List of Papers Browse by Subject Areas Author List
Abstract
Accurate recognition of anatomical structures is essential for safe and effective minimally invasive surgery (MIS), yet it remains underexplored in surgical computer vision due to limited annotated data and methods tailored primarily to natural scenes. In this work, we present a combined dataset and model framework to advance anatomy-aware perception in MIS. First, we introduce ATLAS-120k, a large-scale clip-level semantic segmentation dataset comprising over 120,000 annotated frames from 100 surgical videos spanning 14 procedures and multiple modalities, including laparoscopic and robot-assisted surgery. The dataset captures substantial procedural variability and was created using a scalable annotation pipeline that integrates expert manual labeling, automated propagation, iterative refinement, and surgeon verification to ensure high-quality annotations. Second, we propose ATLAS (Anatomy Recognition with Context Learning using Foundation Representations), a video semantic segmentation model specifically designed for surgical anatomy recognition. Unlike conventional approaches that emphasize object tracking, ATLAS leverages foundation-model embeddings together with lightweight temporal reasoning to incorporate contextual cues such as procedure type, surgical phase, and short-term visual memory. This design enables temporally consistent and accurate predictions while maintaining real-time feasibility. Together, the dataset and model establish a practical foundation for robust surgical scene understanding and support the development of clinically applicable guidance systems for minimally invasive surgery. The models, dataset annotations and annotation platform are publicly available at: https://github.com/TimJaspers0801/ATLAS.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3502_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/TimJaspers0801/ATLAS
Link to the Dataset(s)
https://huggingface.co/datasets/TimJaspersTue/ATLAS-120k
BibTex
@InProceedings{deRon_Surgical_MICCAI2026,
author = { de Jong, Ronald L. P. D. AND Jaspers, Tim J. M. AND Vervoort, Raf A. H. AND Bakker, Aron F. H. A. AND Li, Yiping AND Tolenaar, Jip L. AND Ruurda, Jelle P. AND Brinkman, Willem M. AND Pluim, Josien P. W. AND Breeuwer, Marcel AND de Geus, Daan AND van der Sommen, Fons},
title = { { Surgical Anatomy Recognition with Context Learning using Foundation Representations } },
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
his paper presents two complementary contributions: (1) ATLAS-120k, a large-scale surgical anatomy segmentation dataset covering 14 procedures and 120k annotated frames, and (2) ATLAS, a video semantic segmentation model that extends EoMT/VidEoMT with procedure and phase context queries built on surgical foundation model embeddings. The dual dataset-plus-model contribution is well-motivated, and the experimental results demonstrate clear improvements over competitive 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.
- Substantial dataset contribution. ATLAS-120k is the first surgical anatomy segmentation dataset spanning 14 procedures with both laparoscopic and robot-assisted videos, representing a meaningful advance over existing single-procedure datasets.
- Well-motivated method design. The incorporation of procedure and phase context queries directly targets a genuine domain gap, namely that surgical anatomy segmentation requires procedural awareness beyond pure visual tracking, and the motivation is clearly articulated.
- Comprehensive evaluation. The comparison covers natural-image, medical, and surgical foundation models with detection, segmentation, and temporal consistency metrics, providing a thorough picture of the state of the art.
- Real-time feasibility. 64 FPS on an H100 is a practically meaningful result for intraoperative guidance applications.
- 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.
- Annotation quality validation is insufficient. The annotation pipeline relies heavily on automated propagation with manual correction, yet no inter-annotator agreement metric is reported. Surgeon review is described only for the first frame of each clip, leaving the quality of propagated frames unverified beyond the correction step. Given that ATLAS-120k is a primary contribution, a more rigorous quality analysis, including an estimate of propagation error rates and correction frequency, is essential.
- Class imbalance and exclusion criteria are underreported. Training is restricted to 30 of 42 classes by excluding categories absent from all splits and merging semantically similar ones. The paper does not specify which classes were excluded or merged, nor does it report per-class performance.
