List of Papers Browse by Subject Areas Author List
Abstract
Controllable medical video generation has achieved remarkable progress, but it still lacks reliability, which requires the alignment of generated contents with physical priors and faithful clinical manifestations. To push the boundaries from mere controllability to reliability, we propose DepthPilot, the first reliable framework for colonoscopy video generation. This work takes a step toward trustworthy generation through two synergistic paradigms. To achieve explicit geometric grounding, DepthPilot devises a prior distribution alignment strategy, injecting depth constraints into the diffusion backbone via parameter-efficient fine-tuning to ensure anatomical fidelity. To enhance intrinsic nonlinear modeling under these geometric constraints, DepthPilot employs an adaptive spline denoising module, replacing fixed linear weights with learnable spline functions to capture complex spatio-temporal dynamics. Extensive evaluations across three public datasets and in-house clinical data confirm DepthPilot’s robust ability to produce physically consistent videos. It achieves FID scores below 15 across all benchmarks and ranks first in clinician assessments, bridging the gap between “visually realistic” and “clinically reliable”. Moreover, DepthPilot-generated videos are expected to enable reliable 3D reconstruction, facilitating surgical navigation and blind region identification, and serve as a foundation toward the colorectal world model.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5550_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/5550_supp.zip
Link to the Code Repository
https://github.com/JunhuFu/DepthPilot
Link to the Dataset(s)
N/A
BibTex
@InProceedings{FuJun_DepthPilot_MICCAI2026,
author = { Fu, Junhu AND Chen, Ke AND Guo, Weidong AND Liang, Shuyu AND Xu, Jie AND Ma, Chen AND Wang, Kehao AND Lin, Shengli AND Li, Zeju AND Wang, Yuanyuan AND Guo, Yi AND Li, Shuo},
title = { { DepthPilot: From Controllability to Reliability in Colonoscopy Video Generation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes DepthPilot, a framework for colonoscopy video generation that conditions diffusion-based synthesis on depth priors to improve geometric consistency, while introducing an adaptive spline denoising module to better model complex spatio-temporal dynamics. The method is evaluated on multiple public datasets and demonstrates strong quantitative as well as clinician-rated performance.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The main strength of the paper lies in its well-motivated formulation and its clear focus on improving the realism of colonoscopy video generation. Both the PDA strategy and the ASD module are designed toward this goal: PDA introduces depth-guided geometric constraints to improve anatomical plausibility, while ASD aims to better capture complex spatio-temporal patterns and reduce blur or incoherence. The paper is also supported by relatively solid experiments across multiple datasets, including clinician evaluation and ablation studies.
- 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.ASD design is insufficiently justified. It remains unclear why the proposed spline-based ASD design is preferable to conventional nonlinear components already used in modern backbones. Important design choices, such as the number and placement of ASD blocks, are also not sufficiently explained.
2.Method description needs clarification. Some implementation details remain unclear. In particular, the tokenization process in Fig. 2 are not clearly explained.
3.Novelty of synthesized videos is unclear. The conditions, including depths and reference images, are derived from real training data. It is unclear whether the synthesized videos are truly novel. Providing evidence, such as a case study, would help demonstrate that the synthesizer is not merely replicating the real dataset.
4.Sensitivity to depth priors is not analyzed. The method relies heavily on estimated depth priors, but the paper does not examine how errors in depth estimation may affect generation quality.
5.The limitations of the method should be discussed.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
I would recommend weak accept, pending a satisfactory rebuttal. The proposed framework shows promising novelty and strong empirical performance, although I still have several major concerns that should be addressed more carefully.
- 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.
The rebuttal addresses some of my original concerns. However, the number and placement of ASD blocks are not fully supported by ablation. The novelty of the synthesized videos relative to the training set is also not directly demonstrated, for example through a feature-level nearest-neighbor analysis. Nevertheless, combined with the well-motivated formulation, the strong empirical performance, the clinician evaluation, and the released code, the paper is on balance slightly above the acceptance threshold.
