List of Papers Browse by Subject Areas Author List
Abstract
Accurately recognizing surgical action triplets, comprising instrument, verb, and target, is critical for advancing intelligent healthcare systems. However, it remains challenging due to the complex temporal dynamics and intricate semantic dependencies involved. Existing methods often fall short in simultaneously capturing long-range temporal dependencies and local multi-scale features, and they typically overlook explicit semantic correlations among triplet components. To address these limitations, we propose a multi-stage Dilated Convolution-Mamba Network (DCM-Net) for surgical action triplet recognition. Specifically, we propose a Dilated Convolutional Mamba (DCM) block with a parallel architecture: one branch employs a State Space Model (SSM) to capture long-range dependencies, while the other utilizes dilated convolutions with multiple dilation rates to extract multi-scale local temporal features. Moreover, to enhance semantic consistency, we present a Task-aware Graph Fusion (TGF) module, which comprises a two-stage graph structure. Each stage is constructed with corresponding nodes, enabling the model to learn both intra-cluster correlations and inter-layer relationships between coarse and fine semantic representations. Extensive experiments on the CholecT45 dataset demonstrate that our method effectively captures both temporal evolution and semantic interdependencies, achieving state-of-the-art performance.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2862_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/qikuo10/DCM-Net
Link to the Dataset(s)
N/A
BibTex
@InProceedings{MenQin_MultiStage_MICCAI2026,
author = { Meng, Qingke AND Tang, Zhangyong AND Zhu, Xuefeng AND Wu, Xiao-Jun AND Zhou, Tao},
title = { { Multi-Stage Dilated Convolution-Mamba Network for Surgical Action Triplet Recognition } },
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
(1) The authors propose Dilated Convolution–Mamba block to solve the dilemma between global context modeling and local feature retention. (2) The authors propose a task-aware graph fusion module to model semantic correlation among triplet components.
- 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 structured and easy to read. (2) The model architecture is clear and well presented in Fig. 1. (3) Model’s superior performance on open dataset CholecT45.
- 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 proposed DCM block appears to be a general-purpose context modeling module and lacks surgical domain-specific insights. (2) No comparison between the proposed Task-aware Graph Fusion Module and other fusion methods (for example, transformer-based). It is unclear whether the performance improvement comes from the effectiveness of GCN or simply from an increase in model parameters. (3) If the authors want to prove that their model produces higher confidence in predicting the correct categories, more examples should be provided in Fig. 2. (4) Typo for title of Fig. 1(b) “Diated” -> “Dilated”
- 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.
(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?
Overall, the paper is well organized. The method is clearly presented and demonstrates strong performance. However, the model appears to offer limited novelty, and the experiment is not sufficient.
- 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 proposed method offers very limited novelty and surgical insights.
Review #2
- Please describe the contribution of the paper
1.The manuscript presents a novel approach to jointly capture long-term temporal semantics and local features at multiple scales, followed by a graph based fusion approach ultimately building robust features for the triplet recognition task. 2.The method shows improvement on the cross-validation setting of CholeT45 signifying the importance of joint modeling of local and global features.
- 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 relatively easy to follow. 2.The design of MS-DCM block is straightforward with dilated convolution and Mamba modules capturing short multi-scale and long temporal features effectively utilizing their complementary feature semantics. This serves as the foundation for the next set of blocks towards triplet recognition. 3.Graph fusion further modulates features by enforcing interaction between the triplet and their components. 4.The design choice is well motivated and reported results show improvement over contrastive based approach in CurConMix.
- 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.In section 2.1, it has been stressed that the first DCM block has N layers and the rest M layers. The manuscript does not clarify and provide rationale on this specific choice as in the experiment settings N is 6 and M is 5. 2.The Task-aware graph fusion can be improved as it is hard to understand the whole sequence of feature modeling starting with the GCN Fusion paragraph. Specifically, it should first start with how the node is constructed, followed by what one cluster looks like and then so on.
