Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Surgical action automation has progressed rapidly toward achieving surgeon-like dexterous control, driven primarily by advances in learning from demonstration and vision-language-action models. While these have demonstrated success in table-top experiments, translating them to clinical deployment remains challenging: current methods offer limited predictability on where instruments will interact on tissue surfaces and lack explicit conditioning inputs to enforce tool-action specific safe interaction regions. Addressing this gap, we introduce AffordTissue, a multimodal framework for predicting tool-action specific tissue affordance regions as dense heatmaps during cholecystectomy. Our approach combines a temporal vision encoder capturing tool motion and tissue dynamics across multiple viewpoints, language conditioning enabling generalization across diverse instrument-action pairs, and a DiT-style decoder for dense affordance prediction. We establish the first tissue affordance benchmark by curating and annotating 15,638 video clips across 103 cholecystectomy procedures, covering six unique tool-action pairs involving four instruments (hook, grasper, clipper, scissors) and their associated tasks: dissection, grasping, clipping, and cutting. Quantitatively, our task-specific architecture achieves an average symmetric surface distance (ASSD) of 20.6 px, substantially outperforming vision/vision-language baseline models (U-Net: 44.3 px and Molmo-VLM: 60.2 px) in dense surgical affordance prediction. By predicting tool-action specific tissue affordance regions, AffordTissue provides explicit spatial reasoning for safe surgical automation, potentially unlocking explicit policy guidance toward appropriate tissue regions and early safe stop when instruments deviate outside predicted safe zones.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5918_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{MakAiz_AffordTissue_MICCAI2026,
        author = { Maksutova, Aiza AND Seenivasan, Lalithkumar AND Ding, Hao AND Xu, Jiru AND Yu, Chenhao AND Jing, Chenyan AND Shen, Yiqing AND Unberath, Mathias},
        title = { { AffordTissue: Dense Affordance Prediction for Tool-Action Specific Tissue Interaction } },
        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 a new prediction task in surgical intervention, surgical action affordance, which is formulated as a dense prediction problem to localize instrument–tissue interaction regions. The authors propose a vision–language framework that takes a language query along with a buffered video stream as input to predict an affordance map corresponding to a given surgical action. The model is trained and evaluated on a dataset collected from five different sources, where key frames are annotated with both textual queries and affordance maps.

    Overall, the proposed framework aims to provide improved spatial understanding of surgical videos and shows potential for downstream applications such as surgical guidance.

  • 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 introduces a novel dataset that covers a large number of videos from diverse sources, which is valuable for advancing research in this area.

    2)The ablation study is comprehensive and evaluates the contributions of the proposed components in a systematic manner.

    3)The overall structure of the paper is clear, and both the methodology and experimental setup are well described.

  • 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 reported DICE score is relatively low (0.124), which raises concerns about the effectiveness of the method. It would be helpful to understand why this is the case, especially in comparison to SAM-based approaches that achieve higher DICE scores. Additionally, more details are needed on how SAM (or SAM3) is prompted in this context.

    2)How about alternative dense prediction baselines? For example, will standard architectures such as U-Net be able to be adapted to this task and how they compare with the proposed method?

    3)From a clinical perspective, the definition of the affordance regions is not entirely clear. Since this is a newly introduced dataset, more details are needed regarding the annotation protocol, including how consistency and reliability are ensured across videos/annotators.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    1)Reproducibility: It would be beneficial to clarify whether the code and dataset will be publicly released.

    2)In the Results section, the statement “the intensity discrepancy does not pose a problem as of now” requires further elaboration. It would be useful to discuss whether this might become an issue in other scenarios or datasets.

    3)The paper would benefit from reporting computational details, such as training time and inference time, to better assess its practical applicability.

  • 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?

    Please see the main weaknesses.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors have addressed several of my main concerns in the rebuttal. I encourage the authors to incorporate these additional details into the final manuscript, as they improve the clarity and credibility of the work. I also encourage them to open-source the curated dataset and code, since this would greatly benefit reproducibility and future research in the topic.



