List of Papers Browse by Subject Areas Author List
Abstract
Critical View of Safety (CVS) assessment has become a central safeguard against bile duct injury during laparoscopic cholecystectomy (LC). Current deep learning-based solutions achieve strong performance but rely on massive backbone architectures (up to >1B parameters) that hinder real-world deployment, specifically in resource-constrained operating rooms. We challenge the assumption that such scale is necessary. Because CVS assessment is confined to a specific surgical phase and anatomical region, its visual representation occupies a compact, low-dimensional manifold. Guided by the hypothesis that this reduced semantic dispersion enables substantial model compression without sacrificing accuracy, we introduce EdgeCVS, a phase-gated distillation framework that compresses a high-capacity teacher into a 5M-parameter edge-efficient student. By incorporating anatomical spatial priors and phase-gated pseudo-labels, the compact model recovers complex decision boundaries. EdgeCVS achieves mAP close to the most recent state-of-the-art models while reducing the parameter count to 1.5%, enabling real-time inference (40 FPS) on standard CPUs.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1745_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/IMSY-DKFZ/edgecvs
Link to the Dataset(s)
SAGES-CVS: https://huggingface.co/datasets/CAMMA-public/SAGES_CVS_Challenge_2024
Endoscapes: https://github.com/CAMMA-public/Endoscapes
Cholec80: https://github.com/CAMMA-public/TF-Cholec80
Phakir: https://zenodo.org/records/15740620
BibTex
@InProceedings{YamAmi_EdgeCVS_MICCAI2026,
author = { Yamlahi, Amine AND Hennighausen, Jakob AND Hansen, Pascal AND Leeb, David AND Maier-Hein, Lena},
title = { { EdgeCVS: Democratization of surgical AI with a Distilled Edge-Deployable Critical View of Safety (CVS) model } },
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 studied the feasibility of a real-time CPU-based surgical AI for automated CVS assessment. The method is distillation from big to small models.
- 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 clinical motivation is clear 2.The paper is well written and easy to understand.
- 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.For such an application focused paper on tiny ML, the authors should compare with all the relevant state-of-the-art. Since, the primary problem of CVS is segmentation, comparing the performance with the Neural Cellular Automata (NCA) that typically has only thousands of paprameters, is a must. A bit of search in Scholar revealed this paper where a foundation model is distilled into the NCA: https://link. springer. com/article/10.1007/s11548-025-03425-x 2.Figure 2 and 3 contains too much information that is not very easy to comprehend. Consider re-designing those.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Missing comparison with state-of-the-art.
- 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.
My concerns are not addressed properly. Either the authors sidestepped (specific point about a paper example rather than the general NCA as a method for segmentation/ classification) or used flowery language to convolute the triviality of the method. As such, I do not see any reason for this paper to be accepted.
Review #2
- Please describe the contribution of the paper
The paper proposes a lightweight model for CVS assessment that can run in real time on CPU hardware. The main idea is to compress a large teacher model into a 5M-parameter student by combining segmentation-based supervision, soft predictions from the teacher, and pseudo-labels from additional unlabeled data. The practical contribution is strong: the final model is much smaller (reducing the parameter count to 1.5%) while staying close to teacher performance and being much faster for deployment.
- 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 practical problem. Real-time CPU deployment is very relevant for surgical AI. The experiments are structured around clear ablations. In particular, Figure 4 shows the effect of the proposed components on models of different sizes, and Table 1 shows that the final 5M model remains competitive while being far more efficient in terms of inference speed.
Concretely, for the proposed 5M model, performance increases from 58.8 to 63.7 with segmentation supervision, then to 65.2 with distillation, and finally to 68.9 with additional pseudo-labeled data in mAP (teacher at 71.4).
- 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 methodological novelty is more limited than the paper presentation sometimes suggests. The final gain comes from combining known ingredients: segmentation-based auxiliary supervision, teacher-student learning, and pseudo-labeling. Among these, the gain from distillation itself seems modest (1.5%) compared with the gains from segmentation supervision and adding more pseudo-labeled data.
I also found the framing of RQ1 weaker than RQ2 and RQ3.The question of whether the CVS task has lower variability seems partly expected from the fact that the task is already restricted to a specific surgical phase and anatomical region. In addition, the “manifold” analysis is done in the embedding space of a specific pretrained encoder, so this is not an intrinsic property of the raw data itself, but a property of that chosen representation.
