List of Papers Browse by Subject Areas Author List
Abstract
Online surgical phase recognition aims to identify temporal segments in long surgical videos under real-time constraints. Unlike generic action recognition, it faces challenges such as redundancy, noisy frames, abrupt transitions, and low inter-class variability. We propose PhaseMamba, a two-stage framework that progressively learns from short clips to full procedures. Stage 1 pretrains a local phase learner to capture fine-grained temporal patterns and transitions, while Stage 2 integrates these representations to train an end-to-end model for efficient online inference. Both stages leverage Mamba-based architectures for scalable long-sequence modeling. A phase-order refinement strategy is introduced to further improve robustness under low-confidence predictions. Experiments on Cholec80 and AutoLaparo demonstrate state-of-the-art performance.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1619_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)
Cholec80 dataset: https://camma.unistra.fr/datasets/
AutoLaparo dataset: https://autolaparo.github.io/
BibTex
@InProceedings{MohSha_PhaseMamba_MICCAI2026,
author = { Mohamed, Shaheer AND Gammulle, Harshala AND Fernando, Tharindu AND Sridharan, Sridha AND Lucey, Simon AND Fookes, Clinton},
title = { { PhaseMamba: Two-Stage Learning from Short Clips to Full Procedures for Online Surgical Phase 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
In the phase recongition task in surgical video analysis, the paper proposes a two-stage framework that progressively learns from short clips to full procedures. Stage 1 pretrains a local phase learner to capture fine-grained temporal patterns and transitions, while Stage 2 integrates these representations to train an end-to-end model for efficient online inference. Both stages leverage Mamba-based architectures for scalable long-sequence modeling. Experiments on Cholec80 and AutoLaparo datasets demonstrate state-of-the-art performance in the phase recognition for surgical videos.
- 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.
- This paper proposes a novel two-stage framework for online surgical phase recognition that progressively learns from short local windows to full procedures based on Mamba. The paper further introduces a phase-order refinement strategy to correct low-confidence mispredictions during real-time inference.
- The authors compare their proposed method with existing state-of-the-art techniques using Cholec80 and Autolaparo, two representative datasets for surgical video analysis, and find that theirr proposed method achieves state-of-the-art 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.
- Although the authors claim that the proposed method is online and real-time, I (reviewer) believe that this assumption breaks down when the Phase-Order Refinement Module is used. Section 3.3 states that at time t, Equation (9) is used to calculate the optimal sequence Y* from time 1 to time t using the Viterbi algorithm, based on the data (class probabilities and class transition probabilities) from time 1 to time t. In other words, as time progresses, the past phase determination results are updated. This is considered neither online nor real-time. Although it is not explicitly stated how the final phase determination result is obtained, it appears that the Y* value reached when the video sequence reaches its final frame is used as the phase determination result. In that case, the comparison with existing methods shown in Table 1 is not fair. For example, the results of the proposed method without the Phase-Order Refinement Module, as shown in Table 4, are Accuracy 92.2%, Precision 91.1%, Recall 91.8%, and Jaccard 84.2%. When compared with the results of the existing methods in Table 1, these cannot be considered state-of-the-art (SOTA).
- Section 3.1 describes the details of Stage 1, and Section 3.2 describes the details of Stage 2.However, both sections merely describe the model architecture; there is no logical explanation of why this architecture is adopted or why it yields the characteristics of the proposed method claimed by the authors.
- 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
- Figure 1, which illustrates the proposed method, is inaccurate. Specifically, while the figure suggests that the Local Phase Prediction Module in Stage 2 corresponds to the purple-hatched area in Stage 1, based on the explanation in Section 3.2, it appears that the correct interpretation is that the purple-hatched area in Stage 1 has had the Phase Prediction Head and Transition Head removed.
- In Sections 3.1 and 3.2, the symbols are not explained, making the results difficult to reproduce. Examples include “Mamba->”, “Mamba<-” and “flip” in Equation (2), the symbol (circle + center dot) in Equation (3), “σ” in Equation (5), and “Permit” in Equation (9).
- 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?
Although the authors claim that the proposed method is online and real-time, the use of the Phase-Order Refinement Module undermines this assumption, suggesting that it is neither online nor real-time. In that case, the comparison with existing methods shown in Table 1 is not fair, and the results of the proposed method cannot be considered state-of-the-art.
- 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.
My main concern was that it appeared the Phase-Order Refinement Module did not adhere to the online requirement. However, based on the rebuttal, which clearly states that the Phase-Order Refinement Module does comply with the online requirement and that this will be documented in the revised version, I conclude that the above concern has been addressed.