Review #2

  • Please describe the contribution of the paper

    This paper introduces AffordTissue, a multimodal framework for predicting tool-action specific tissue affordance regions as dense heatmaps during laparoscopic cholecystectomy. The proposed architecture combines a frozen Video Swin Transformer for spatiotemporal encoding of 256-frame temporal windows, a frozen SigLIP 2 language encoder for tool-action conditioning, and a trainable AdaLN-based decoder adapted from DiT for dense heatmap prediction. The work also curate the “first” tissue affordance benchmark, annotating 15,638 video clips from 103 cholecystectomy procedures across five source datasets, encompassing six tool-action pairs. Evaluation against VLM baselines (Molmo, QWEN) and SAM3 shows strong improvements in spatial alignment metrics (ASSD, PCK), with the proposed approach achieving 20.6 px ASSD compared to 60.2 px for the best-performing baseline. Impressive and extensive ablation studies validate the contribution of each architectural component.

  • 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.
    • Framing tissue affordance prediction as a dense heatmap estimation task is a genuinely novel and well-motivated formulation that bridges surgical scene understanding and robot safety, going meaningfully beyond existing semantic segmentation approaches.
    • The large benchmark dataset is a substantial contribution in its own right and could enable future research in this area if released.
    • The design choice of conditioning on tool-action language prompts rather than hardcoding tool categories is principled and enables generalization to unseen instrument-action combinations.
    • The ablation studies are thorough and cover language encoder choice, vision backbone, decoder type, temporal context, and prompt structure, giving a clear picture of what each component contributes.
    • The dual-mode safety application is well-articulated and clinically meaningful and repurposing the AdaLN decoder from diffusion modeling for dense heatmap prediction is an interesting and practical architectural adaptation that is clearly explained.
  • 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.Annotation Protocol Is Insufficiently Rigorous for a Safety-Critical Task: The ground truth affordance regions are defined by four manually annotated keypoints forming a polygon, from which a Gaussian heatmap is generated centered at the polygon centroid. This annotation protocol raises several concerns: (1) four keypoints are a very coarse representation of a tissue interaction zone, particularly for irregular or elongated anatomical structures, (2) the paper provides no information about annotator qualifications; whether annotations were performed by surgeons or trained laparoscopic experts, how many annotators were used, or what inter-annotator agreement looks like. For a task explicitly framed around surgical safety, the reliability and clinical validity of the ground truth is paramount. If affordance regions are inconsistently or inaccurately annotated, any model trained on them inherits those errors, and the safety claims become difficult to defend. Recommendation: Report inter-annotator agreement statistics (e.g., keypoint displacement variance or IoU between annotators) if annotated by humans, specify annotator qualifications, and consider whether four keypoints are sufficient to capture the geometry of the affordance region or whether a more expressive annotation format such as a freehand polygon or brush mask would be more appropriate.

    2.The Comparison Against Baselines Is Not Fully Fair: The proposed model is a task-specific architecture trained end-to-end on 15,638 annotated clips for the exact task of tissue affordance heatmap prediction. The baselines (Molmo-VLM, QWEN-8B, and SAM3) are general-purpose foundation models fine-tuned only with regression heads for vertex prediction or segmentation, and their outputs are post-hoc converted to heatmaps via Gaussian centering. This comparison is inherently asymmetric: the proposed model has a significant architectural and training advantage by design. Furthermore, the heatmap conversion from predicted polygon vertices introduces an additional source of error for the baselines that does not apply to the proposed model. The conclusion that “task-specific architecture outperforms large-scale foundation models” is expected and unjustifiable given this setup and does not tell us much about the core architectural choices. Recommendation: Include at least one more competitive baseline, for example, a simpler task-specific model such as a UNet or ViT-based heatmap predictor trained on the same data, to isolate the contribution of the specific architecture choices (AdaLN decoder, SigLIP conditioning, Swin backbone) from the benefit of task-specific training in general.