3.The manuscript should also stress why fusion of multi-scale features are selected as a cluster. The edges could easily be set up between nodes belonging to two different clusters (F3 and F4) but with the same label (e.g. V). The choice of using fusion features is not clear and then from the ablations it is not clear if it contributes meaningfully or is it just over-engineering.
4.The manuscript mentions that Pn goes to classification head. It should clarify which Pn is used for classification - is it the fused features from F3 and F4 as there are other fused features present as well and the Figure 1 suggests that three fused features go to the classification head. As mentioned, this section needs to be properly motivated and written clearly as the flow is a bit disconnected.
5.Task-aware classification head is just multi-task learning and should not be claimed as a newly introduced strategy as this has been employed in Rendezvous, RiT, MT4MTL-KD, TDN, etc.
6.Even though the method scores highly on the cross validation set of CholecT45, the strength of the model could have been evaluated further on the challenge test of CholecTriplet2021.
- 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
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.
(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 manuscript presents a method to exploit multi-scale local features and global features through parallel branches of dilated convolution and SSM blocks. The results show improvement on just the cross validation set of CholecT45.However, there are sections where more clarity in the design choice, rationale and the overall flow needs to be properly highlighted.
- 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 authors provide compelling rebuttal and further offers sufficient justification to the queries raised. The authors agreed on improving the Task-aware graph fusion method in the manuscript for better readability.
Review #3
- Please describe the contribution of the paper
This paper proposes DCM-Net for surgical action triplet recognition. The method consists of three main parts: a frozen Swin Transformer backbone for spatial feature extraction, a multi-stage Dilated Convolution-Mamba (DCM) temporal module for modeling both long-range and local temporal dependencies, and a Task-aware Graph Fusion (TGF) module to explicitly model the semantic relationships among instrument, verb, target, and triplet representations. The method is evaluated on the CholecT45 benchmark under the official 5-fold split, and the reported results show consistent improvements over prior methods.
- 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.Surgical triplet recognition is highly relevant for surgical workflow understanding, and the paper clearly identifies three core difficulties: modeling long-range temporal context, preserving local temporal details, and capturing semantic dependencies among triplet components. The proposed architecture is well aligned with these motivations. In particular, combining an SSM/Mamba branch with a dilated convolution branch is a reasonable design to balance global and local temporal modeling, and the graph-based fusion module is a sensible attempt to explicitly encode interactions between instrument, verb, target, and triplet-level semantics.
2.In Tab. 1, the proposed method improves over prior methods across multiple settings, and the best model with SwinB(384) achieves the highest reported APIVT = 42.0, outperforming CurConMix-Ens at 40.7 and TERL-Ens at 38.5.The gains over CurConMix are also fairly consistent across backbone sizes, which supports the claim that the proposed temporal and semantic modeling modules contribute beyond the frozen spatial backbone.
3.The paper includes an ablation study isolating the impact of the temporal modeling block and the graph fusion module. Table 2 shows that replacing the baseline with MS-DCM improves APIVT from 38.8 to 40.3, and adding TGF further improves it to 41.2.This provides some evidence that both components contribute meaningfully to the final performance.
- 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 method novelty is somewhat limited. The proposed method is largely a combination of several existing design ideas: a frozen Swin-based spatial backbone, a hybrid temporal module combining Mamba/SSM with dilated convolutions, and a GCN-based semantic fusion module. While the combination is sensible and appears effective, the overall contribution feels more like a strong architectural integration than a fundamentally new formulation for surgical triplet recognition.
2.All quantitative evaluation is performed on CholecT45 only. For a method whose main claims are about temporal modeling and semantic dependency modeling in surgical triplet recognition more broadly, validation on a single benchmark makes it difficult to judge generalization. This is especially relevant because the method is presented as a general triplet recognition approach, not as one tailored specifically to one dataset.
