If such a scan is already available, the need to estimate the deformation through PhysOCT appears less compelling, as one could directly use the observed image to condition the generative model.
A possible alternative would be to generate plausible future fluid masks. However, in order to meet the constraints assumed by the method, this would itself require modeling the fluid masks under anatomical and physical constraints, which is a non-trivial problem and is not addressed in the paper.
Overall, these limitations suggest that the method may have limited capacity to generate substantially new samples, making its practical impact for data augmentation or training data generation unclear.
W2.While the proposed method shows consistent improvements over baseline generative models in terms of image quality metrics (PSNR, SSIM, FID), it remains unclear how these gains translate into practical benefits for image analysis tasks. In particular, the evaluation does not assess the impact of the generated data on downstream applications, such as segmentation or disease estimation, which are ultimately the relevant use cases in the MICCAI context.
Improvements in image quality and fidelity metrics do not necessarily imply that the images serve as more useful training data or will help to improve the downstream task. In some cases, overly clean or highly constrained images may even reduce variability and limit generalization, while diverse, potentially noisy samples (including less physically consistent ones) can be beneficial for training and prevent overfitting. From a practical perspective, it is therefore unclear whether the additional complexity and computational cost of generating physically plausible deformations leads to measurable improvements in downstream performance.
More generally, and related to W1, it would be important to compare the proposed approach with purely data-driven generative models that do not rely on labeled inputs. While such models may lack explicit physical consistency, they can potentially generate a larger variety of samples, which may lead to improved downstream performance. This aspect is not explored in the current experimental evaluation.
W3.The training procedure for the force prediction network is not fully clear, particularly regarding the correspondence between sampled points across time steps. The method samples a set of points x_i^t from the retinal area at time t, which are then deformed using the predicted forces to obtain their estimated positions at time t+1.The loss (Equation (2)) is defined by comparing these deformed positions of the points at time t with positions x_i^{t+1} of the points sampled at the next time step t+1.
However, it is not clear how the correspondence between the points at times t and t+1 is established. The description and Figure 1 suggest that point sampling is performed independently at each time step based on the retinal region, without relying on a predefined deformation or tracking mechanism. In such a case, there is no obvious one-to-one mapping between the sampled points at t and those at t+1 used as ground truth. In fact, the number of points sampled at t and at t+1 does not even need to be the same.
Since Equation (2) assumes a pointwise correspondence with N points (same for t and t+1), it would be important to clarify how this matching is performed in practice. This detail is critical for understanding the training procedure and ensuring reproducibility.
W4.[Minor] The physical interpretation of the learned force field remains somewhat unclear. In particular, multiple point-wise force configurations can be projected to the same kernel-level forces (via the projection of Eq. (1)). Therefore, the predicted forces may not directly correspond to meaningful biomechanical quantities. Instead, they may act as latent variables (as opposed to physical forces) that parameterize deformation within the solver. While this does not necessarily limit the effectiveness of the method, it suggests that PhysOCT could be better understood as a physics-inspired regularization approach rather than a model of physically accurate forces. It would be interesting (but not critical) to visualize the estimated point-wise forces in a few cases to better assess their plausibility. In any case, I do not consider this to be a major issue with the method.
- 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
The paper does an excellent job at integrating physical modeling with generative image synthesis. To further strengthen the work, it would be helpful to more clearly articulate the practical use cases of the method, given its reliance on existing images and labels (especially fluid masks). In particular, it would be useful to clarify the intended setting in which fluid masks at time t+1 are available. If such masks are derived from existing OCT images, why is the proposed approach preferred over directly conditioning on those images? Conversely, if these masks are generated or predicted, additional details on how they are obtained (and whether their generation follows similar physical constraints) would help clarify the overall pipeline.
It would also strengthen the paper to evaluate the utility of the generated data in downstream tasks (e. g. , training segmentation or disease prediction models), to assess whether the improved image quality translates into practical benefits. Finally, clarifying aspects of the training procedure regarding the correspondence between sampled points across time steps would improve the reproducibility and clarity of the method.
- 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?