    3.The Proposed Task Is Restricted to a Single Procedure and Six Tool-Action Pairs: The entire framework, dataset, and evaluation are scoped exclusively to laparoscopic cholecystectomy, covering only four instruments and six tool-action pairs. While the authors frame cholecystectomy as a good starting point, the broader claims about safe surgical automation and policy guidance are much harder to support when the system has only been demonstrated on one procedure. Surgical affordance needs vary substantially across procedures: robotic prostatectomy, colorectal surgery, and nephrectomy involve very different tissue types, instrument interactions, and safety constraints. The current scope also means the dataset’s tool-action pairs are heavily imbalanced: dissect-hook accounts for over 10,000 clips while cut-scissors has only 343, which may bias both training and evaluation. Recommendation: Discuss the procedure-specificity as an explicit limitation, analyze performance broken down per tool-action pair to expose the effect of class imbalance, and outline a concrete plan for extending the benchmark to additional procedures or robotic platforms as future work.

    4.Temporal Context Ablation Suggests Its Contribution Is Modest, Undermining a Core Design Claim: One of the paper’s key architectural choices is the use of 256 past frames (approximately 10.6 seconds of context) encoded by the Video Swin Transformer to capture tool motion and tissue dynamics. However, Table 5 shows that removing temporal context leads to only a moderate ASSD increase of 21.48% and HD increase of 7.73%, while removing tool specification in the text prompt causes a much larger ASSD increase of 32.81%. This suggests that the single-frame image features, combined with language conditioning, carry most of the predictive signal, and that the temporal context, despite being the most computationally expensive component, provides relatively limited additional benefit. This is not adequately discussed or explained. Recommendation: As an ablation study, provide a more thorough analysis of when temporal context helps and when it does not; for example, by comparing performance on early pre-action frames (where temporal cues should matter most) versus late pre-action frames, and by visualizing what the Swin encoder actually attends to across the temporal window.

    5.The Dice Score Anomaly Is Not Adequately Explained: Table 2 shows that SAM3 achieves a higher Dice score (0.18) than the proposed model (0.124), which is an inconsistency the manuscript acknowledge but dismiss by noting that SAM3 tends to predict nearly the entire tissue region. However, this explanation is not supported by any quantitative analysis, for instance, one would expect a comparison of predicted heatmap coverage area or a breakdown of Dice by tool-action pair. More importantly, the proposed model’s own Dice score of 0.124 is quite low in absolute terms, and the paper’s explanation, that it is caused by intensity distribution differences rather than boundary misalignment, is unconvincing without supporting evidence. A low Dice score in a safety-critical application is not something that should be casually set aside, as it reflects poor probabilistic calibration of the output heatmap. Recommendation: Provide a quantitative breakdown of Dice scores per tool-action pair, and address whether the low dice score has implications for the proposed safety-layer use case.

    6.Safety Claims Are Not Empirically Validated: The paper’s core motivation is surgical safety, specifically, using affordance predictions as a conditioning signal for learned policies and as a safety layer to trigger early robot stopping. These are strong and clinically meaningful claims. However, the paper provides no empirical validation of either use case. There is no simulation or real-robot experiment showing that a policy conditioned on the predicted affordance maps performs better or more safely than one without, and no experiment demonstrating that the predicted heatmaps can reliably detect instrument deviation from safe zones before harmful contact occurs. The entire safety argument rests on the assumption that accurate affordance prediction translates to safety benefit, which may be plausible but unverified. Recommendation: Include at minimum a proof-of-concept simulation experiment demonstrating one of the two proposed safety modes, either policy conditioning or early stop triggering, using the predicted affordance maps, even in a simplified or synthetic setting. Without any empirical grounding, the safety claims remain speculative and should be toned down or clearly labeled as future work.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    Evaluation Is Limited to Cholecystectomy With No Cross-Procedure or Cross-Dataset Testing: All quantitative results are reported on the in-distribution test split of the curated cholecystectomy dataset, with no evaluation on held-out procedures, hospital sites, or imaging conditions not seen during training. Given that the dataset is assembled from five heterogeneous sources (YouTube, Cholec80, HeiChole, CRCD, and SurgVU), it would be informative to evaluate generalization by training on a subset of sources and testing on a held-out source. This would give a clearer picture of whether the model is truly learning anatomically grounded affordance or simply overfitting to the visual statistics of a particular dataset collection.

