List of Papers Browse by Subject Areas Author List
Abstract
Surgical Scene Graphs abstract the complexity of surgical operating rooms (OR) into a structure of entities and their relations, but existing paradigms suffer from strictly dyadic structural limitations. Frameworks that predominantly rely on pairwise message passing or tokenized sequences flatten the manifold geometry inherent to relational structures and lose structure in the process. We introduce TopoOR, a new paradigm that models multimodal operating rooms as a higher-order structure, innately preserving pairwise and group relationships. By lifting interactions between entities into higher-order topological cells, TopoOR natively models complex dynamics and multi-modality present in the OR. This topological representation subsumes traditional scene graphs, thereby offering strictly greater expressivity. We also propose a higher-order attention mechanism that explicitly preserves manifold structure and modality-specific features throughout hierarchical relational attention. In this way, we circumvent combining 3D geometry, audio, and robot kinematics into a single joint latent representation, preserving the precise multimodal structure required for safety-critical reasoning, unlike existing methods. Extensive experiments demonstrate that our approach outperforms traditional graph and LLM-based baselines across sterility breach detection, robot phase prediction, and next-action anticipation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2412_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/2412_supp.zip
Link to the Code Repository
https://github.com/wngTn/TopoOR
Link to the Dataset(s)
MM-OR dataset: https://github.com/egeozsoy/MM-OR
BibTex
@InProceedings{WanTon_TopoOR_MICCAI2026,
author = { Wang, Tony Danjun AND Kim, Ka Young AND Birdal, Tolga AND Navab, Nassir AND Bastian, Lennart},
title = { { TopoOR: A Unified Topological Scene Representation for the Operating Room } },
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 TopoOR, a novel framework for surgical scene representation that moves beyond traditional “dyadic” (pairwise) models to capture the complex, group-based interactions of the operating room. By modeling the OR as a Combinatorial Complex, the authors allow for “polyadic” relationships where multiple entities, such as surgeons, robotic tools, and patients, are treated as a single, unified topological unit.
The approach utilizes a higher-order attention mechanism that processes multimodal data (3D motion, audio, and kinematics) without collapsing it into a simplified “semantic bottleneck,” thereby preserving the precise geometric and topological structures necessary for safety-critical reasoning.
- 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.
Higher-Order Expressivity: The shift from dyadic (pairwise) to polyadic (group-based) modeling is a major contribution. By using Combinatorial Complexes, the model can represent complex group interactions (e. g. , a surgeon, an assistant, and a robotic tool working in unison) as a single mathematical unit.
Avoidance of Latent “Averaging”: The paper introduces a higher-order attention mechanism that keeps different data types (3D poses, audio, robot kinematics) distinct rather than forcing them into a single, shared latent representation. This ensures that modality-specific details aren’t “washed out” during the fusion process.
Hierarchical Reasoning: The architecture allows information to flow through different “ranks” of cells (from individual objects to small groups to large teams).
Theoretical Grounding: The paper provides a formal mathematical bridge between Topological Deep Learning and Surgical Scene Graphs.
- 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.Data Sparsity and Statistical Significance of Higher-Order Relations The manuscript proposes a highly sophisticated topological model for the surgical environment. A primary concern lies in the data requirements for such a framework. While the MN-OR dataset is currently the most comprehensive for this task , it may still be insufficient to validate a polyadic approach. In a standard dyadic model, pairwise events occur frequently enough to provide statistical robustness. However, as the model moves toward higher-order Combinatorial Complexes, the number of unique combinations increases while the number of occurrences per combination naturally decreases. I am concerned that the “long tail” of complex group interactions in this dataset is too sparse to truly evaluate the performance and generalization of the proposed higher-order attention mechanism. 2.Breadth of Evaluation The evaluation is currently limited to a single dataset. While this might be the only relevant dataset, this singular focus limits the validation of the model’s robustness across different surgical specialties or OR configurations. Furthermore, the overall quantitative evaluation appears to be based on a limited number of test cases, which may not fully capture the model’s reliability in edge-case safety scenarios. 3.Automated vs. Manual Hierarchy Construction The methodology behind the “lifting” process requires further clarification. It is not entirely clear if the higher-order combinations (the $k$-cells) are generated fully automatically through spatial/temporal heuristics, or if there was manual intervention involved in selecting which specific hierarchies to prioritize. 4.Supplemental Material and Qualitative Evidence While the supplemental video is visually impressive, its contribution to the scientific validation of the model is unclear. I would appreciate it if the authors could provide more information on it.
