List of Papers Browse by Subject Areas Author List
Abstract
Vision-based surgical skill assessment (SSA) enables objective and scalable evaluation of operative performance. Progress in this field is constrained by the high cost and time demands for manual annotation of quantitative skill scores, as well as the poor generalization of existing regression models to new surgical tasks and environments. Meanwhile, appreciable volumes of unlabeled video data are now available, motivating the development of unsupervised domain adaptation (UDA) methods for SSA. We introduce the first benchmark for UDA in SSA regression, spanning four datasets across dry-lab and clinical settings as well as open and robotic surgery. We evaluate eight representative models under challenging domain shifts and propose CoRe-DA, a novel contrastive regression–based adaptation framework. Our method learns domain-invariant representations through relative-score supervision and target-domain self-training. Comprehensive experiments across two UDA settings show that CoRe-DA is superior to state-of-the-art methods, achieving Spearman Correlation Coefficients of 0.46 and 0.41 on dry-lab and clinical target datasets, respectively, without using any labeled target data for training. Overall, CoRe-DA enables scalable SSA with reliable cross-domain generalization, where existing methods underperform. Our code and datasets are available at https://anonymous.4open.science/r/CoRe-DA-F852.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3901_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/3901_supp.zip
Link to the Code Repository
https://github.com/anastadimi/CoRe-DA
Link to the Dataset(s)
RARP-skill dataset: https://github.com/anastadimi/CoRe-DA
RAH-skill dataset: https://github.com/anastadimi/CoRe-DA
JIGSAWS dataset: https://cirl.lcsr.jhu.edu/research/hmm/datasets/jigsaws_release/
AIxSuture dataset: https://zenodo.org/records/7940583
BibTex
@InProceedings{AnaDim_CoReDA_MICCAI2026,
author = { Anastasiou, Dimitrios AND Caramalau, Razvan AND Xu, Jialang AND He, Runlong AND Tesfai, Freweini M. AND Boal, Matthew W. E. AND Francis, Nader AND Stoyanov, Danail AND Mazomenos, Evangelos B.},
title = { { CoRe-DA: Contrastive Regression for Unsupervised Domain Adaptation in Surgical Skill Assessment } },
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
This paper presents an unsupervised domain adaption framework for video-based surgical skill assessment. The proposed method uses triplets from source and target video to learn domain-invariant representations.
- 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.
- Open-sourced code and dataset. Thank you for making this available
- Interesting considerations of information and gradient flow through different modules to prevent model collapse with contrastive learning
- Thorough comparison with existing methods
- Good performance on unlabelled 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.
- Error is still high relative to the range of the scores
- Since all unlabeled data is used for training contrastive loss, unclear how the relative module will act on unseen data. Granted, this is not uncommon in unsupervised learning, but it’s not a great indicator of generalizability to unseen users
- 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
- There is a lot of information in Fig. 1.It might be helpful to break up the modules into different figures/components to help reader comprehension
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents an interesting idea and demonstrates that it works well. Generalizing to unlabeled domain is an important problem in processing surgical videos where data is plentiful but labeling is expensive. While the overall performance on the task could be better, this work demonstrates significant improvement compared to previous work.
- 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 clarification in the final paper will help the reader understand the contribution.
Review #2
- Please describe the contribution of the paper
This paper introduces the first formulation of Unsupervised Domain Adaptation (UDA) for regression-based Surgical Skill Assessment (SSA). The authors establish a benchmark spanning four datasets (AIxSuture, JIGSAWS, RAH-skill, RARP-skill) across dry-lab and clinical settings, evaluate eight representative models, and propose CoRe-DA — a contrastive regression–based adaptation framework that combines relative-score supervision, dual-head consistency, and target self-training. Reported improvements over SOTA are substantial (+0.13 and +0.26 SCC on the two target domains). The direction is valuable and timely, the benchmark is a meaningful contribution to the community, and the results are convincing. However, several methodological details, fairness-of-comparison issues, and the positioning relative to existing contrastive regression work need clarification.
- 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.Novel and important problem formulation. Existing SSA methods are domain-specific and generalize poorly, while unlabeled surgical video is increasingly abundant. Framing SSA-UDA as regression (rather than binary classification, as in prior work [11, 23]) preserves the granularity required for meaningful feedback to surgeons. 2.Solid benchmark contribution. The four datasets cover genuinely different domain shifts — open vs. robotic, dry-lab vs. clinical, different instruments and procedures. This will be a useful resource for the community. 3.Well-motivated methodology. Exploiting the observation that OSATS/M-GEARS criteria are domain-agnostic, and leveraging relative scores (rather than absolute ones) to learn domain-invariant representations, is a principled choice. The combination of dual-head consistency, stop-gradient, and pseudo-label self-training is coherent. 4.Strong empirical results. Improvements are substantial on both settings with low variance across seeds (±0.03, ±0.05). Fig. 4 is particularly compelling — competing methods collapse predictions toward the mean, whereas CoRe-DA preserves score ordering across the full range. 5.Thorough ablations. Table 3 systematically evaluates each loss term, the stop-gradient mechanism, and background mixing, which is appropriate for the MICCAI standard.