Review #2
- Please describe the contribution of the paper
The authors present a two-stage framework for online surgical phase recognition. Stage 1 learns transition-aware local temporal representations from short clips, while Stage 2 fuses these features with frame-level representations. They further introduce a phase-order refinement module to improve temporal coherence under ambiguous conditions.
- 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 authors provide detailed related work discussion. (2) The authors provide detailed comparison experiments on two open datasets and ablation studies.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
(1) Contributions 2 and 3 might not be presented as separate contributions, as they appear to further elaborate on Contribution 1 rather than stand as fully independent contributions. (2) Suggest to rotate the Stage 2 diagram in Fig. 1 so that the text is properly oriented for readability. (3) As an online surgical phase recognition model, the authors did not report the processing speed even though they highlight the importance of real-time prediction. (4) In Stage 1, why frame-level spatial feature is not fused with the feature X_out before sending to phase and transition heads (like what you did in Stage 2)? (5) In Stage 2, how you deal with the predictions from the phase prediction head in local phase prediction module? If these predictions differ from the outputs of the Phase Prediction Head (inside the Green box in Fig.1) in Stage 2, what would happen?
- 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?
Authors present detailed experiments for the model. However, I think the motivation and introduction of the model design is not sufficiently clarified. Besides, as an online surgical phase recognition model, the authors did not justify the real-time performance.
- 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 design motivations behind Stage 1 and Stage 2 models are not explicitly explained, and the proposed design lacks sufficient experimental validation. Why fusing frame-level features at stage 1 won’t help in transition-aware pretraining? Why phase predictions from stage 1 are not used for the final decision?
Review #3
- Please describe the contribution of the paper
1.Proposes a two-stage framework for online surgical phase recognition, learning local temporal patterns from short clips and integrating them for long-range sequence modeling. 2.Introduces a feature fusion mechanism to combine frame-level and local temporal representations. 3.Applies a phase-order refinement strategy based on transition priors to improve temporal consistency.
- 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.Addresses an important and practical problem in online surgical phase recognition. 2.Proposes a coherent two-stage framework that combines local and long-range temporal modeling. 3.Provides comprehensive experiments, including comparisons and ablation studies. 4.The overall methodology is clearly presented and well-structured, making the paper easy to follow.
- 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.Unclear alignment with the online setting. Although the method is motivated by online surgical phase recognition, its design does not fully reflect a strictly online paradigm. The introduction of a two-stage training pipeline and additional post-processing steps makes the overall framework more complex, and less streamlined compared to prior online methods such as SKiT, which adopt a more direct and efficient end-to-end design. ______________ 2.Inconsistency of Viterbi-based refinement with online inference. The phase-order refinement module relies on Viterbi decoding over the sequence, which may introduce additional latency (especially for long videos). More importantly, this sequence-level optimization may revise past predictions y_{1:t}, whereas a strict online setting should ideally output only y_t at time t without revisiting earlier outputs. This raises concerns about whether the proposed approach fully adheres to the online constraint. ______________ 3.Potentially unfair comparison with prior methods. The reported results include the Phase-Order Refinement Module, which functions as a post-processing step rather than an integral part of the model. For a fair comparison with prior work, it would be more appropriate to also report results without this refinement, as existing baselines are not evaluated with similar post-processing. ______________ 4.Unclear contribution and necessity of the pretraining stage. The effectiveness of the Stage 1 pretraining is not sufficiently justified. The ablation study only compares against a frame-level baseline, without evaluating stronger single-stage alternatives. It remains unclear whether the two-stage design is necessary, or whether similar performance could be achieved with a fully end-to-end training strategy.
- 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.