- 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
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?
The surgical domain is characterized by high levels of spatial and relational complexity. The authors address this by presenting a novel framework grounded in a rigorous mathematical foundation, utilizing topological deep learning to model higher-order group interactions. While the evaluation faces inherent limitations due to the availability of large-scale annotated datasets for such complex tasks, the study represents a significant and innovative step forward. By moving beyond standard graph-based representations, this work provides a compelling contribution to the advancement of surgical scene understanding.
- 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
Review #2
- Please describe the contribution of the paper
The paper proposes a strategy to derive the relations between the entities in OR scene and a novel way to represent these relations in a hierarchical fashion building on the concepts of scene graphs. Different multi-modal data from the entities are aggregated based on their hierarchical relations using a proposed attention network which produces pooled features of a surgical scene suitable for downstream applications.
- 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 mathematical background of the method is clearly explained. 2.Ablations in Tab.3 clearly show the merits of the proposed frameworks in handling various modalities. 3.Overall the work is novel and has sufficient ablations to support the merits.
- 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 are no significant drawbacks but few issues need to be addressed to maintain the scientific standards. The details are provided in the additional comments section along with few suggestions related to the work.
- 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
1.In Tab 1.performance of the models for sterility breach is same for all latent and relational model. This is kind of unusual and ideally the reason for this should have been discussed in the paper. The reviewer would like to point out to look at this aspect consider the Hasse diagram in Fig3 which leads to the conclusion that for this task just pair wise relations suffice and any additional information such as un-/structured / hierarchical relations do not matter. 2.Similarly, other tasks might need different levels of information. Hence just like the ablation in Tab3, in the future extend version of this work it would be nice if an ablation is performed to see which rank’s relations i.e., which level of hierarchical relations are needed for which tasks. 3.Instead of Fig 1, Fig 2 should have been used as an opening figure. This is because the papers starts with higher-order relations right in the abstract but a proper clarity as to what is higher-order and what other orders exists comes only after seeing Fig2, the text also addresses this at a later point but unclearly. Hence, it is suggested to move to Fig 2 to Fig1 and for better clarity few suggestions are provided for both Fig2 and Fig 1 below. 3.Suggestions to Fig2: a. Block III, the hierarchy is visible, so just adding that rank 2 nodes and edges are higher-order relations would be easy to understand in relation to abstract and introduction. 4.Suggestions to Fig1: a. Keep the center 3d diagram, but push it to preferably left. b. Instead of multiple 3D views (the large cluster of hierarchical graphs makes the visualizations messy) pick one view and show all three ranks cells for that view on the size. If space permits pick 2 view and show them side by side by stacking the ranks vertically. Note that for larger ranks there is no need to show the entire entity a 3D version of the graph with spherical nodes would suffice. Remember that a reader needs to be walked through the paper with specific visual cues to understand the paper and each figure/table need to be placed and designed for this purpose alone. 5.There are few minor issues: a. In abstract is it “modality-specific features throughout” or “through”. b. The equations (1,2,3) and definitions are not according to general mathematical notations of tensors and variables. Please refer to any standard DL text and update the notations, as it would be easy for the reader to follow the text and relate to the type of variable/tensor/data in content. c. It seems like the paper was written in last minute with a lot of LLM outputs, kindly remove the ‘-‘s generated from LLMs, as they are un-necessary and do reflect on the scientific writing standards.
- 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 concept is novel and the paper has all factors required to prove the working on the concept along with its superiority to prior works. However, there are few minor issues which must be addressed before the paper is publised.
- Reviewer confidence
Very confident (4)
- [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
Review #3
- Please describe the contribution of the paper
The paper proposes a method for surgical scene reasoning via higher-order structures, hierarchically modeling features into several ranks. Via Higher Order Attention Networks, features are exchanged between instances and through ranks, resulting in improved performance for various downstream classification tasks.
- 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 presented approach brings technical novelty while achieving state-of-the-art performance in next action prediction and robot phase recognition. 2.The ablations show how the method utilizes all input modalities for improved estimations, including the introduction of temporal edges.
- 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.Although I do understand that the MM-OR benchmark is rather new, the presented results are sparsely presented. Comparison with a wider variety of methods, reporting of multiple metrics, scores per class, and standard deviations will make it easier to interpret the impact of the presented results. 2.While a limitation is discussed that addresses surgical scene understanding benchmarks, it would be good to discuss/show the limitations of the presented method as well. 3.For example, it seems that the method requires priors by predefined domain-specific instances and relations. To get the full picture on the applicability of the method, I think it would be good to elaborate on the potential challenges that this poses.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 presents a novel and relevant method for surgical scene understanding. While state-of-the-art performance is shown on various benchmark tasks, the presented results could be more elaborate and convincing.