    To strengthen the quality of this work, perform a leave-one-source-out evaluation. e.g, training without YouTube or SurgVU videos and testing exclusively on those. This will help to assess cross-domain generalization within the cholecystectomy setting, and report performance separately for each source domain in the test split.

  • 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 task formulation is genuinely novel, the benchmark dataset is a real contribution, and the ablation studies are thorough. However, the safety claims, which are the paper’s primary motivation, are entirely unvalidated empirically, the baseline comparison is architecturally unfair, the annotation protocol lacks clinical rigor, and the method is locked to a single procedure. The gap between the paper’s ambition and what is actually demonstrated is significant. I would be open to upgrading my score if safety claims are substantially toned down and a fairer baseline is added.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    The rebuttal improves the clarity of several design choices and partially strengthens the empirical grounding of the work, particularly by providing additional information on the annotation pipeline, introducing a U-Net baseline, and offering a more detailed explanation of temporal context benefits in early pre-action phases where degradation is more pronounced. The clarification that annotations were performed by trained students with clinician oversight, combined with semi-automatic tracking constraints, partially addresses concerns about annotation quality, although the lack of quantitative inter-annotator agreement metrics still leaves uncertainty about ground-truth reliability in a safety-critical setting. Similarly, the added U-Net comparison helps mitigate the concern that gains are purely due to task-specific architecture, although it still does not fully resolve whether the proposed model’s improvements stem from architectural innovations versus strong supervision and design bias, since comparisons remain limited to relatively weak baselines. The rebuttal also provides a more concrete argument for temporal modeling utility, showing stronger degradation in early pre-action scenarios, which is a meaningful refinement over the original aggregate ablation and partially addresses the concern that temporal context was under-contributing. However, several critical issues remain unresolved or only rhetorically addressed.

    First, the Dice score explanation remains unconvincing: attributing discrepancies to Gaussian post-hoc fitting versus direct prediction does not resolve the underlying concern that evaluation metrics reflect different distributional assumptions, and no additional analysis (e.g., coverage statistics or calibration curves) is provided to substantiate the claim that low Dice is not indicative of poor spatial alignment or safety-relevant performance. Second, the safety motivation remains entirely unvalidated; while the rebuttal clarifies the intended affordance interpretation, there is still no empirical evidence (simulation or real-world) that predicted affordance maps improve downstream robotic safety or decision-making, leaving the central application claim speculative; if scaled to paper acceptance, this most be mentioned as a limitation. Third, the dataset generalization argument is weak: asserting that clip-level diversity within cholecystectomy cases makes leave-one-source-out evaluation redundant is not convincing, as source-wise domain shifts (e.g., YouTube vs clinical datasets) are known to induce substantial distributional differences, and no empirical cross-source or cross-domain validation is provided to support this claim. Finally, although annotation quality is partially clarified, there is still no inter-annotator agreement or variability analysis, which is particularly important given the coarse keypoint-based representation of complex tissue regions.

    Overall, the rebuttal strengthens methodological transparency but does not fully address concerns about evaluation rigor, safety validation, and generalization, which remain substantial; as a result, the work is improved but still borderline and would likely require another revision cycle or stronger empirical validation before it can be confidently accepted.



Review #3

  • Please describe the contribution of the paper

    The paper proposes a method to predict the safer regions of instrument operation in laparoscopic cholecystectomy procedure. To aid this task the paper contributes a new dataset with image -tool and action - affordance map triplets for around 15K video snippets from around 100 cholecystectomy videos collection from various existing public datasets.