3.The paper also motivates itself partly through the limitations of Transformers in long-sequence modeling and the advantages of Mamba-style temporal modeling, but it doesn’t provide any efficiency or complexity analysis. There are no results on runtime, memory, FLOPs, or parameter overhead for the proposed temporal module. Since one of the main motivations for using Mamba is its linear complexity and long-range modeling capability, I would have expected at least some quantitative evidence that the proposed design is more efficient or scalable than relevant alternatives. At present, the benefit is only indirectly argued, not demonstrated.
4.The paper’s discussion of class imbalance and long-tail recognition is not fully backed up by the experiments. The introduction emphasizes severe long-tailed distributions and introduces a weighted multi-task loss to address this, but the experimental section does not include any tail-class analysis, per-class breakdown, or comparison showing that the proposed loss particularly helps rare but safety-critical classes.
5.The method relies heavily on a frozen CurConMix/Swin-based spatial backbone, and the main contribution is effectively in the temporal-semantic head on top of that backbone. This isn’t necessarily a problem, but it does mean the paper should be framed more explicitly as improving the temporal-semantic modeling stage rather than as a full end-to-end rethinking of triplet recognition.
- 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.
(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 think this is a solid paper with strong benchmark performance. The empirical gains are real, the architecture is coherent, and the ablations support the usefulness of the main modules. At the same time, the paper is held back by limited conceptual novelty, validation on only one dataset, and missing analyses on efficiency and long-tail behavior, despite these being part of the paper’s motivation.
- 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.
I thank the authors for the detailed rebuttal. The response addresses several of the main concerns raised across the reviews. In particular, the additional comparison of TGF with MLP and Transformer variants helps clarify that the gain is not simply due to more parameters. The explanation and ablation of the DCM layer settings also make the architectural choice more convincing.
I also appreciate the additional results on CholecTriplet2021 and the efficiency comparison with full-attention and window-attention Transformers. These results partially address my concerns about generalization beyond CholecT45 and the lack of complexity analysis. The ablation of the frequency-weighted loss also provides some support for the claim on handling class imbalance.
However, I still think the methodological novelty is moderate. The method remains largely a strong integration of existing components, including a frozen Swin backbone, Mamba/dilated convolution temporal modeling, and graph-based semantic fusion. Some explanations, especially for the TGF design and task-aware heads, should have been clearer in the original paper. The improvement on CholecTriplet2021 is also useful but still relatively modest.
Overall, the rebuttal improves my assessment. The paper presents a coherent method with strong benchmark performance, and most of my concerns have been at least partially addressed. I would therefore keep my recommendation as weak accept.
Author Feedback
Thanks for comments.
To R1 1.DCM: The DCM block uses a general-purpose architecture, while it is specifically tailored to capture surgical-domain temporal patterns, from long procedural dependencies to short instrument-tissue interactions.
2.Comparison with TGF: We compare the TGF module against an MLP with a similar parameter count and a Transformer variant under the same setup. The MLP achieves 40.7% mAP, slightly behind our 41.2%. The Transformer variant yields 40.4% mAP but requires 11.86M parameters, more than double our 5.28M. These results demonstrate the accuracy and efficiency advantages of our module.
To R2 1.DCM Layers: The asymmetric choice of N=6, M=5 stems from the complementary strengths of dilated convolution and Mamba. The six initial dilated layers provide a receptive field of 127, covering full action segments while complementing Mamba’s global semantics. The three subsequent five-layer refinement stages smoothly expand the receptive field from 127 to 313 for longer actions, without conflicting with Mamba’s global modeling. An ablation study reveals an inverted-U trend in mAP: 39.7% (3,2), 41.2% (6,5), 40.5% (9,8), and 40.2% (11,10), confirming (6,5) as the optimal setting.
2.Presentation of TGF: We will update texts in the final version.We construct the graph in three steps. First, from each of the three feature sequences (shallow, deep, fused), four 1×1 projectors extract four semantic nodes: Instrument, Verb, Target, and Triplet, resulting in twelve nodes. Second, nodes from the same sequence form a cluster with full intra-cluster connections, ensuring cross-component consistency per scale. Third, inter-cluster edges connect nodes of the same semantic type across clusters, forming chains (e.g., Instrument nodes across shallow, deep, and fused scales), where the fused nodes serve as anchors to absorb complementary information. After two graph convolution layers, the updated fused nodes go to separate residual-connected classification heads, and the Triplet node of the fused sequence is projected back for the next FPN stage.
