List of Papers Browse by Subject Areas Author List
Abstract
NA
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2089_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/2089_supp.zip
Link to the Code Repository
https://anonymous.4open.science/r/OCS_MAMBA-D3AE
Link to the Dataset(s)
Cholecseg8k dataset: https://arxiv.org/pdf/2012.12453
Cholect50 dataset: https://github.com/CAMMA-public/cholect50
SSG-VQA dataset: https://github.com/CAMMA-public/SSG-VQA
BibTex
@InProceedings{ShuMd_OCSMAMBA_MICCAI2026,
author = { Shuvo, Md Rezowan H. F. AND Mekala, M. S. AND Elyan, Eyad},
title = { { OCS-MAMBA: Object-Centric, Logic-Guided Approach for Surgical Action Understanding } },
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
The authors presented OCS-Mamba, a neuro-symbolic architecture for surgical video understanding targeting triplet recognition (instrument, verb, target) and action anticipation.
- 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 greatest strength of this paper is the integration of individual existing components in the creation of a solid systems. The factored Graph-Mamba design is particularly well-motivated and effective. Additionally, the RAG safety head is a meaningful step toward possible deployment making the work even more interesting. Ablation clearly show the contribution of the losses and how mamba improves long-horizon prediction, strengthening claims.
- 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.
The manuscript could benefit by considering the following aspects:
- The is no formal analysis about Mamba working better than Transformers here. You should be doing a clear evaluation evaluating attention against Mamba and demonstrate how result are better or worse. This is something already shown in the field of animal action recognition (e. g, https://www. sciencedirect. com/science/article/pii/S1574954124004977) where authors shown that Mamba is indeed better than using Transformer (linear attention/transformer) and at the same time requiring less FLOPs and computational resources. You should do something similar, that would strengthen the method.
- The manuscript would benefit in a more extensive description of the RAG safety, I believe that is a very important component, especially in the surgery field. However, there is no well evaluation/description. Does it improve accuracy or only interpretability? Can you show real failure cases prevented?
- The limitations are described in a very short line in the conclusion section “Current constraints are tailored to cholecystectomy and require adaptation for other procedures, and the fixed slot budget may not generalise to scenes with more entity types. ”. The authors should try to understand what are possible limitations of the method and its usability and write a deeper analysis. This would improve the paper’s clarity and credibility.
- Finally, I feel like you give the spotlight to Mamba in your title, however DinoV2+slot attention does a huge amount of work in our framework. I would consider to reconsider the name for publication, like this is somewhat misleading and not completely honest. I would recommend to also test some other image backbone based on Mamba if you want to keep that title.
- 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 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
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?
Although the manuscript could benefit from some deeper analysis, I believe that this is a very good contribution to the field and should be accepted.
- 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
1.The manuscript provides a mixture of object-centric features along with long-term interaction semantics modeling using Mamba based state space model (SSM) for the task of triplet recognition together with the anticipation task at short temporal horizons. 2.In order to train the model, the manuscript exploits segmentation dataset CholecSeg8k along with CholecT50 and SSG-VQA datasets for triplets and scene graph connectivity. 3.The reported results show improvement on CholecT50 cross validation and on the triplet anticipation task over SOTA triplet models augmented with simple LSTMs.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper is well organized and easy to follow. 2.The concept of utilizing object-centric knowledge has been proven beneficial for surgical workflow tasks but it is more helpful for fine-grained classification tasks. 3.The idea of using slots for tool-target together inherently allows to model interactions in the slot attention followed by factorized processing of spatial and temporal features aptly utilizing the long-term context. 4.The ontology loss idea used as a negative constraint acts as a regularizer and forces the model to adhere to allowed constraints. 5.The use of anticipation tasks together with triplet recognition mutually benefits both as reported in the results.