Review #2
- Please describe the contribution of the paper
The paper presents DepthPilot, a diffusion-based framework for interpretable colonoscopy video generation. The main idea is to enforce physical consistency and explicit geometric grounding by aligning the generated content with strict anatomical constraints. This is achieved in two primary ways: first, a Prior Distribution Alignment (PDA) strategy injects monocular depth sequences into the diffusion backbone via parameter-efficient fine-tuning. Second, an Adaptive Spline Denoising (ASD) module replaces fixed linear weights with learnable B-spline functions to better capture complex, nonlinear spatio-temporal dynamics. The method operates within a conditional latent diffusion architecture, utilizing the depth prior alongside a reference image to guide the generation process. Experiments on three public datasets and an in-house clinical database suggest improvements over prior work in visual realism, temporal coherence, and anatomical fidelity, achieving state-of-the-art FID scores and high clinician assessments.
- 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.
Integration of the ASD module: The introduction of learnable B-splines (inspired by the KAN architecture) into the diffusion model represents a sound engineering choice aimed at better capturing the complex, nonlinear spatio-temporal dynamics of the colon. Effective geometric conditioning (PDA): The distribution alignment strategy is a practical and effective solution to prevent training collapse when injecting external depth maps into the pre-trained backbone. Strong quantitative results: The model demonstrates high empirical performance, achieving competitive metric scores, including an FID below 15 across all evaluated 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.
W1.Misalignment of Core Terminology (“Interpretability”): The core premise of the paper relies heavily on the term “interpretable framework.” However, in the broader Machine Learning and Explainable AI (XAI) community, “interpretability” refers to the ability to understand the internal decision-making process of a model. In this work, the Prior Distribution Alignment (PDA) enforces anatomical grounding and physical consistency by conditioning the generation on depth maps. While valuable, this does not make the generation process intrinsically explainable. This semantic discrepancy misleads the reader regarding the actual nature of the methodological contribution. W2.Unsubstantiated Claims Regarding the “Lightweight” Architecture: The authors describe their depth prior encoder as “lightweight” and “parameter-efficient.” Unfortunately, the manuscript lacks the quantitative evidence necessary to support these claims. Without a detailed computational complexity analysis - such as the number of trainable parameters in the encoder versus the frozen backbone, or the impact on training time and inference speed (FPS) - the architectural efficiency of the proposed method remains unverified. W3.Arbitrary Architectural Choices (B-Splines): The introduction of “cubic B-splines” is justified by the authors as a balance between representation power and efficiency. However, this architectural choice appears arbitrary as it is not supported by an empirical ablation study. The absence of a comparison between different spline orders (e.g., linear, quadratic, cubic) or varying grid intervals makes it difficult to ascertain if this configuration is indeed optimal for the task. W4.Unaddressed Vulnerability to Depth Estimation Failures: The DepthPilot framework is heavily dependent on an external, self-supervised monocular depth estimation model. Colonoscopy videos frequently contain challenging artifacts such as specular reflections (glare), fluids, and rapid camera movements, which often cause depth estimators to fail. The current manuscript does not evaluate the robustness of the generative model under these challenging conditions or when conditioned on noisy depth priors, leaving a significant gap in assessing its real-world clinical feasibility. W5.Ambiguity in Data Splits and SOTA Comparisons: The authors note that datasets were randomly partitioned into an 80/20 train/evaluation split. However, it is unclear if the baseline models (e.g., ColoDiff, FEAT-L) in Table 1 were evaluated using the exact same data split and random seed. Without ensuring identical data distributions for the test sets, the reported improvements in highly sensitive metrics like FID and FVD are difficult to fully validate and reproduce. W6.Speculative Claims Regarding 3D Reconstruction: The manuscript claims that DepthPilot-generated videos “are expected to enable reliable 3D reconstruction.” This is a strong assertion that is not supported by any experiments, visual demonstrations, or metrics within the paper. Including such speculative claims without empirical backing diminishes the scientific rigor of the presented conclusions. W7.Unclear Visual Communication in Figure 1: The visual design of Figure 1 is confusing and does not effectively communicate the limitations of existing methods. The use of only negative indicators (“crosses”) makes it visually difficult to understand what successful generations or positive aspects should look like in this context. W8.Limited Statistical Detail in the Clinician Score (CS): While including clinical assessment is a strong conceptual point, the execution lacks statistical depth. The Clinician Score is based on the subjective ratings of only three specialists, which constitutes a very small sample size. Furthermore, reporting only the average score without providing the variance or distribution of the ratings makes it challenging to interpret the statistical significance and reliability of this metric. W9.Incomplete Ablation of the ASD Module: The ablation study for the ASD module (Figure 5) successfully highlights its impact on spatial consistency (reducing local blurring and structural degradation). However, it does not provide visual or quantitative evidence regarding how the module affects temporal consistency over a longer sequence of frames, which is a critical aspect of evaluating video generation models. W10.Absence of a Dedicated Limitations Section: The manuscript does not include a section discussing the limitations of the proposed approach.