3.Multi-scale features as a cluster: A fused node cluster is necessary-separate f_3 and f_4 clusters only exchange semantics without real fusion. We build a coarse f_fusion (f_3 + f_4) and let GCN refine it. Without it, raw clusters lack distinct triplet semantics. Adding the fusion improves mAP from 40.9% to 41.2%, confirming its anchor role.
4.Classification head: All GCN fusion outputs go to classification heads, but only the lowest-level (f1+P1) features are used for final prediction; the rest act as auxiliary supervision during training. Unlike prior methods that share undecoupled features across heads, our task-aware heads receive decoupled, semantically distinct features from the GCN module.
5.CholecTriplet2021: Our model obtains 37.2% mAP against CurConMix’s 36.0%, indicating the effectiveness on this dataset.
To R3 1.Novelty: 1) We propose the DCM block, a Mamba-dilated convolution hybrid that balances global context and local features, to capture surgical temporal patterns, unlike existing CNN or Mamba models. 2) We propose a task-aware graph fusion module that explicitly models semantic correlations among triplet components while aligning cross-granularity features during fusion. This novel design enables the model to learn both intra-cluster correlations and inter-layer relationships between coarse and fine semantic representations. 3) Results validate its effectiveness.
2.CholecTriplet: Our model obtains 37.2% mAP on CholecTriplet 2021, surpassing CurConMix’s 36.0%.
3.Efficiency: The basic Transformer (full-attn) 37.86M params, 54.16 GFLOPs; window-attn Transformer (win=64) 37.87M, 39.04 GFLOPs; our DCM module 34.21M, 34.04 GFLOPs. DCM shows clear efficiency advantage, indicating the involved Mambda’s linear complexity.
4.Frequency-Weighted Loss: mAP improves from 40.5% to 41.2% with FWL, indicating its effectiveness to handle class imbalance.
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.
The rebuttal should address four decision-affecting concerns. First, limited methodological novelty: the DCM block, GCN fusion, and task-aware classification head are recombinations of components already used in the literature; the rebuttal should make the specific contribution explicit. Second, the Task-aware Graph Fusion needs comparison against alternative fusion designs and a parameter-matched control, since the gain may stem from added capacity rather than the GCN itself. Third, the paper provides no efficiency analysis (runtime, FLOPs, parameters) and no per-class or tail-class evidence to support the long-tail motivation. Fourth, evaluation on the official CholecTriplet2021 challenge test split would substantiate the SOTA claim beyond CholecT45 cross-validation, and the rationale for the asymmetric stage depths in the DCM block should be provided.
- 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 paper combines a dilated convolutional Mamba block with a task-aware graph fusion module for surgical action triplet recognition. The reviewers raised concerns about novelty, the need for a parameter-matched control to isolate the contribution of graph fusion, the lack of an efficiency analysis, and the lack of evaluation on the official CholecTriplet2021 split. Two reviewers endorsed acceptance after the rebuttal; one maintained Reject on novelty and surgical insights. The contribution is a careful architectural integration rather than a fundamentally new formulation, but the rebuttal makes the integration choices defensible.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
After rebuttal, two reviewers are satified with the response but one reviewer thinks the novelty is limited. The pros include the clear model design, good performance and convincing ablation study. The cons include limited novelty and experimental verification. The rebuttal adds new experimental results which is not allowed. So the recommendation is Reject.
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.
The rebuttal addresses several decision-relevant concerns by adding parameter-matched TGF comparisons, CholecTriplet2021 results, DCM efficiency analysis, and justification for the asymmetric DCM depth setting. Although the methodological novelty remains moderate and the design is largely an effective integration of existing temporal and graph modules, the empirical performance and post-rebuttal reviewer balance support acceptance as a borderline/weak-accept paper.