- 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.Slot Attention, famous in the unsupervised object discovery, generates slots that are unordered so the slots can bind to any object in the scene. How are the 21 queries handled to make sure that each query (e.g. grasper or gallbladder) correspond to the object segmentation mask of the same grasper or gallbladder. There is a mention of Presence predictor but does it mean each query is hardcoded to a specific class? Within the same paragraph Ot = ot,k is unclear and should have proper format. 2.The slots that are generated when visualized helps the readers to understand how the slot attention have captured the true object semantics - even though it is not perfect but it should roughly capture the parts of the instrument or anatomical structure to make more sense later in the graph message passing and the ontology constraints part. The lack of this information makes the manuscript a bit incomplete. 3.Overall, the manuscript should make it clear what it means by temporal. Is it short context e.g. 32/64 frames or is it all the frames of a video? 4.In the “Anticipation via State-Rollout” section, there seems to be a disconnect in the phrasing. First the future slots are mean pooled to get a single feature and then I, V, T are separately classified - does it mean no triplet head is present here? Then the manuscript mentions last-frame slots - last frame with reference to which group of frames? Is it a video clip with fewer frames? The manuscript will greatly benefit if these clarifications are properly provided. 5.The use of RAG guardrail - is it connected to the model training and/or evaluation pipeline? Because it is mentioned once before section 3, but not discussed later. The manuscript should provide clarification on this. 6.The implementation details regarding the value of feature dimension D, Mamba parameters, number of epochs or iterations, batch size, GPUs, etc. are missing which renders a big question on the reproducibility of the said experiments. 7.Minor weakness - Is there a rationale for choosing LSTM over autoregressive transformer for anticipation baselines?
- 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 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.
(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?
The manuscript proposed a method that jointly performs triplet recognition and anticipation by leveraging object-centric features with ontology based constraints and shows good improvement over baselines on both tasks. However, there are some key issues with the components proposed or it is not straightforward from the manuscript such as the slot attention module and the anticipation rollout to task head. On top of these, the lack of implementation details even though the OCS-Mamba contains a lot of important building blocks - Slot Attention, feature dimension, and Mamba, makes the reproducibility aspect of this work challenging.
- 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 #3
- Please describe the contribution of the paper
This paper introduces OCS-Mamba, a novel object-centric, logic-guided framework for surgical action understanding from video. The method specifically targets the challenges of long procedural durations and complex instrument-verb-target(IVT) interactions in robotic surgery. The proposed approach called Neuro Symbolic Surgical Graph Mamba (OCS-Mama), integrates spatial, structural, and temporal reasoning through a 3 stage pipeline: (1) semantic entity extraction using hybrid DINOv2 and Slot Attention backbone, (2) efficient modeling of long-range temporal dependencies via a linear-complexity Mamba State Space Model, and (3) enforcement of valid surgical relationships using a semantic grammar and logic-constrained loss, complemented by a RAG safety for procedural guideline validation. Empirical results demonstrate that OCS-Mamba outperforms ensemble baselines on public datasets, with notable improvements in short-horizon action anticipation and robust performance in complex surgical scenarios. The work is relevant in the field of robotic surgery and surgical training.
- 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 three stage framework moves away from pixel-centric approach and decomposes video frames into discrete semantic entities and focuses on object-centric slots to explicitly track interactions between instruments and targets. The model is hence less prone to hallucinations and can reason about physical and procedural validity of actions
- The model enforces surgical safety through a semantic grammar and logic-constrained loss, directly penalizing physically impossible or unsafe instrument interactions. The addition of a retrieval-augmented generation (RAG) safety head provides interpretable guidance aligned with procedural guidelines, which is a significant step toward clinical feasibility and real-world deployment
- Evaluation on multiple public datasets supports the generalizability of the approach. The ablation studies further validates the effectiveness of each module, supporting claims of improved memory efficiency and safer predictions.
- 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.
Overall the paper is well-written and very clear. Some comments on weaknesses:
- The current approach in terms of logic constraints and ontological matrices for indicating mechanically invalid predictions are tailored to cholecystectomy. The authors acknowledges these constraints require adaptation for other surgical procedure. This thus limits the immediate applicability of the OCS-Mamba method to broader surgical domains
- The evaluation is conducted on public datasets such as CholecT50 and CholecSeg8k, thus it is not real how well the demonstration of clinical feasibility or deployment in real surgical environments.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a novel framework for surgical action understanding that enforces clinical safety constraints, resulting in more anatomically plausible and interpretable predictions. Experimental results show that the method outperforms SOTA on public dataset. The work is supported by strong experimental and validation results, and it is overall relevant to the conference in terms of surgical AI.