- 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 weak accept because the paper has clear practical value and strong relevance for deployable surgical AI. The authors show that a 5M-parameter model can remain competitive for CVS assessment while running in real time on CPU hardware, which is an important result. The experiments are solid and include meaningful ablations.
My concerns are mainly about novelty and positioning: the method is a combination of known ideas, and the effect of distillation alone is fairly small. I also found RQ1 weaker than the rest of the paper.
- 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.
Among all the reviews, the main concern appears to have been raised by Reviewer 1 regarding the missing comparison with an NCA segmentation model. However, the problem addressed in this paper is a multilabel classification problem, while segmentation is used only as an auxiliary task. Therefore, I do not consider this concern to be grounds for rejection.
My other concerns, namely the limited novelty and the limited effect of distillation, cannot be addressed in the rebuttal. Nevertheless, in my opinion, the work remains solid enough from the perspective of practical deployment, both in terms of accuracy and inference time.
Review #3
- Please describe the contribution of the paper
The paper’s main contribution is EdgeCVS, a practical framework for compressing a very large surgical AI model into a tiny edge-deployable model for Critical View of Safety (CVS) assessment during laparoscopic cholecystectomy, without losing much accuracy. The paper shows that accurate surgical AI for CVS can be made small, fast, and deployable on ordinary hardware by exploiting the task’s low visual complexity and using targeted distillation plus pseudo-labeling.
- 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 particularly strong aspect is the claim that CVS assessment lies on a compact, low-diversity visual manifold because it is restricted to a specific surgical phase and anatomy. This is interesting because it gives a principled reason why aggressive compression should work here, rather than treating model shrinking as just an engineering trick.
- The paper makes especially good use of unlabeled videos by restricting pseudo-label generation to the Calot’s triangle dissection phase, where the signal is most relevant. This is a strong data-centric idea: instead of adding arbitrary more data, the method adds task-aligned data density, which is more meaningful for a compact model.
- Compressing from a 305M-parameter teacher to a 5M-parameter student while staying close to SOTA performance is a major strength. The paper does not just show a smaller model works; it shows that a model with about 1.5% of the teacher’s size can still recover much of the teacher’s decision quality.
- The deployment story is unusually strong. The paper reports real-time CPU inference above 40 FPS, which directly supports the claim that the method could run on ordinary hospital hardware rather than requiring specialized GPUs.
- The method is compared against recent CVS challenge teams and prior Endoscapes baselines, and remains competitive despite using vastly fewer parameters. This strengthens the claim that the contribution is not just efficiency in isolation, but an efficiency–accuracy frontier improvement.
- 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’s strongest practical result is real-time CPU inference, which is important. But there is no prospective study, no user study with surgeons, no analysis of whether the system changes decisions in the OR, and no evidence that it improves safety outcomes. So, the “democratization of surgical AI” claim is only partially demonstrated.
- A big part of the gain comes from densifying the training manifold using teacher-generated classification and segmentation pseudo-labels. However, the paper does not provide much calibration, confidence filtering, or error analysis of those pseudo-labels. That matters because a compact student can also inherit systematic teacher mistakes.
- 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 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
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?
Although the paper does not yet demonstrate prospective clinical utility and lacks a thorough analysis of pseudo-label quality, it presents a technically sound and practically meaningful step toward deployable surgical AI by showing that phase-gated, task-aligned distillation can produce a CPU-real-time CVS model with competitive accuracy. For that reason, I judged as weak accept.
- 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.
The author’s rebuttal is reasonable, and they have made revisions in response to the comments. There is no change to the initial decision of “weak accept.”
Author Feedback
We thank all reviewers for their constructive feedback. We address the major concerns below.
Task definition (R1): R1 states “the primary problem of CVS is segmentation” and suggests missing benchmarking against tiny ML segmentation models. We respectfully clarify: the benchmarked task is CVS multi-label classification - predicting whether each Critical View of Safety (CVS) criterion is met, evaluated on SAGES-CVS and Endoscapes benchmarks, which define our SOTA scope. Segmentation is used exclusively as an auxiliary training signal; the segmentation head is discarded at inference and no segmentation metrics are reported. The NCA method (eNCApsulate) suggested by R1, addresses segmentation and depth estimation on capsule endoscopy - a different task, anatomy, modality, and benchmark - with no published results on any CVS dataset, making the comparison infeasible.
