List of Papers Browse by Subject Areas Author List
Abstract
Surgical phase recognition models struggle with rare phases—states under-represented in training annotations, occupying <5% of training frames. We identify the coverage-gap paradox: increasing rare-phase training videos fails to proportionally improve accuracy because prototype dilution averages out intra-class variation, a phenomenon we observe consistently across three representative classifier families (k-NN, linear probes, and prototypes). To bridge this gap, we propose a lightweight test-time adaptation (TTA) framework (20K parameters, <0.1 ms/frame) whose core mechanism is an adaptive pseudo-label threshold (q-th percentile) that avoids the empty-pool failure of fixed thresholds, combined with temporal smoothing for pseudo-label precision filtering and auto-guided annotation (B=5 frames). On three datasets spanning gastric bypass (MB140, 13 phases), cholecystectomy (Cholec80, 7 phases), and cataract surgery (Cat-101, 10 phases)—our method improves rare-phase accuracy by +29.2/+31.5/+33.3 pp, outperforming temporal convolutional network (TCN), gated recurrent unit (GRU), and TENT-style baselines.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6471_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/powersimmani/tta-rare-surgical
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ParHo_TestTime_MICCAI2026,
author = { Park, Ho-min AND Tozzi, Francesca AND De Muynck, Robbe AND Kim, Narim AND Rashidian, Niki AND Willaert, Wouter AND De Neve, Wesley AND Vankerschaver, Joris},
title = { { Test-Time Adaptation for Rare Surgical Phase Recognition: Bridging the Coverage-Gap Paradox } },
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
A method to improve recognition performance for rare phases in the recognition of surgical videos is proposed.
- 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 propose a test-time adaptation framework centered on an adaptive pseudo-label threshold that resolves the empty-pool failure of fixed thresholds, with temporal smoothing serving as a precision filter for downstream pseudo-label selection. The proposed method assigns phase numbers using the nearest neighbor algorithm and adaptively adjusts representative points at test time; its key feature is its low computational complexity.
- Among three representative surgical image datasets (MultiBypass140, Cholec80, and Cataract-101), the method demonstrates improved recognition performance for rare phases compared to conventional methods on two of these 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.
There are numerous points where the presentation of experimental results and the discussion are insufficient, making it impossible to assess the effectiveness of the proposed method.
- In the experimental results for “Causal variants” in Section 4.2, only the results for Cholec80 are shown; results for the other two datasets are not presented.
- In Table 1, the experimental results for “Combined” and “Oracle” on the Cholec80 dataset are not shown. Furthermore, no reason is given for this omission.
- In Table 1, for the Cataract-101 dataset, the proposed method (Combined) achieves a 33.3-point improvement in accuracy over the Baseline for the rare phases, but a 10.8-point decrease in accuracy for the overall phases. Such a significant drop in accuracy for the overall phases appears to be a major issue, yet no discussion or explanation regarding this is provided.
- In Table 2, results for Cataract-101 are not shown.
- In Table 2, the experimental results for Cholec80 show a significant decline in “Ours (Combined)”, resulting in lower performance than the conventional methods; however, no discussion of this is provided.
- Please rate the clarity and organization of this paper
Poor
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
In addition to the major weaknesses, there are concerns regarding the following points, and it is hoped that the authors will address them.
- The text is often written in a prose style, which is inappropriate for a scientific paper. This is particularly evident in Sections 4.2 through 4.4.
- There is no explanation of the experimental conditions used to obtain the results in Table 1 and elsewhere, specifically regarding the value of w.
- There is no explanation of what the “self-training method” in Table 1 entails.
- Section 4.4 presents experimental results for “Cross-center transfer,” but the experimental conditions are not specified.
- The meaning of “structural embedding limitation” in the “Limitations” section of Chapter 6 is unclear.
- 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?
As indicated in item 7, there are numerous points where the presentation of experimental results and the discussion are insufficient, making it impossible to assess the effectiveness of the proposed method.
- 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.
In the first review, it was pointed out that a discussion was needed regarding the reason why “Ours (Combined)” performed significantly worse than “Ours (Temporal)” in the results for Cholec80 shown in Table 2, and why its performance was lower than that of conventional methods. In response, the authors state that this is due to the same cause as the significant decline in Cat-101’s Overall results in Table 1 (the Stage-4 oracle-drift mechanism). However, since Overall performance declined in Table 1 while Rare performance improved, and Rare performance declined in Table 2, this explanation is considered insufficient for the discussion.
