List of Papers Browse by Subject Areas Author List
Abstract
Online Surgical Phase Recognition (SPR) models can reach high frame-wise accuracy, yet their predictions often lack temporal stability, fragmenting workflow understanding and reducing the reliability of downstream assistance. We show that this instability is not random noise but arises from two mechanisms: early misclassifications corrupt temporal feature states and propagate forward to form error cascades, and phase transitions follow evidence-accumulation dynamics whereas most online SPR systems rely on memoryless frame-wise decisions, making them sensitive to transient confidence fluctuations. We propose a unified Train–Inference–Evaluation framework that explicitly stabilizes temporal inference dynamics using model-agnostic, plug-and-play components. For training, the Temporal Error-Cascade (TEC) loss suppresses error onset and mitigates forward error propagation by stabilizing temporal feature evolution. For inference, the Evidence-Gated Transition Predictor (EGTP) enforces evidence-driven state transitions, allowing phase changes only when accumulated evidence exceeds a confidence boundary. For evaluation, we introduce the Temporal Fragmentation Index (TFI), a
reliability-aware metric that quantifies instability-induced temporal disagreement beyond conventional frame-wise and token-based measures. Experiments on Cholec80 and AutoLaparo across three representative backbones show that the proposed framework substantially improves temporal stability and reduces prediction fragmentation, while maintaining or modestly improving frame-wise performance. Code is available at https://github.com/MRUIL/StableSPR.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0190_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/MRUIL/StableSPR
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiuYan_Stabilizing_MICCAI2026,
author = { Liu, Yang AND Zhu, Ning AND Peng, Jingjing AND Chen, Xiwu AND Granados, Alejandro AND Wang, Guotai AND Ourselin, Sébastien},
title = { { Stabilizing Temporal Inference Dynamics 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
The authors present an approach to stabilize temporal inference dynamics in surgical phase recognition by a) introducing a novel loss for training SPR backbones that penalizes frequent phase switching, and b) including an evidence-gated transition predictor that prevents frequent phase switching. In addition, the authors introduce a novel metric to accurately capture the instability of SPR across entire video segments.
- 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 authors address an important issue: stability in SPR which is relevant to the community. The proposed modules for stabilization, the loss function and the evidence based gate, are well motivated and methodologically sound. The introduction of a novel metric to better capture temporal inference instability is an important addition which benefits the community. The experiments using two different data sets and three different state-of-the-art SPR backbones, and their results illustrate the validity of the proposed contributions. The results clearly show that temporal stability has been drastically improved.
- 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 main weakness is that improvements obtained with the proposed approach with respect to Accuracy and Precision are modest. This is however not the main claim of the manuscript. The temporal stability is indeed improved significantly.
- 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
The legibility of the labels in some figures should be improved. In Fig. 2a, EGHS probably refers to EGTP.
- 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 relevance to the community, the novelty of the proposed elements and the convincing experiments and results.
- 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 framework for dealing temporal instability in surgical phase recognition in surgical videos. It includes a loss function for learning, an additional step for inference and a metric for validation. It has been tested on two publicly available data sets with ablation studies and shown improvements in several metrics.
- 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 objective of the paper is relevant. However, temporal instability is on issue between other ones in surgical workflow recognition from surgical videos The proposed solution is generic and can be used within several architectures as shown in the paper. The proposed solution includes a coherent framework covering training, inference and validation. Quantitative validation performed with several metrics on two publicly available datasets show the added value of the approach.
- 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 paper is based ont the statement that “true surgical phase transitions are inherently evidence-accumulation processes”. It is true but it is not the only reason. Instability can simply come from a wrong or ambiguous definition of a phase or since the information is not available in the video: something happens out of the field of the surgical video. In the Metrics section and in the introduction section, one of the first (if not the first) publications mentioning this issue and suggesting corresponding new metrics is not cited: Dergachyova, O., Bouget, D., Huaulme, A., Morandi, X., & Jannin, P. (2016). Automatic data-driven real-time segmentation and recognition of surgical workflow. Int J Comput Assist Radiol Surg, 11(6), 1081-1089.Results: Analysis without standard deviation and without statistical analysis is not acceptable. The statement of order of magnitude improvements is oversold. Extension: And what about step recognition ? Is there the same issue ? Is your solution compatible ? Will it provide the same improvements ?
- 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?
A well written paper but could be improved especially as regards to the results section.
- 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
This paper addresses an important problem in online surgical phase recognition, predictions may be frame-wise accurate but temporally unstable, which can fragment workflow understanding and reduce the reliability of downstream assistance systems. To address this, the authors propose a unified train–inference–evaluation framework consisting of: (i) the Temporal Error-Cascade (TEC) loss, which emphasizes the onset of consecutive errors during training to reduce forward error propagation; (ii) the Evidence-Gated Transition Predictor (EGTP), which models phase changes as an evidence-accumulation process during inference rather than relying on instantaneous argmax decisions; and (iii) the Temporal Fragmentation Index (TFI), a new metric intended to quantify instability-induced temporal disagreement beyond frame-wise accuracy and Edit Score. The framework is presented as model-agnostic and plug-and-play, and experiments on Cholec80 and AutoLaparo with three backbones suggest improved temporal stability. The combined addage of TEC and EGTP to each of the three backbones resulted in higher performance and stability across both datasets.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
A major strength of the paper is that it focuses directly on temporal stability, which is a clinically relevant perspective for online surgical workflow recognition. The proposed framework is also conceptually clean, with separate components for training, inference, and evaluation, and the methods themselves are easy to understand and appear broadly applicable across several representative SPR backbones, including Trans-SVNet, SKiT, and Surgformer. The introduction of the TFI metric further strengthens the paper, as it highlights a real limitation of Edit Score by distinguishing between predictions that may have similar compressed token sequences but very different temporal fragmentation. The qualitative and quantitative results consistently show that EGTP strongly reduces flickering and rapid phase switching across datasets.