Missing tiny ML SOTA comparison (R1): EdgeCVS is not a tinyML models benchmarking study, it is a data-centric, dual-modality distillation framework that exploits task-specific manifold structure to enable extreme compression. The framework is backbone-agnostic; EdgeNeXT-Small was selected as a strong representative within the established 2.5-10M parameter edge deployment range. The correct evaluation is therefore vertical: how close does the student get to the teacher and SOTA on CVS benchmarks (Table 1 & Figure 4). To the best of our knowledge, no lightweight method has published results on SAGES-CVS or Endoscapes.
Manifold analysis (R3): We acknowledge that measuring dispersion in a pretrained embedding space is representation-dependent and falls short of intrinsic dimensionality (ID) metrics. However, classical ID analysis is not directly applicable to video data, as estimates are dominated by temporal autocorrelation between neighboring frames, violating the i.i.d. assumption these metrics require. Using the EVA02-Large embedding space, the source of the distillation signal itself, was our most principled available proxy. We will acknowledge this representation dependency as a limitation in the final manuscript.
Methodological novelty (R3): While individual components are established, their contribution lies in unification: MTL, KD, and phase-gated semi-SL are not independent modules but three expressions of a single data-centric dual-modality distillation process, coupling soft classification and segmentation predictions across seen and unseen phase-gated data to probe edge model compression limits. This unification produces a non-trivial result: in general vision tasks, compressing a large model to a small one keeps a significant gap, EdgeCVS substantially closes it on the CVS task. The modest +KD gain in isolation is precisely what reveals this: the bottleneck is not signal quality on existing annotated data but manifold coverage on unseen data, distinguishing EdgeCVS from standard distillation. The contribution is both a deployable surgical AI solution and a transferable hypothesis: task-specific manifold structure determines compression potential, a principle we demonstrate on CVS and propose for similar phase-gated tasks.
Pseudo-label quality (R4): To mitigate teacher noise, pseudo-label generation averaged logits across multiple training epochs to reduce variance. Both classification and segmentation pseudo-labels were generated as soft labels encoding the teacher’s probabilities, propagating the teacher’s uncertainty rather than noisy hard pseudo-labels. The consistent gains from +MTL and +SemiSL across all four model sizes in Figures 3 & 4 suggest soft pseudo-labels encode reliable signals. We will update the manuscript accordingly.
Prospective clinical validation (R4): The current work establishes the technical foundation, a deployable model retaining large teacher performance, that we hope will enable and inspire such clinical studies. We will revise the manuscript to frame clinical validation as future work enabled by this contribution.
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.
This paper studies the feasibility of a real‑time CPU‑based surgical AI for automated CVS assessment, with a clear clinical motivation and a strong emphasis on deployability. Reviewers acknowledge that demonstrating a 5M‑parameter model that remains competitive while being far more efficient, achieving real‑time CPU inference, is practically relevant. However, there is consensus that the methodological novelty is limited, as the approach largely combines known components; segmentation‑based supervision, teacher–student learning, and pseudo‑labeling, and that the gain from distillation itself appears modest relative to other components. For an application‑focused paper on tiny ML, a key concern is the missing comparison with relevant lightweight state‑of‑the‑art models, which weakens the positioning of the contribution. Additional concerns include that the manifold analysis is representation‑dependent, that analysis of pseudo‑label quality is limited, and that claims regarding broader impact are only partially demonstrated. Authors should address each concern raised by the reviewers, clarify their positioning, justify methodological choices, and appropriately contextualize claims.
- 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 convincingly addresses one major concern by clarifying that the task is classification (not segmentation), making some suggested comparisons less relevant. However, other key issues remain: limited novelty (combining known components like distillation, segmentation supervision, and pseudo-labeling), lack of comparison to lightweight baselines, and limited analysis of pseudo-label quality and broader impact. These points cannot be fully resolved by clarification alone. Despite this, the reviewers lean towards Acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
While there is clear motivation for the work, the paper still lacks technical novelty and clinical validation. The rebuttal addresses some of the raised concerns but not adequately to clear the major issues. Therefore, I recommend a rejection.
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.
After reading the paper and reviews post-rebuttal my opinion is for the paper to be accepted. I believe the work deals with an important and open problem of (edge deployment) with a method that presents interesting novelty. The validation is comprehensive.