Review #2
- Please describe the contribution of the paper
The paper fixes surgical AI’s failure on “rare phases” by identifying the “coverage-gap paradox,” which proves that simply adding more training data actually dilutes the model’s understanding. To solve this, they built a super lightweight Test-Time Adaptation (TTA) framework that ditches rigid confidence thresholds for an adaptive percentile threshold, guaranteeing the model actually captures rare events for self-updating. Combined with a quick 5-frame human check, this massively boosts rare-phase accuracy by ~30 percentage points across three different surgical 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.
1.The proposed Test-Time Adaptation framework is exceptionally lightweight, requiring under 20,000 parameters and processing at less than 0.1 milliseconds per frame, making real-time intraoperative deployment highly feasible 2.Its shift from a standard fixed-confidence threshold to an adaptive q-th percentile threshold is a brilliant fix that guarantees rare-phase updates and acts as the primary driver of the model’s success.
- 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.
While the method improves rare-phase recognition, it causes a severe degradation in overall accuracy on the Cataract-101 dataset, dropping by up to 13.7 percentage points. The authors acknowledge this as a structural limitation for datasets with a high number of classes and overlapping clusters. 1.The framework relies on pseudo-labels that have very low precision, explicitly noted as 11-25% on the MB140 dataset. This means the prototypes are being updated with a large proportion of incorrect data, leaving theoretical convergence bounds an open question.
2.The ablation study used to justify hyperparameter choices is fragile, as it was conducted on only a single fold and a single random seed. The authors concede that values from a single seed can deviate from the true multi-seed means by up to 12 percentage points.
3.The study lacks a systematic evaluation of different foundation model choices, relying solely on LemonFM and DINOv2 for feature extraction. Because the “coverage-gap paradox” and the success of the Test-Time Adaptation framework fundamentally depend on the geometric properties of the embedding space, failing to benchmark across diverse foundation models leaves it unclear whether the observed baseline failures and subsequent accuracy gains are universal or merely artifacts of the specific embeddings used.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
While this brilliantly achieves massive ~30% accuracy gains on rare phases across three diverse surgical datasets and is highly feasible for real-time deployment, the approach is ultimately hindered by severe trade-offs in overall accuracy on complex datasets, reliance on low-precision pseudo-labels, and fragile statistical validation.
I would recomment the author to clarify the experiment setting to make sure the random seed is not the cause of the improvement.
A discussion about the justification behind general and surgical foundtion model will be nice.
- 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.
The rebuttal address my concern
Review #3
- Please describe the contribution of the paper
The paper makes three main contributions. First, it identifies the “coverage-gap paradox”, namely that increasing the number of training videos containing rare-phase labels does not proportionally improve rare-phase recognition because prototype averaging dilutes intra-class variation. Second, it proposes a lightweight source-free TTA pipeline built around an adaptive percentile-based pseudo-label threshold, combined with temporal smoothing and optional annotation of only a few selected frames. Third, it validates this approach on MB140, Cholec80, and Cataract-101 datasets, showing substantial gains in rare-phase accuracy and competitive comparison against several temporal baselines.
- 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 paper tackles a genuinely important and under-addressed failure mode in surgical phase recognition: poor performance on rare or sparsely annotated phases. This is clinically meaningful because the paper is not merely optimizing overall accuracy, but focusing on phases that are easy to miss and often most affected by data scarcity.
The paper clearly identifies and empirically characterizes a “coverage-gap paradox” linked to prototype dilution across multi-center surgical videos, a problem formulation that resonates with practical annotation scarcity and across-hospital heterogeneity.
The central adaptive percentile threshold for class-conditional pseudo-label selection is simple, computationally negligible, and directly addresses empty-pool failures of fixed thresholds in rare classes.
The framework integrates temporal precision filtering and a minimal-budget, auto-guided annotation strategy in a practical TTA pipeline (including a causal, real-time variant).
Evaluation on three distinct surgical domains (gastric bypass, cholecystectomy, cataract) with multiple settings (within- and cross-center) demonstrates robustness.
Factorial analysis isolates the contribution of the adaptive threshold (dominant driver, +22 pp on MB140) and shows interactions with temporal smoothing for precision improvements.
Addresses a clinically relevant failure mode (rare, short phases with low label coverage) that conventional long-tail or class-balanced losses cannot resolve when video-level class coverage is lacking.
The solution is practical: source-free, low-parameter, fast, and easily layered atop foundation-model features for deployment.