- 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.Insufficient differentiation from prior contrastive regression work Section 2.2 explicitly builds on CoRe+GART [29] and follows [5] for the relative regressor design (concatenated pooled features + MLP). The authors claim to be “the first to explore contrastive regression as an adaptation strategy,” but the novelty relative to a naive combination of CoRe/GART with target self-training is not clearly isolated. Is the contribution the triplet sampling (source–exemplar–target), the consistency loss between R_abs and R_rel, or both? 2.Fairness of test-time background mixing The test-time transformation mixes source-exemplar backgrounds into target videos (λ = 0.25). While using source information at test time is not formally forbidden in UDA, this raises two concerns: （1）Table 3 shows a non-trivial drop (SCC 0.47 → 0.43) when this component is removed. （2）Were the competing baselines (MDD, DARE-GRAM, RSD, CO2A) evaluated with any comparable test-time augmentation? 3.Exemplar count and inference cost Testing uses M = 10 exemplars per target video, requiring 10 forward passes averaged at inference. The paper reports monotonic improvement from M = 2 to 12 (plateauing at 12), but does not report inference time or FLOPs. This is a meaningful cost for clinical deployment. Please add an efficiency comparison against baselines.
- 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?
This paper addresses an important and underexplored problem — UDA for regression-based surgical skill assessment — and contributes a valuable multi-dataset benchmark together with strong empirical results (+0.13 and +0.26 SCC over second-best baselines, with low variance across seeds). Fig. 4 convincingly demonstrates that CoRe-DA preserves score ordering while competing methods collapse toward the mean, and the ablations adequately isolate the contribution of each loss component. However, three concerns prevent a higher score. First, the methodological novelty is not clearly isolated from prior contrastive regression work: the relative regressor design largely follows CoRe/GART [29] and [5], and it remains unclear whether the gains come from the triplet sampling, the dual-head consistency loss, or simply from combining contrastive regression with target self-training — a direct CoRe/GART + pseudo-labeling baseline is missing. Second, the fairness of the test-time background mixing is unclear: Table 3 shows a non-trivial contribution from this component (SCC 0.47 → 0.43), and it is not stated whether competing baselines received comparable test-time augmentation. Third, inference cost is not reported despite using M = 10 exemplars per target video (10 forward passes averaged), which is a meaningful consideration for clinical deployment. On balance, the benchmark contribution and empirical strength outweigh these concerns, and the paper will be useful to the SSA community. I would be willing to raise my score if the authors convincingly address points 1 and 2 in the rebuttal.
- 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 paper presents a meaningful and well-motivated contribution, and the experimental results are promising. I believe the work has sufficient novelty and technical merit for publication at MICCAI.
Review #3
- Please describe the contribution of the paper
This paper introduces the first benchmark for unsupervised domain adaptation in surgical skill assessment regression, covering two UDA settings across four datasets. It also proposes CoRe-DA, a contrastive regression-based adaptation framework that combines relative-score supervision, self-training on unlabeled target videos, and exemplar-based testing to improve cross-domain skill prediction.
- 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 addresses an important and underexplored problem in surgical AI: adapting skill assessment models to new procedures and environments without labeled target data. The benchmark contribution is valuable, since prior work in this area is limited and often focuses on classification rather than regression. The proposed method is well motivated, technically coherent, and consistently outperforms strong baselines across both UDA settings. The paper is also generally well organized, with useful ablations and an additional semi-supervised experiment.
- 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 cross-dataset label harmonization deserves stronger justification. Although the source and target domains share overlapping OSATS/M-GEARS components, reducing them to a single summed score may obscure differences in rubric interpretation, score calibration, and component importance across datasets.
A key part of CoRe-DA is exemplar-conditioned inference at test time, together with background mixing. This is an interesting design choice, but the paper does not fully analyze how sensitive performance is to exemplar selection, the mixing ratio, and the additional test-time computational cost.
The proposed method combines several reasonable components, including relative-score supervision, target self-training, and background mixing, but the current analysis does not fully disentangle which component drives the largest gains. This is especially important because the main ablation is limited to a single adaptation setting and a single run.
The empirical evaluation focuses on the summed final skill score, whereas the underlying rating systems (e.g., OSATS and M-GEARS) are multi-component. Reporting adaptation performance for individual skill dimensions would provide a more informative and clinically meaningful analysis.
The statistical evidence could be strengthened. For example, the paired significance analysis appears to be based on predictions from a single run, whereas significance assessed across repeated runs or dataset splits would better support the robustness claims.
The benchmark settings conflate multiple sources of domain shift, including task variation, operative environment, and surgical platform. A more fine-grained analysis of which shift factors are most challenging would improve the scientific insight of the study and clarify the method’s expected range of generalization.