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the paper presents a technically promising hybrid architecture and achieves strong quantitative results, several critical methodological and terminological flaws outweigh its strengths. The core claim of proposing an “interpretable framework” is fundamentally flawed; the PDA strategy enforces geometric and physical consistency via depth conditioning, which does not equate to interpretability in the context of Explainable AI. This severely misrepresents the paper’s actual contribution. Furthermore, the introduction of cubic B-splines lacks empirical validation, as no ablation study compares different spline orders against computational overhead. The framework also heavily relies on external depth estimation but fails to analyze how the generative process behaves under common clinical failure cases, such as specular reflections causing noisy depth maps. Finally, it remains unclear if the 80/20 data split strictly aligns with the evaluation protocols of the baseline methods, casting doubt on the absolute fairness of the metric comparisons. Ultimately, the paper overpromises with unfounded claims of interpretability and 3D reconstruction while under-evaluating its core novelties. To be considered for acceptance, the authors must revise the narrative to focus on geometric grounding, provide the missing ASD ablations, and demonstrate the model’s robustness to imperfect inputs.
- 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 authors addressed all of my concerns. From my perspective, the main issue was the use of the term “interpretability.” In the machine learning field, interpretability has a well-established meaning and typically refers to systems whose internal mechanisms can be understood and explained. In this case, the authors expressed their willingness to replace “interpretability” with “reliability,” which I believe is a much more appropriate term for this work. I am concerned that the use of “interpretability” could mislead readers, whereas “reliability” better reflects the actual contribution of the paper.
Review #3
- Please describe the contribution of the paper
Prior Distribution Alignment (PDA) Strategy: A parameter-efficient fine-tuning approach that injects depth-based physical priors into a latent diffusion backbone. This ensures anatomical fidelity and geometric grounding by aligning depth embeddings with the backbone feature distribution, preventing catastrophic forgetting while enforcing physical constraints.
Adaptive Spline Denoising (ASD) Module: Inspired by the Kolmogorov-Arnold representation theorem, this module replaces fixed linear weights in the denoising network with learnable B-spline functions. This enhances nonlinear representation capacity to capture complex spatio-temporal dynamics of irregular intestinal structures under geometric constraints.
Clinical Validation: The work introduces Clinician Score (CS) evaluated by three senior clinicians, demonstrating that the generated videos are not only realistic (FID < 15 across all benchmarks) but also clinically interpretable in terms of anatomical plausibility and lesion characteristics.
- 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.
Novel Conceptual Advance: The transition from “controllability” to “interpretability” represents a significant conceptual advancement in medical generative AI. While existing methods focus on mask or class conditioning, DepthPilot enforces physical priors (depth) to ensure generated content complies with anatomical constraints (e.g., hyperplastic vs. adenomous polyp characteristics, regional occurrence patterns).
Parameter-Efficient Integration: The PDA strategy’s use of distribution alignment and selective fine-tuning (only ASD blocks) while freezing the pre-trained Stable Video Diffusion backbone is technically elegant. It prevents catastrophic forgetting while effectively injecting geometric constraints.
Adaptive Spline Architecture: The ASD module’s use of learnable B-spline activation functions (cubic B-splines) instead of fixed linear weights is novel for video diffusion models. This addresses a fundamental limitation of current architectures in modeling irregular, scale-varying anatomical structures and rapid camera motion typical in endoscopy.
Strong Clinical Validation: Unlike many generative papers that rely solely on metrics like FID/FVD, this work includes clinician assessment (CS) evaluating physical interpretability, anatomical plausibility, and clinical manifestations. The ranking “first in clinician assessments” lends credibility to the interpretability claims.
Broad Compatibility: The framework accepts depth priors from diverse sources (real video estimates, simulation, phantom data), making it practical for various clinical scenarios including rare lesion synthesis and surgical planning.
Comprehensive Evaluation: Testing across four datasets (three public + one in-house clinical dataset with ethics approval) demonstrates robustness 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.