- 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 adaptive percentile pseudo-labeling idea, while effective, is conceptually modest and related to established class-wise confidence/percentile selection in UDA/self-training; the novelty is more in application to surgical rare-phase TTA than in core methodology.
Prototype updating can adversely shift decision boundaries, substantially degrading overall accuracy in higher-class settings (e.g., Cat-101), indicating brittleness in complex feature geometries.
The method presumes non-empty class-specific predicted pools to compute percentiles; a fallback strategy for empty-pool scenarios is not fully specified.
Comparisons to strong temporal backbones (e.g., MS-TCN, Trans-SVNet, LoViT) under an adapted rare-label regime are not provided; only post-hoc smoothing and a TENT-style baseline are shown, potentially underestimating competitive strong baselines for rare phases. The paper cites SPA as complementary, but does not empirically compare against it. Given the paper’s framing around test-time adaptation for surgical phase recognition, the absence of a stronger surgical-specific adaptation baseline weakens the claim of state-of-the-art relevance.
Reported overall accuracies are quite low on Cholec80 relative to the literature (owing to the label-masking regime and prototype classifier), making it hard to contextualize results versus established systems.
Prior class-wise adaptive pseudo-labeling in source-free UDA/self-training (e.g., CBST and related class-conditional thresholding) is not cited or contrasted. Also, broader long-tailed action/phase literature and prototype-based adaptation work could be more comprehensively connected to the paradox and method.
There is a meaningful trade-off with overall performance, especially on Cat-101, where rare-phase gains come with a large drop in overall accuracy. Since phase recognition systems are usually deployed for complete workflow understanding, this is not a minor issue. The paper describes it as structural, but the practical consequence is that the proposed method may improve the corner case while harming the full task.
Finally, the “Combined” setting uses oracle annotation of B=5 frames, which is useful as an upper-bound/practical hybrid scenario, but it also means some of the most impressive gains partly rely on human input. That is not a flaw by itself, though the framing should remain careful when emphasizing adaptation performance.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
I am not convinced whether the broader conceptual claim is fully substantiated. The paper would be stronger if the “coverage-gap paradox” were framed a bit more cautiously as an observed failure mode of prototype-based rare-phase recognition under sparse coverage, rather than as a more universal classifier-agnostic principle. Likewise, the empirical story would be more convincing with stronger comparisons to recent surgical adaptation methods and with a more explicit discussion that Cholec80 reflects annotation scarcity more than naturally rare phase occurrence.
- 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?
This paper addresses a meaningful and underexplored problem in surgical video analysis: recognition of rare or sparsely annotated phases. Its central technical contribution, adaptive percentile-based pseudo-label thresholding for lightweight source-free test-time adaptation, is simple, well motivated, and empirically effective. The method is also appealing from a deployment standpoint because it is compact, fast, and compatible with real-time settings. The experiments are reasonably broad and include useful analyses such as factorial decomposition and cross-center transfer. These are substantial strengths. On the other hand, I am not fully convinced by the breadth of the paper’s conceptual claims, especially the framing of the coverage-gap paradox as classifier-agnostic, since the evidence is mostly limited to prototype-oriented or shallow classifiers over fixed features. The comparison set is also not as strong as it could be for a paper centered on surgical TTA, and some statistical conclusions are weakened by pseudo-replication. In addition, the large overall-accuracy drop on Cat-101 raises an important practical concern.
- Reviewer confidence
Somewhat confident (2)
- [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 was already leaning toward a weak accept, and the authors’ rebuttal helped alleviate several of my concerns.
Author Feedback
We thank the reviewers and meta-reviewer. We restrict the response to presentation amendments of existing results; algorithmic extensions are future work.
A. Cat-101 overall-rare trade-off (shared). Under the Stage-4 oracle update, rare-class prototypes on Cat-101 drift toward boundary regions of the embedding space. This drift raises rare-recall but reduces the margin against neighboring common classes, costing -10.8 pp overall under Combined and -13.7 pp under Oracle (Tab. 1). The mechanism is the same in both cases: it is structural to the source-free regime with a frozen encoder on a 10-class topology (Sec. 4.2), not a missed hyperparameter. Anchoring each rare-class prototype to its base counterpart by bounding the cosine drift, or applying a confidence-margin gate at prediction time, both preserve the source-free guarantee and are a clearly scoped future direction.