- 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 are delighted that our paper has been provisionally accepted and would like to sincerely thank the reviewers and the Meta-Reviewer for their thoughtful, relevant, and constructive feedback. We are especially pleased that the reviewers recognised the novelty and practical relevance of our object-centric methods.
Reproducibility and Implementation Details (R2, Meta-Reviewer): We agree with the reviewers that some of the implementation details were missing. In the camera-ready version, we will clarify the key architectural and training settings. We also intend to make the source code and full configuration details available upon acceptance.
Slot Attention Clarification (R2, Meta-Reviewer): We thank the reviewers for pointing out that the slot semantics require clearer explanation. The 21 slots are not manually hard-coded to specific classes. Rather, they are object-centric proposals whose semantics are learned through segmentation supervision, presence prediction, and IVT recognition losses. The presence predictor does not impose a fixed slot-to-class assignment, but encourages the slot set to represent relevant instruments and anatomical targets. We use Slot Attention as a structural inductive bias (object-like decomposition into K competing representations), but replace its unsupervised reconstruction objective entirely with task-specific surgical supervision. We will refine the description in the revision and also clarify the notation around $O_t$.
Mamba Temporal Modelling (R1, Meta-Reviewer): Our motivation for adopting Mamba stemmed from the long procedural durations and dense temporal dependencies in surgical videos. Mamba offers linear complexity, efficient state propagation, and improved long-context scalability. The submitted experiments already include relevant Transformer/attention-based baselines, including FUTR, MT4MTL, TERL, and CurConMix. We clarify that OCS-Mamba does not simply replace Transformers with Mamba; DINOv2 remains the visual backbone, while Mamba is used within the factored Graph-Mamba stage for efficient temporal-state propagation over object-centric slot features. We will make this design motivation clearer in the camera-ready version. The title reflects the role of Mamba-based temporal reasoning within the broader object-centric and logic-guided framework.
Anticipation Rollout and RAG Safety Module (R1, R2): Temporal modelling is performed over sequential frame groups rather than full videos simultaneously, where future slot states are recursively propagated before IVT decoding. We acknowledge that this section currently lacks sufficient clarity and will revise it accordingly.
The RAG safety module is used only as an inference-time interpretability and procedural verification component. It is not used during training or optimisation and does not contribute to the reported performance metrics. Its role is to provide post-hoc safety-oriented feedback for predicted actions. A more extensive evaluation of RAG-guided safety verification, including clinical validation and failure-case analysis, remains important future work.
Generalisability and Limitations (R1, R3): We acknowledge that the ontology constraints are currently tailored to cholecystectomy, the slot budget is fixed, and the evaluation is based on public datasets. These points, together with the need for prospective clinical validation before deployment, will be briefly emphasised in the discussion of future work and limitations.
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.
While two reviewers are strongly in favor of acceptance due to the framework’s impressive systems-level integration and innovative logic-guided constraints, R2 explicitly disagrees with a recommendation to reject. All reviewers recognize the paper’s strengths, specifically the effectiveness of the factored Graph-Mamba design for long-horizon prediction, the well-motivated use of object-centric slots, and the clinical potential of the RAG safety head for interpretable surgical guidance. However, there are concerns regarding a critical lack of implementation details—such as Mamba hyperparameters and feature dimensions—which severely hinders reproducibility. Furthermore, the authors must clarify the mechanics of the Slot Attention module regarding how unordered slots bind to specific semantic entities, provide a formal comparative analysis justifying the choice of Mamba over Transformer baselines, and offer a more robust evaluation of the RAG safety head’s ability to prevent real failure cases while discussing the limitations of generalizing these logic constraints beyond cholecystectomy.