  • 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 concept of the paper as a whole is novel and is clinically very important. 2.The paper is well presented and the concept is exposed in sufficient detail. 3.The paper provides sufficient ablations to highlight the contributing components in the proposed method.

  • 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.

    As such there no major technical weakness that the reviewer could identify in this work. However, there are some minor clarifications required along with few questions and suggestions to the authors which are listed in the other comments section.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    1.It appears that the paper uses ‘video clips’ and ‘cases’ interchangeably. However, it would provide more clarity to a reader if either of them is used and if ‘cases’ is preferred then kindly define them properly. Also, a short description as to how/ on what basis these cases are extracted from videos would be helpful. If not, and cases corresponds to the complete videos then kindly mention it or use one term consistently across the paper. 2.The abstract refers to the vision encoder capturing multiple-view points. However, there is no reference to this kind of data in the training setup. The term its self is kind of ambiguous given the task and data, so unless there are specific steps taken to address the view point issues or any proof from evaluating this specific aspect it would not be fair to comment that the model implicitly address this. 3.How valid is it to assume that the affordance map is a gaussian distribution centered around the segmented polygon enclosing the object of interest? A comment on this assumption and its limitations is required. As the affordance maps vary of each object/tissue and it is important that these are 3D objects and the affordance maps would ideally vary across that 3D space of tissue. Also, they are highly influenced by the surrounding tissue. Say for example, any dissection on a tissue closer to a artery would have a low safe operating region- higher risk region in the affordance map and the risk region would be to the side of the artery. Similarly, dissection of adhesions in the abdominal cavity from previous surgery would be low risk and have more uniformly distributed affordance maps. The point that the review is trying to make is, affordance maps ideally should be 3D, take in to account nearby spatial information to have any relevant clinical impact. Having said that this work is a reasonable staring point in that direction but this limitation in framing the problem should be highlighted somewhere in the paper for providing the readers an overall idea. 4.In continuation to the above point, ideal affordance maps should be transparent and comment on the operational safety of the tissue region which is directly beneath/occlude by the tool. And these affordance maps should ideally be shown from the start of action to the end i.e., from before tool insertion to after tool retraction and transparently showing the occluded tissue regions as well. As these are ideal concepts and possible future work a short description of this point in the limitations would make the paper complete. 5.As the current work only focuses on affordance maps, the prior-art and over emphasis on robotic automation could be avoided, and that space could be used for mentioning the limitations and applications, where a brief mention of the advantage this method brings to the surgical robotic automation would be neat. 6.Similarly, overly describing method components which are not modified in significant way are just space consuming, could use that space to discuss about the performance of method or provide actual details which would help in replicating the work. 7.The text prompt description in Fig.1 and Sec.2.2(a) do not match. If there is a common text prefixed to the surgical triplet than kindly mention that in the text as well.

  • 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?

    The paper addresses a critical problem in surgical data science and a proposes a method to solve it via a dataset contribution tailored for that task. However, given the complexity of the problem there are few gaps that the review could point out to in the problem formulation, once a proper justifications to them are provided and they are addressed as limitations of the problem formulation in the paper then its could be accepted.

  • 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.