- 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.Experimental protocol and reporting: It is currently difficult to assess how reliable the reported gains actually are. In particular, the paper states that Cholec80 is split into 40 training and 40 testing videos, but it does not explain how hyperparameters such as alpha and k are selected in the absence of a validation split. In addition, the paper reports only single-run numbers without standard deviations or multiple-seed experiments, even though many of the reported gains in accuracy-related metrics are small and may not be statistically meaningful; this is especially important because some prior work in this area (https://arxiv.org/pdf/2305.13961) has reported variances on Cholec80 which might suggest that several of the perfomance metrics reported could be statistically indistinguishable. This also goes in line with recent findings regarding overclaiming model accuracy when statistical variability is not considered (https://arxiv.org/pdf/2505.04720) which in particular affects the validity of the “improved accuracy” claim in this study.
2.Benefit of EGTP architecture: The empirical comparisons do not fully compare EGTP against simpler or more established temporal smoothing alternatives. Since EGTP functions as a post-processing comparison to standard temporal smoothing or structured prediction methods, such as transition constraints or probabilistic sequence smoothing approaches, would be important to show that the benefit comes from the proposed mechanism rather than from generic temporal regularization alone. Furthermore, because surgical workflows follow a strict procedural order, the authors should evaluate whether simply restricting impossible phase transitions (e.g., using a transition prior matrix derived from the training set) would achieve similar stability improvements. Without this baseline, it is unclear if the dynamic evidence accumulation of EGTP is truly necessary or if a simpler transition constraint would suffice.
3.Interpretation of results: The results reveal trade-offs that are not discussed clearly enough. While EGTP strongly improves TFI and Edit Score, some precision, recall, and Jaccard results on Cholec80 are better without EGTP, suggesting that stronger temporal smoothing may suppress true transitions or harm class-wise performance in some settings. Similarly, in the EGTP versus E-Const comparison, the proposed method is argued to be better mainly due to slightly higher accuracy and more consistent threshold scaling, but in some cases E-Const yields substantially lower TFI, so the paper should better justify why the chosen balance between marginal accuracy gains and much larger stability differences is preferable for practical deployment.
4.Missing discussion: A major structural weakness is the absence of a discussion section which is necessary to bring the achieved results into perspective of related work and discuss shortcomings and future work. A summarizing conclusion is not sufficient for this. Furthermore, the experiment section should be renamed to Experiments and Results to highlight where the results are presented in the paper.
- 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
- All figures and tables should be self-explanatory. In particular in the caption of figure 1 the terms p, P, E, y^, M should be introduced
- The authors claim that “On AutoLaparo of Surgformer, EGTP reduces TFI from 10.67 to 0.48” while table 1 only shows a reduction from 10.67 to 0.96.- While the newly introduced TFI metric conceptually addresses the temporal disagreement within segments compared to Edit Score, its benefit doesn’t seem to show in the performed validation in which Edit Score features a similar “nearly one order of magnitude” change. It would be good to discuss / highlight scenarios in the context of SPR where TFI is beneficial.
- 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?
This paper addresses a highly relevant clinical problem by proposing an elegant, model-agnostic framework (TEC, EGTP, and TFI) to explicitly tackle temporal instability in online surgical phase recognition. However, the experimental methodology currently lacks the rigor required to fully support the authors’ claims. Most critically, the paper does not define a validation split for the Cholec80 dataset or generally clarify the subsets used, raising significant concerns about hyperparameter tuning (such as alpha and k) and potential data leakage on the test set. Additionally, the absence of variance estimates or standard deviations across multiple random seeds makes it impossible to verify the statistical significance of the reported marginal accuracy gains, which is a known issue in this field. Finally, EGTP operates as a temporal post-processing decision policy but fails to compare against established temporal smoothing baselines, such as Hidden Markov Models (HMMs), Conditional Random Fields (CRFs), or basic surgical workflow transition priors. While the core ideas and the new TFI metric are strong, the authors must clarify their data splitting protocol, provide variance estimates, and include standard baselines in the rebuttal for this work to confidently cross the acceptance threshold.
- 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
N/A
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 tackles a clinically important problem. Reviewers consistently agree on the relevance, soundness, and generality of the approach. The key contributions are well-motivated and practically impactful, with clear and consistent improvements in temporal stability across multiple datasets and backbones. The introduction of TFI is also a valuable addition for the community. While minor concerns remain (e.g., statistical reporting and additional baselines), they do not undermine the strength of the core contributions. Overall, the paper shows clear novelty, solid validation, and high practical relevance. Therefore, I would suggest early acceptance.