B. R1 - missing data and conditions. Cholec80 Combined/Oracle in Tab. 1.Cholec80 uses label masking: videos stay in training but rare-phase labels are masked. Oracle annotation re-introduces them, making “Combined/Oracle” semantically circular under this protocol. MB140 and Cat-101 use video-removal and are unaffected. Cholec80 Combined decline in Tab. 2.Same Stage-4 oracle-drift mechanism as in §A applies. Cat-101 in Tab. 2.The MV/TCN/GRU/TENT comparison was run on MB140 and Cholec80 only; the Base-vs-Temporal-vs-Combined analysis on Cat-101 in Tab. 1 supports the same claim that temporal post-processing alone cannot bridge the coverage gap. Cholec80 causal. Tab. 4 already reports causal results (71.2% vs. 70.8% bidirectional, w=15); the small gap supports cross-dataset extension as future work. Clarifications. w=15 means 15 frames at 1 fps; Self-Training is TENT-style entropy minimization on a prototype classifier with a fixed p=0.95 threshold; “structural embedding limitation” means the frozen-encoder representational ceiling under the source-free constraint.
C. R2 - precision, ablation, FM diversity. 11-25% pseudo-label precision. The adaptive percentile is designed to prevent rare-class empty pools (factorial Tab. 3); we accept lower per-frame precision for guaranteed coverage. Temporal smoothing and prototype-centroid geometry recover precision downstream. “Single-seed” wording. The wording was misleading: the factorial in Tab. 3 in fact uses 5 seeds x 5 folds, matching the headline +22 pp claim. The “up to 12 pp deviation” note refers to a separate per-fold sub-analysis in the hyperparameter ablation. FM diversity. LemonFM (surgical) and DINOv2 (general) span two distinct families and TTA gains are consistent across both (Tab. 1); broader FM sweeps are future work.
D. R3 - prior work, baselines, framing. CBST / class-conditional thresholding. Our method differs by being source-free TTA over fixed prototypes (no retraining), providing a per-class empty-pool guarantee via the adaptive percentile, and running at <0.1 ms/frame. Stronger temporal baselines. Tab. 2 already compares MV, TCN post-hoc, GRU-smooth, and TENT-style on the same features. MS-TCN and Trans-SVNet are full end-to-end retraining methods outside the source-free regime we target; our claim is that temporal post-processing alone cannot bridge the coverage gap (Sec. 4.3), which Tab. 2 evidences. “Classifier-agnostic” framing. We agree this overstates the result; the supported claim is “observed across three representative classifier families (k-NN, linear, prototype)” (Sec. 3.2). Combined uses oracle B=5.The Temporal rows in Tabs. 1-2 are the fully source-free results, and we will sharpen this distinction.
Camera-ready amendments. Tab. 1 footnote on label masking; CBST (Zou 2018), SHOT, and SPA citations in Sec. 2; softened “classifier-agnostic” claim; sharper Combined-vs-Temporal distinction in Sec. 3.4; corrected Limitations note on factorial seed counts. We intend to extend our code release with reproducibility scripts upon acceptance.
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.
There are mixed reviews (2WA, 1WR). Reviewers are concerned with the presentation of experimental results and the limited discussion. It would be great to clarify them during rebuttal.
- 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.
This paper explores improving recognition of rare or sparsely annotated phases. The main merit of the work is the proposed lightweight source-free test-time adaptation strategy based on adaptive percentile pseudo-labeling. The method directly targets the empty-pool failure of fixed confidence thresholds for rare classes. It looks computationally very efficient and is practical for real-time deployment. The evaluation across multiple surgical datasets and feature backbones provides useful evidence that the approach can improve rare-phase recognition in several settings. The rebuttal also clarified several important points, including the seed/fold setup and the role of the Combined versus Temporal settings. There are still remaining concerns, including performance degradation on more complex datasets and the reliance on low-precision pseudo-labels. Despite these limitations, I think the problem formulation and practical efficiency are sufficiently valuable for the MICCAI community. I would suggest accept.
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.
After reviewing the rebuttal letter and reviewers’ post-rebuttal comments, I recommend acceptance as the post-rebuttal largely addresses reviewers’ concerns.
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.
The reviewers highlight concerns regarding insufficient experimental results and discussion, making it difficult to assess the method’s effectiveness. Key issues include missing baseline evaluations, lack of explanation for significant performance drops, reliance on fragile ablation studies, and limited benchmarking across diverse models or stronger baselines. Despite addressing concerns in the rebuttal, the core issues around robustness, novelty, and broader applicability remain unresolved.