    N/A



Author Feedback

We thank reviewers for acknowledging our strengths﹣the novel task formulation (R4, R5), clinical importance (R4, R5), unique architecture and ablations (R1, R4, R5), novel dataset (R1, R4, R5), and clarity (R1, R5)﹣and for their feedback. 1.Clarity on annotation protocol (R1, R4, R5) Interaction zones are defined using task-specific keypoints﹣4 or 10 (grasping) points depending on action﹣to capture geometry accurately. We employed a two-stage annotation pipeline: medical/engineering students performed primary annotations, and audited by a core research team in consultation with clinicians. Quality/reliability is additionally enforced through a mandatory onboarding process (30 successfully completed cases). To minimize variance, we utilize semi-automatic tools (point tracking) which anchor predictions to anatomical landmarks and strictly reject “drift” (where points follow instruments instead of tissue). While affordance is annotated as uniformly safe polygons, we model them as gaussian distributions for dense prediction, where distance to boundary represents a transition toward non-affordance. Critically, this probability gradient is a mathematical modeling choice, not a reflection of varying safety levels. Therefore, the task is satisfied if the model recovers the area of the region, even if intensities deviate from the gaussian prior. 2.Affordance definition (R1, R4, R5) Tool-tissue affordance is the determined contact area on a tissue surface for a specific tool-action pair. We envision it to navigate the instrument toward regions where it can optimally and safely act to successfully execute an action. 3.Additional baseline (R1, R4) U-Net performs worse under the same training setup, validating efficiency of our architectural design. Model | ASSD | HD U-Net | 44.39 | 98.24 4.DICE score clarity (R1, R4) The DICE score discrepancy arises because SAM3’s binary output is post-hoc fitted to a Gaussian, a mathematical artifact that artificially matches the target distribution. Our model, however, predicts continuous probability values directly. While our model accurately localizes the affordance areas (low ASSD, HD), it receives a lower DICE score because it does not perfectly replicate the idealized Gaussian distribution. We argue the intensity mismatch that lowers our DICE score is a mathematical artifact from design choice, not a failure, and precise spatial localization is far more critical for surgical safety than the exact fit of a probabilistic prior. 5.Dataset collection and model generalizability (R4, R5) We distinguish between case: full recordings from unique operation and clip: temporal windows extracted from cases to represent singular tool-action pairs. Each case (that includes all its clips) is assigned exclusively to either the training / test set to prevent data leakage. Because each case is an independent table-top study, they represent a diverse distribution. This variety allows for a robust validation of model generalization, as the individual procedures present unique environmental conditions that make a leave-one-source-out evaluation redundant. 6.Role of temporal context (R4, R5) Our temporal window captures tool-tissue dynamics across viewpoints to model deformation and intent, also providing robustness against occlusions and motion blur. Our analysis shows that temporal context is critical in the early pre-action phase: a single-frame model suffers a 39% degradation in ASSD and 19% in HD. Model | ASSD | HD Baseline Model with Temporal Context on early pre-action frames | 22.93 | 85.71 Model w/o Temporal Context on early pre-action frames | 31.95 | 102.70 This proves temporal context is essential for providing reliable guidance at the start of a maneuver.



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 introduces tissue affordance prediction as a dense heatmap estimation task conditioned on tool–action language prompts, alongside a sizable benchmark curated from laparoscopic cholecystectomy videos. Reviewers agree that the task formulation is novel and clinically meaningful, and that the dataset and ablation studies represent a solid contribution. However, several issues limit the strength of the current claims. The annotation protocol lacks sufficient clinical rigor for a safety‑motivated task (coarse keypoint polygons, unclear annotator expertise, no agreement analysis), and the baseline comparisons are asymmetric, comparing a task‑specific model against general‑purpose foundation models. Absolute performance remains problematic, particularly the low Dice score, which is insufficiently analyzed despite being critical in a safety context. The work is also restricted to a single procedure with heavy class imbalance, and the claimed safety applications (policy conditioning, early stopping) are not empirically validated, leaving a gap between motivation and evidence. To strengthen the paper, the authors should ensuring the claims are fully supported by the evaluation, clarify and justify the annotation and evaluation choices, discuss the limitations of procedure scope and metrics. They should also refrain from grand safety claims and over emphasis on robotic automation unless supported by concrete validation. Authors should address each concern raised by the 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.

    While the authors clarified the annotation pipeline, added a U‑Net baseline, and better explained temporal modelling, some concerns remain unresolved. In particular, the low Dice score is not convincingly addressed, annotation reliability lacks inter‑annotator agreement evidence, generalization is unsupported (no cross-source evaluation), and the safety claims remain unvalidated. Overall, reviewers lean towards Acceptance.



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.

    The reviewers’ concerns have been successfully addressed in the rebuttal, making the work a solid contribution especially for what concerns methodological applications in CAI.



back to top