- 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
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 recommend a weak accept because the paper addresses an important and underexplored problem, introduces a useful benchmark, and shows clear improvements over strong baselines. My main reservations are that the label harmonization across datasets is not fully analyzed, the contribution of individual components is not completely disentangled, the evaluation focuses only on the summed score, and the statistical evidence could be stronger. Overall, the work is meaningful and timely, but some analysis and validation could be strengthened.
- 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 R1, R2, R3, MR. We are glad our work is “interesting-R1”, “valuable and timely-R2”, “well motivated, technically coherent-R3”, “addresses a novel, important, underexplored problem-all” with “strong, convincing results-all”. Answers follow and will be added in final paper.
R1-High errors: CoRe-DA yields much lower errors/SD than all methods (Tab.1). MAE of 4.29 is reasonable for [6,30] JIGSAWS range, given no target labels for training. On RARP-skill it even beats fully supervised baselines (MAE 1.93 vs 2.10 [4],suppl.).
R1-Generalizability: Using all unlabeled target data to train and test is standard in UDA [8, 30] given small datasets. Future work to study this with larger ones.
R1-We will restructure Fig.1. R2-Differentiation from prior contrastive work: We are the first to propose contrastive regression as an UDA strategy, introducing novel components:(1) triplet sampling enabling cross-domain contrastive regression, (2) cons. loss between R_abs and R_rel, (3) self-training (L^T_cons with stop-grad). Each one is essential (per ablations,Tab. 3). [5,29] do not address domain adaptation (pure AQA models).
R2-Fairness of test-time BG mix, non-trivial drop: Test-time BG mix is an integral design contribution of CoRe-DA, not applicable to competing methods as it requires exemplar anchoring (to extract the BG). Applying a global transform of the average source BG is of no value. Note, it’s not stochastic test-time augmentation (multiple random passes, often used to boost performance); it’s a deterministic, single-pass transformation to partially align low-level visual statistics across domains. From Tab.3 even without BG mix, CoRe-DA (SCC: 0.43,MAE: 4.57) outperforms all methods of Tab.1. R2,R3-Inference cost, exemplar selection, mix ratio: SSA is a post-op application, not real-time, so inference cost is not critical. GFLOPs are: Source-Only/MDD/DARE-G./RSD 670.638 (same model inference), Contra-S 2177.518, ViSA 912.435, CO2A 889.276, CoRe-DA (M=10) 1760.424.
We empirically found random exemplar sampling to be slightly worse than uniform. Very small mix ratio values have minimal effect (mirroring ablation without BG mix), while values >0.25 distort target frames, degrading performance.
R2,R3-CoRe-DA component gains, and disentanglement: Triplet sampling and self-training are inherently coupled and cannot be evaluated in isolation. Accurate Δy_T-E_hat predictions require R_rel to train on source-exemplar pairs, so they work jointly by design. Our ablations systematically remove one component at a time and evaluate the resulting performance drop to isolate each component’s contribution (incl. that of cons. losses). Single seed reporting is standard for ablations [3,9,16].
R3-Cross-dataset label harmonization: Our settings share identical label components (6 OSATS for AIxSuture-JIGSAWS, 6 M-GEARS for RAH-skill-RARP-skill); we do not transfer OSATS to M-GEARS as it violates UDA assumptions. We only use dry-lab-dry-lab or clinical-clinical pairs to reduce rubric/calibration mismatch. We assume marginal distribution shift with invariant label-conditional distribution (standard in UDA); relaxing this would require target labels (outside scope). OSATS and M-GEARS weight components equally, so summing introduces no bias.
R3-Statistical evidence: We include paired t-test on all 4 runs for both UDA settings (p-value ranges: JIGSAWS/ RARP-skill). CoRe-DA’s improvements are statistically significant (or marginally) against nearly all cases: Source-Only (0.018-0.058/<0.05), Contra-S. (<0.05/0.000-0.070), ViSA (<0.05/<0.05), MDD (<0.05/<0.05), DARE-G. (0.024-0.056/<0.05), RSD (0.043-0.155/p<0.05), CO2A (0.043-0.084/<0.05).
R3-Shift factors, individual skill components: All cases are suturing tasks. AIxSuture-JIGSAWS shifts platform (open vs robotic) and environment, RAH-skill-RARP-skill shifts environment/procedure. Fully isolating individual shifts and individual skill components scoring are important future directions
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.
Should go forward for rebuttal from the authors.
- 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.
.
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.
All reviewers recommend acceptance.
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 the main concerns by clarifying novelty relative to prior contrastive regression work, explaining the role and fairness of test-time background mixing, and providing additional details on inference cost and statistical validation. Some limitations remain (e.g., reliance on exemplars at test time and partial disentanglement of components), but these are well acknowledged and do not undermine the main contribution. The strong benchmark, clear problem formulation, and consistent improvements support the paper’s value. Reviewers agree on Acceptance.