(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 lean toward a weak accept. The paper addresses a relevant problem and presents a clear and well-structured framework, with consistent but modest improvements over strong baselines. However, there are concerns regarding the strict adherence to the online setting (e.g., Viterbi refinement) and the fairness of comparisons. Overall, the work is solid but would benefit from clearer justification of its design choices.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers and AC for their constructive feedback. We first address the common concern from R1 and R3 regarding the phase-order refinement module. We acknowledge the unclear explanation. The term “Viterbi refinement” and Eq. 9 were misleading, as they implied full-sequence decoding with revision of past predictions. In our intended inference, at time t, the model uses only current and past information and emits the current prediction. Refinement applies only when the current prediction is low-confidence and differs from the previously emitted phase. In this case, training-set transition priors are used to penalise implausible transitions. It does not access future frames or revise past predictions. Thus, the method remains online; however, we agree that the terminology and equation should be clarified. For R1 and R3’s concern on comparison without refinement, PhaseMamba achieves 92.2% accuracy, 91.1% precision, 91.8% recall, and 84.2% Jaccard on Cholec80.While accuracy trails SKiT/SurgFormer (93.4%) without refinement, the refinement module is an online-compatible component, not offline post-processing: it activates conditionally only when the current prediction is low-confidence, runs causally without buffering the full sequence, and adds negligible latency. It is therefore architecturally integral and a fair inclusion in comparisons. Notably, even without refinement, Jaccard is SOTA, surpassing all prior methods, critical as it penalises both missed and over-predicted phases unlike accuracy, which is biased by dominant phases under class imbalance. We thank R2 for raising runtime/real-time performance. Due to page limitations, this discussion was not included in the original submission. Under the same temporal-module-only benchmarking setting as SKiT, PhaseMamba shows approximately 2× lower inference latency than SKiT and is substantially faster than LoViT. This supports the efficiency motivation for using a Mamba-based temporal module in online surgical phase recognition. This efficiency advantage also supports the design rationale questioned by R1, R2, and R3.Regarding R3’s concern on two-stage necessity: Table 3 directly addresses this. Removing Stage 1 entirely (frame-only baseline) drops Jaccard by 10.7 pts and accuracy by 4.3 pts, demonstrating that end-to-end single-stage training is substantially inferior. Regarding R2’s question on Stage 1 fusion: Stage 1 operates on short clips for transition-aware pretraining in isolation; fusing frame-level features at this stage would conflate local transition learning with global appearance, undermining its specialised role. Contributions 2 and 3 describe distinct innovations: transition-aware local pretraining (Stage 1) and gated online fusion with causal Mamba (Stage 2), each independently validated in Tables 3 and 5.We would like to emphasise that the refinement module is applied only as a conditional online correction for low-confidence implausible predictions. Regarding R2’s questions on Stage 1 and Stage 2 interactions, the phase and transition heads in Stage 1 are auxiliary and used only during pretraining to supervise the local Mamba module. In Stage 2, the Stage 1 module is frozen and used purely as a local feature extractor; its auxiliary predictions are not used for the final decision. The final prediction is produced solely by the Stage 2 prediction head after fusing local and frame-level representations. Therefore, there is no conflict between Stage 1 predictions and Stage 2 predictions during inference. We thank all the reviewers for the minor comments on Figure 1, symbol definitions, contribution clarity, and reproducibility details. We are confident that the clarifications on online validity, the ablation evidence in Tables 3–5, and the runtime results collectively address the reviewers’ main concerns, and we will incorporate all updates in the final version.
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 Phase-Order Refinement Module uses Viterbi decoding that may revise past predictions, which appears inconsistent with a strict online setting; the authors should clarify how the final predictions are obtained at inference and whether past outputs are revised. To support fair comparison with prior online methods, results without the refinement module should be more clearly highlighted. The motivation behind the architectural choices in Stages 1 and 2 should be better explained. The necessity of the two-stage design relative to a single-stage end-to-end alternative should also be justified. Finally, the authors should clarify the undefined symbols in the equations and correct the inaccuracies in Figure 1.
- 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 addressed the most critical reviewer comments and clarified the raised issues.
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 paper proposes a two-stage Mamba-based framework for online surgical phase recognition that includes a phase-order refinement module. In the original reviews, both reviewers shared a core concern that the Viterbi-style refinement appeared to revise past predictions, thereby violating the online assumption. The rebuttal clarifies that the refinement is invoked only when the current prediction is low-confidence and inconsistent with the previously emitted phase. It uses training-set transition priors and neither accesses future frames nor revises past predictions. R1 explicitly indicated that this resolves the main concern. R2’s remaining concerns about the motivation for Stages 1 and 2 are partially addressed by an ablation showing a clear drop when Stage 1 is removed and by clarifying that Stage 1 heads serve an auxiliary pretraining role. The paper also reports competitive results on the standard benchmark even when the refinement module is removed. On balance, the substantive correction to the online-compatibility framing and the supporting ablations justify acceptance. The camera-ready version must rewrite the relevant section and equation to clearly describe the conditional, causal refinement procedure.
Meta-review #3
- 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, there are one Accept and one Reject. The other reviewer did not provide post-rebuttal recommendation. After reading all the reviews and author rebuttal, I think there is still large room to improve the paper’s presentation and strengthen the experimental verification. So the recommendation is Reject.