Limited Architectural Novelty: While the PDA and ASD components are novel, the underlying architecture remains largely based on Stable Video Diffusion [1] with standard VAE and CLIP encoders. The contribution lies in the conditioning strategy and activation functions rather than fundamental architectural innovation.
Dependency on Depth Estimation Quality: The method relies on monocular depth estimation [32] as input. The paper does not analyze failure modes when depth estimation is inaccurate (e.g., in poor lighting conditions or with severe artifacts), which could limit reliability in real-world deployment.
Small-Scale Clinical Validation: While the inclusion of clinician scores is commendable, the assessment by only three clinicians limits statistical power. No inter-rater reliability metrics (e.g., Cohen’s kappa) are reported to quantify agreement between clinicians.
Computational Efficiency Concerns: The paper mentions training on A100 GPUs but does not provide inference time statistics or memory requirements. Given the use of learnable spline functions which may increase computational complexity compared to standard linear layers, efficiency comparisons with baseline methods would strengthen the paper.
Limited Generalization Analysis: The evaluation focuses exclusively on colorectal regions. The claimed compatibility with phantom and simulation data is mentioned but not empirically validated in the provided text.
Temporal Consistency Metrics: While FVD is reported, specific metrics for temporal consistency (e.g., optical flow smoothness or frame-to-frame depth consistency) are not explicitly analyzed, which is crucial for surgical navigation applications claimed in the conclusion.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
Depth Quality Analysis: Please include an analysis of how depth estimation errors propagate to the generated videos. This is crucial for clinical safety.
Clinician Agreement: Report inter-rater reliability statistics for the clinician scores to validate the consistency of the clinical assessment.
Computational Metrics: Include inference time and memory footprint comparisons with baselines, as spline-based activations may have different computational characteristics than standard convolutions.
Temporal Consistency: Consider adding specific temporal consistency metrics (e.g., LPIPS between consecutive frames, or optical flow divergence) to complement FVD for the surgical navigation use case.
Phantom Validation: If possible, include results using phantom depth data to substantiate the claim of broad compatibility with simulation/phantom sources.
Ethical Considerations: While you mention ethics approval, briefly clarify whether the clinicians rating the videos were blinded to the synthetic vs. real nature of the content to prevent bias in the CS evaluation.
- 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?
Positive Factors:
Clinical Impact: The paper addresses interpretability in medical video generation—a crucial step toward trustworthy AI for data augmentation and surgical training. The clinical validation with expert ratings provides evidence of real-world utility. Technical Innovation: The combination of parameter-efficient fine-tuning with physical prior injection (PDA) and learnable spline activations (ASD) represents a thoughtful, theoretically-grounded approach to handling the unique challenges of endoscopic video (irregular anatomy, rapid motion). Comprehensive Evaluation: Strong quantitative results (FID < 15, best FVD scores) across four datasets, combined with qualitative demonstrations of anatomical fidelity (Fig. 4), support the claims. Reproducibility: Provision of anonymized code facilitates verification. Minor Concerns:
The clinical validation involves only three clinicians; larger studies would strengthen the clinical claims. Computational efficiency details are missing, which is important for real-time surgical applications. The reliance on pre-trained Stable Video Diffusion limits the architectural novelty to the conditioning and activation components.
- 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 all reviewers for your valuable feedback. Reviewers recognize our motivation and novelty (R2: “well-motivated formulation”, R4: “novel conceptual advance”), reproducibility (R2-4), and clarity (R2, R4). We address main concerns below.
I. Concept Clarification 1.R3 on interpretability Current interpretability reflects decision-making reliability of trained models, but ill-suited for generative models. Here we explore clinical interpretability to assess alignment with physical priors and clinical manifestations, where the physically grounded paradigm enhances generation-process understanding. Our definition may be bold, but we regard it rational. If “interpretability” is still deemed unsuitable, we can adopt “reliability” instead. 2.R4 on architectural novelty Our novelty lies in a generative model grounded in physical prior and clinical manifestation. Additionally, ASD module is an architectural change: beyond activation replacement, it turns neuron-edge weights into learnable splines. 3.R2 on tokenization Our tokenization follows ViT-style patch embedding: feature maps are split into patches and flattened into vector sequences.