The paper presents an interesting combination of physics-based deformation modeling and diffusion-based OCT image synthesis. It shows moderate but consistent improvements in image quality metrics. The approach is technically sound and the integration of a differentiable physical model into a generative framework is appealing.
However, the practical utility of the method remains insufficiently supported. The reliance on labeled inputs (including masks at both t and t+1) appears to limit the diversity of the generated data and raises questions about its applicability in realistic scenarios. In addition, the evaluation focuses on image quality metrics, but does not assess whether the generated data improves clinically relevant downstream tasks.
As a result, I think that the lack of evidence demonstrating practical impact limits the strength of the contribution. I would therefore place the paper slightly below the acceptance threshold but remain open to revising my assessment based on the rebuttal.
- 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 successfully addressed most of my concerns.
However, I still believe the reliance on future masks greatly limits the practical reach and clinical significance of the method. In my view, generating physically and clinically plausible future lesion evolution is itself a more difficult and clinically relevant problem. While PhysOCT enforces anatomically consistent image deformation conditioned on a prescribed future mask, the realism of the future pathology is effectively delegated to an external editing or mask-generation process that is not subject to similar anatomical or physical constraints. In other words, the framework applies sophisticated physical modeling to the image synthesis stage, but not to the generation of the pathological trajectories that ultimately drive the simulation.
Nevertheless, the paper is technically strong and methodologically novel. The integration of differentiable biomechanical modeling with diffusion-based OCT synthesis is well designed, and the evaluation across multiple datasets and backbones is thorough. Although I believe the practical impact is narrower than initially suggested, I still consider this a valuable contribution.
Review #2
- Please describe the contribution of the paper
The paper proposes PhysOCT, a novel physics-guided framework for longitudinal OCT synthesis. To provide physically meaningful structural priors, PhysOCT introduces a force prediction network that learns lesion-induced forces from pairs of consecutive lesion masks. It then employs a mesh-free GMLS-based solver to propagate retinal deformation using the predicted generalized forces, the current sampled point positions, and precomputed GMLS operators, thereby estimating the point-wise locations, the deformation field, and the predicted next-step retinal layer prior. These physics-derived priors are further used as conditions for diffusion-based OCT synthesis. Extensive experiments covering two major diffusion backbones(DDPM and LDM), and three longitudinal OCT datasets covering three retinal diseases (AMD, CNV, and DME), demonstrate improved simulation accuracy and better visual generation quality. The ablation and sensitivity studies confirm the importance of each simulation components, thereby systematically validating the effectiveness of the proposed framework.
- 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 technical innovation is well motivated and conceptually novel. The paper introduces a new perspective on OCT synthesis by formulating lesion evolution as a differentiable physical deformation process, thereby enabling the derivation of physically interpretable structural priors for OCT B-scan generation. 2.The methodology of PhysOCT is mathematically grounded and insightful. The framework estimates generalized forces from point-wise forces via the principle of virtual work, and then simulates retinal geometry through a differentiable solver, resulting in physically grounded conditions for reliable OCT generation. 3.Extensive experiments on diverse disease(AMD/CNV/DME) OCT images and two major generation framworks(DDPM and LDM) further demonstrate its improvement and generalizability.
- 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.
There are also several weaknesses of this manuscript: 1.The major weaknesses is the incomplete disclosure of dataset size, especially for the private CNV cohort. 2.The paper’s claims would benefit from a more rigorous justification through ablation studies on image generation conditions. Although the experimental results are relatively self-contained from the physical simulation perspective,the generation pipeline in PhysOCT involves multiple inputs, including the warped OCT sketch, the predicted layer prior, and the target fluid structure. Since the target fluid mask may provide a strong structural prior for image generation, disentangling the contribution of each conditioning signal would help clarify the performance gains and make the paper’s claims more convincing. 3.The presentation of Table 2 has the potential to be misleading. The entries “100× parameters” and “0.01× parameters” appear to correspond to a sensitivity analysis of the Young’s modulus E, but this is not explicitly specified in the table and may cause ambiguity for readers. Since the simulation relies on the point-wise discretization of the retinal layers, a further sensitivity analysis with respect to the sampled points numer N would substantially improve the transparency of the method. 4.In practice, the training of PhysOCT still relies on point-level supervision extracted from the target-layer for the force predictor. As such, the claim that the method operates “without dense paired structural supervision at scale” may require further clarification and more explicit justification. 5.Including downstream utility validation, such as evaluating the anatomical consistency between generated images and the ground truth or assessing the diagnostic performance of models trained on the synthesized data, would further enhance the rigor of the study and better demonstrate its practical value. 6.The authors should also discuss the limitations of the proposed method. For example, under what circumstances would this method fail?