- Reviewer confidence
Somewhat confident (2)
- [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 the reviewers for the constructive comments and respond to the main points below.
Statistical capacity (R1, W1): We agree that sparse higher-order events are a central challenge in OR modeling. Our construction is designed to mitigate this issue by using higher-order structure as an inductive bias rather than treating every polyadic tuple as an independent class. This is particularly important in low-data regimes: known geometric and clinical priors constrain the hypothesis space, preserve modality-specific structure, and reduce the need to relearn domain relations purely from data. In HAT, the projections W_Q, W_K, and W_V are shared across cells of all ranks; only a small learnable rank-pair bias, b_{rk(y),rk(x)}, modulates source-target routing. Capacity, therefore, scales with the number of cell instances seen in training, not with the count of unique polyadic combinations.
Cell construction (R1, W3): Per §2.3, all k-cells are produced deterministically from a small, fixed rule set: kinematic trees for intra-entity skeleton edges, a spatial-proximity threshold for inter-entity edges, and a compact list of domain-specific semantic links (e.g., Technician-Robot). No per-scene or per-sample manual intervention is involved.
Sterility-breach parity (R2, c1): We agree with R2 that sterility breaches primarily depend on spatial proximity and therefore do not require the full higher-order hierarchy. As stated in §2.4, sterility breach is computed by “rule-based spatial heuristics directly over the 3D entities,” not by a learned task head. All 3D-grounded methods, therefore, share the same evaluation and produce identical scores; the parity reflects the evaluation protocol. Per R2’s comment 2, a per-rank ablation isolating each rank’s contribution per task is a valuable direction for extended work.
Method limitations and prior dependence (R4, W2-W3): Beyond the dataset constraints, the principal methodological limitation is dependence on upstream perception (pose, depth, segmentation); errors there propagate into the cell structure. This constraint will ease as OR perception models improve. The proximity threshold can be replaced by a learnable soft kernel with no architectural change. The semantic-link prior is small and intentional, but new procedures (e.g., laparoscopy: endoscope/tissue, tool/tool, surgeon/console) require re-specifying the link set. We will expand the Limitations section accordingly.
Evaluation breadth (R4, W1; R1, W2): Cross-institution and cross-procedure validation is important. MM-OR, however, is currently the only multimodal OR benchmark at this scope, and we follow the Macro-F1 protocol of [20] for a like-for-like comparison with the MM2SG baseline, which itself does not report per-class scores. Within that constraint, Tab. 3 isolates each modality’s contribution, and the three tasks cover spatial safety, current-state recognition, and temporal anticipation. Cross-procedure validation remains a key direction as comprehensive OR datasets emerge.
Supplemental video (R1, W4): The video visualizes the constructed combinatorial complex on a single time frame: 3D bounding boxes, human skeletons, and rank-1/rank-2 cells rendered as edges and faces. It conveys the explicit hierarchical structure on which the model operates, which is difficult to view from static images.
Writing and notation (R2, c3-5): The notation in Eqs. 1-3 combines standard topological deep learning conventions for cells, ranks, and incidence neighborhoods with standard query-key-value attention. On punctuation, we use em dashes primarily for appositive or parenthetical clauses, though we will reduce nonessential instances in the camera-ready version. To keep the MICCAI revision focused, we retain the current figure layout and will consider R2’s restructuring suggestions in a subsequent journal extension.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 presents TopoOR, a framework for surgical scene understanding that models higher‑order, polyadic interactions in the operating room using Combinatorial Complexes and higher‑order attention over multimodal signals. Reviewers agree that the approach is novel and well grounded, noting that it meaningfully advances beyond traditional dyadic scene representations and enables hierarchical reasoning across multiple relational ranks. Empirical results and ablations indicate that leveraging higher‑order structure and multimodal information improves performance on several downstream tasks. Reviewers also point out limitations, including evaluation on a single, sparse benchmark, relatively limited quantitative breadth, and the need for clearer exposition of how higher‑order relations are constructed and what assumptions they impose. Nonetheless, these constraints are largely attributed to current data availability rather than deficiencies in the proposed formulation, and the work is regarded as a strong contribution to surgical scene modelling.