II. Method and Experiment 1.R2, R3 on ASD clarification and ablation Conventional nonlinear components: fixed activation function (ReLU/GELU); ASD: learnable splines on neuron edges (Fig.2), thus enabling adaptive nonlinear modeling of spatio-temporal dynamics. Testing quadratic (FVD=306, 57M learnable params), cubic (FVD=289, 62M), and quartic (FVD=275, 91M) splines on Colonoscopic dataset, we chose cubic as a trade-off. ASD blocks are placed at bottleneck (Fig.2) to leverage rich semantic features without dimension change. For ablation, local blurring in partial frames of Fig. 5(b) shows temporal incoherence, and FVD (Tab.2) inherently reflects cross-frame continuity. We visualized more cases, observing inter-frame incoherence w/o ASD. 2.R2 on video diversity, R4 on simulation data analysis Condition inputs (depth, reference image) can additionally derive from simulation/phantom besides real training data. We had conducted experiments with Unity3D simulation [31], and DepthPilot also generates authentic videos (FVD/FID/IS/CS: 356/15.0/3.64/3.77). Moreover, even conditioned on real data, IS and CS still reflect synthesized diversity. 3.R2, R3, R4 on depth estimation error DepthPilot is robust: depth estimation only provides a basic physical constraint; estimation error can even enhance generation diversity. Moreover, depth prior from simulation/phantom requires no estimation, thus eliminating this issue at its root. 4.R3, R4 on clinical evaluation CS complements objective scores (FVD, FID, IS). Clinicians were blinded to real/synthetic videos, with Fleiss’ kappa=0.72 showing good agreement. We will involve more clinicians for further evaluation. 5.R4 on temporal consistency metrics FVD computed on spatio-temporal features reflects inter-frame consistency. Additionally, we measured depth MAE between adjacent frames, DepthPilot achieves 28.76, outperforming second-best method (FEAT-L, 42.31).
III. Other Details 1.R3, R4 on computational efficiency Using parameter-efficient fine-tuning, DepthPilot has ~1,680M parameters (3.7% trainable, 62M). Training (batch size=4) requires ~22GB memory, and single A100 inference reaches 27.25 FPS. 2.R3 on data split All methods used identical 8:2 train/eval split, which will be clarified. 3.R3 on successful cases in Fig.1 Fig.1 aims to illustrate limitations of current controllable generative methods: lack of physical constraints and clinical manifestations, which motivate DepthPilot. Successful cases are shown in Fig.4.4.R3 on reconstruction prospect Integrating physically grounded videos into training naturally mitigates feature mismatch and camera pose estimation error in 3D reconstruction. Our work provides such videos, thus making reconstruction a reasonable prospect. If it is still regarded an overclaim, we can remove it in the revision.
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 was reviewed by three experts in the field and received the following recommendations: two Weak Accepts and one Reject. While the reviewers noted several strengths, including promising quantitative results and the integration of effective ASD and PDA modules, they also raised several concerns, including unclear or potentially misleading presentation, insufficient evaluation, and limited justification of the model design. The authors are encouraged to address these concerns in the rebuttal.
- 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.
Based on the reviewers’ post-rebuttal feedback, the paper is recommended for acceptance to MICCAI 2026.The authors are encouraged to incorporate the necessary changes discussed during the rebuttal process into the final camera-ready version.
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.
Although the ablation study is not sufficient to evaluate all of the proposed components, the reviewers have positively assessed the novelty of the method and the validation results. Therefore, I believe this paper has value and is suitable for presentation at MICCAI.
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.
I recommend acceptance. The paper proposes DepthPilot, a depth-guided diffusion framework for colonoscopy video generation, with two main components: prior distribution alignment for incorporating depth-based geometric constraints, and adaptive spline denoising for modeling nonlinear spatio-temporal dynamics. The reviewers generally recognized the motivation, empirical performance, clinician evaluation, and released code.
The initial concerns were mainly about terminology, justification of the ASD design, robustness to depth estimation errors, data split clarity, computational cost, and clinical evaluation details. The rebuttal addressed these points reasonably well. In particular, the authors clarified the use of “interpretability” and agreed to adopt the more appropriate term “reliability” if needed; provided additional ASD ablation results, computational statistics, data split clarification, clinician blinding and agreement information, and a temporal consistency metric. These responses substantially improve confidence in the method and evaluation.