- 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
Here are a few additional comments on the manuscript, I would appreciate it if the authors could, if possible, share their thoughts on the following points:
1.In my view, the main contribution of this work lies in providing a physically grounded structural reference for conditional generation, while the generative model itself (e.g., DDPM/LDM) mainly serves as an image renderer that complements texture and appearance. To some extent, this may not fully exploit the potential of generative models. Since the both physical principles and generative modeling may offer strong generalization capacity and promising applications in disease-risk modeling and pathological progression analysis. It would be beneficial if the authors could bring discussions on this perspective and outline possible future directions for such applications.
2.Compared with purely deep learning-based methods, the longer inference time remains a practical challenge for the application of simulation-based approaches. I am curious whether this physically guided paradigm could be distilled into a learning-based module to enable more efficient deployment.
- 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?
The paper is well-written, the techinial contribution is novel and solid, the experimental results are convincing, and the conclusion is self-contained.
- 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 author has addressed my concerns.
Review #3
- Please describe the contribution of the paper
The main contribution of the paper is the introduction of PhysOCT, a physics-guided plug-in framework designed to improve the anatomical accuracy of conditional Optical Coherence Tomography (OCT) image generation. Specifically, it addresses the common issue where generated OCT scans look realistic but fail to show the physically correct deformation of retinal layers around changing lesions. PhysOCT solves this by predicting forces induced by changes in lesions and simulating how these forces physically deform the surrounding retinal layers over time, and subsequently use these as a structural guide for the diffusion models to synthesize anatomically consistent and realistic longitudial OCT scans.
- 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.Novel hybrid formulation: It bridges neural networks and biomechanics by predicting lesion-induced point-wise forces and propagating them through a physics solver to ensure anatomically consistent layer deformations 2.Modular Architecture: PhysOCT is designed as a flexible conditioner that integrates with existing diffusion models to improve generation fidelity. 3.Stable Differentiable Solver: The framework ensures stable end-to-end training by predicting point-wise forces and utilizing an SPD linear system to avoid numerical failures. 4.Efficient: It provides minimal computational overhead 5.Evaluated thoroughly for MICCAI standards on three different longitudinal disease datasets (CNV, DME and AMD).
- 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.Fixed biomechanical parameters: Tissue elasticity may vary widely between individual patients and age groups - though I understand that this is needed for now for numerical stability. 2.Limitation to 2D: I understand that this is mainly due to a computational barrier, yet I want to mention it: Currently it is strictly limited into 2D, by predicting 2D forces, but this ignores crucial out-of-plane forces, which are present in the three dimensional retinal volumes. 3.No comparison against SOTA generative models, like SD3, FLUX.
Minor comments: Please use vector graphics wherever it is possible. The plots in Figure 2 and Figure 3 are currently rasterized, which do not look so nice. Also, if possible, please include the full-scale images in the work (downscaled only in PDF), so that upon zooming in, the small details become visible. I find this very important for publications about generative models.
- 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.
(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?
Nicely done work. However, I miss somewhat some clinical downstream task evaluation, to prove that this synthesized data actually improves downstream clinical algorithms.
- 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.
My few concerns were answered. I think this work is without doubt an accept.
Author Feedback
Thank all reviewers and the AC for their constructive feedback. We address the main concerns below. (1) R1-W1, Meta-Reviewer: Concern that the method requires a real t+1 scan, limiting practical utility. PhysOCT is positioned as a physics-consistent OCT image editing model, not an image predictor. At inference, the t+1 fluid mask comes from: (i) clinician-specified or algorithmically edited fluid masks derived from a known mask, e.g., augmenting rare lesion morphologies (similar to FESM [20]); (ii) a mask prediction network, which addresses a lower-dimensional structure prediction task than full-image synthesis. In both cases, no real t+1 scan exists or is necessary. PhysOCT’s core value lies in turning a mask-level edit into an anatomically consistent image edit, a guarantee that purely data-driven generators cannot provide. (2) Meta-Reviewer, R3-W2: Concern about the method’s 2D limitation and extensibility to 3D. The formulation extends naturally to 3D: both the physics simulator and force prediction can be lifted to 3D. The main obstacles are computational cost and the scarcity of paired longitudinal OCT volumes. A 2.5D extension is a feasible intermediate, with GMLS interpolation keeping the cost increase manageable. (3) R1-W3: Lack of clarity on point correspondence between t and t+1.Points at time t are uniformly sampled pixel-wise within the retinal region defined by layer_t, excluding the lesion region. At time t+1, each sampled point retains its relative position within its inter-layer band, and is mapped via linear interpolation between layer_t and layer_{t+1}. The resulting t+1 coordinates serve as supervision for the force prediction network. (4) R1-W2, R2-W5, R3: Absence of downstream task evaluation. We agree downstream evaluation is important. As R2 highlights, “anatomical consistency between generated images and ground truth” is the key evaluation axis, which is exactly what our deformation metrics (Rel-L2/MAE, Table 1) and layer-overlay analysis (Fig. 3) directly measure. Common OCT segmentation tasks (fluid or layer) are not the most sensitive probe of PhysOCT’s contribution: PhysOCT can introduce lesion/layer diversity by editing the t+1 fluid mask, but since the generated images are derived from existing scans, diversity-driven augmentation is not its strongest fit; meanwhile, the layer-lesion structural consistency that PhysOCT contributes does not directly translate into segmentation gains. A more suitable downstream task is longitudinal dataset augmentation for temporal prediction. These tasks demand high realism in predicted images, and require augmented training data anatomically realistic and physically coherent. This is exactly what PhysOCT delivers. By analyzing fluid-evolution distributions and generating under-represented trends via targeted mask editing, PhysOCT can augment longitudinal datasets in ways unattainable by appearance-based generation. (5) R2-W1: Incomplete disclosure of the CNV dataset size. The CNV cohort includes 39 patients and 49 paired longitudinal 3D cubes (Optovue, 400×640×400). 936 paired 2D B-scans with substantial lesion change are split 8:2 at the patient level. The force predictor uses this set; the generative model is additionally trained on remaining labeled paired B-scans from the same training patients. (6) R2-W2, R3-W3: Need for ablation on conditioning signals and comparison with SOTA generators. Our main contribution is the PhysOCT framework, not the backbone. Table 3 acts as the ablation: DDPM/LDM baselines use fluid_{t+1} and the current image, while +PhysOCT adds the warped sketch and layer prior, with consistent PSNR/SSIM/FID gains and boundary-aligned error reduction in Fig. 3.SD3/FLUX, as natural-image foundation models, differ from the OCT setting in both training scale and domain, making fair comparison infeasible; meanwhile, PhysOCT is designed as a pluggable module, orthogonal to backbone choice, hence our comparison focus on standard LDM/DDPM.
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.
This paper received mixed scores 3, 5, 5.Concerns include the method’s generalizability (e.g., reliance on labeled inputs including masks at both t and t+1), and 3D adaptability. AC invites the authors for a rebuttal to clear the concerns raised by reviewers.
- 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.
Although there remain some concerns about the practical impact and clinical usefulness, all reviewers agree that this work is novel and well-suited for MICCAI. Therefore I recommend accept.
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.
Following a constructive rebuttal phase, the authors addressed the reviewers’ primary concerns regarding the practical utility of the mask conditioning and the dataset specifications. This led to the consensus acceptance.
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.
Reviewers agree positively on this paper and find the physics-guided formulation technically interesting and well motivated. The rebuttal addressed the major concerns, especially regarding the intended use case, extensibility, and evaluation scope.