- Ablation contains an inconsistency. In Table 3, adding temporal query propagation to the in-domain pretrained model increases AP (0.57→0.59) but decreases mDice (0.43→0.36) and mVC metrics. This counterintuitive drop in segmentation quality is not discussed. Understanding this trade-off is important for interpreting the overall contribution of temporal propagation.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Please see the detailed comments above.
- 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.
N/A
Review #2
- Please describe the contribution of the paper
The paper takes the dataset from the GSViT[19] work and annotates it and introduces a dataset for video anatomy segmentation, which compared to previous datasets is much bigger, includes clips from much more procedures and has images from both MIS and laparoscopy. In addition, the authors introduce a framework for video anatomy segmentation, ATLAS. This framework builds on previous work VidEoMT, and incorporates procedure- and phase- queries, in the pipeline. Each of which have been accompanied by their own cross-entropy based loss. The authors promise to make the model, dataset and annotation tools public upon acceptance.
- 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 presents a very important video anatomy segmentation dataset benchmark that will be beneficial to the entire community. The dataset has much higher quality than previous ones.
The idea of using context in video models, such as phase- and procedure-related information, is interesting.
- 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)The authors claim things that they do not evaluate/demonstrate. For example, they claim to introduce a real-time framework, #INTRODUCTION: “real-time anatomical segmentation”, yet they do not show any time-related metric.
2) The authors do not present any final loss function for their training; instead, they have a block of text in page 7, which is very non-standard, that literally describes the total loss they used. This is very unacceptable. The total loss must’ve been described in the Methods section, and coefficients of the loss function with mathematical symbols, and in the experiment setup, the values of those coefficients must have been outlined using their symbols; this is basic writing etiquette.
3) The main table of results, Tables 2 and 3, all look green and red. In academic papers, usually the use of a color/line is to indicate the best model, and maybe another one to indicate the second-best model, to outline the GAIN compared to the previous best model. Your tables do not look good visually, the color green is everywhere, you have then highlighted in red the worst performing model, I have never seen this in a paper before.
4) The in-domain pretraining was never mentioned in the Methods section, which is considered an important component of the framework, and only in the Experiments section; this is wrong. 5) Clarity and good presentation is important in a paper, I believe Section 2 to not be fully clear. For example, there is no specific pipeline figure to outline the main contribution of the pipeline, which is phase and procedure-specific queries. Fig1.Acts more as showing the difference between current and past approaches. Another example, It is unclear if authors have built upon VidEoMT or EoMT at the end (second paragraph of page 3, as they start talking about EoMT and then switch to VidEoMT in between.
6) In the Experiments section (4), when mentioning the split, the authors do not mention whether they make sure the distribution of procedures remains balanced or similar in the data splits or not, which can highly affect results and interpretations.
7) The authors have chosen to have page 7 with all text, and have the related results on the following page. Best practice would be to balance text/tables, and have the main table (page 8) where the analysis is being presented (page 7). 8) The Dino ablations do not really make sense to perform, as we already knew DinoV3 performs better; the authors could have done other ablations, such as a change of baseline.
- 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 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 think the dataset is a big contribution; on the other hand, the paper has serious presentation/clarification issues, which the authors must explain. In general, the methodology related to the ATLAS model (not dataset) should be further explained.
- 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.
Firstly, I would like to thank the authors for their rebuttal letter.
The authors have acknowledged textual/presentation issues flagged in the comments and have sought to change them in the camera-ready version; minor ambiguities were also clarified.
Review #3
- Please describe the contribution of the paper
This paper introduces (1) a semantic segmentation surgical dataset ATLAS-120k, that spans across multiple procedures and modalities for minimally invasive surgery, and (2) a video semantic segmentation model that extends EoMT with temporal reasoning and context queries.
- 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 introduction of ATLAS-120k addresses an important limitation in existing surgical datasets. The dataset is notable for its diversity, with annotations spanning many procedures and anatomical classes, and for its human-in-the-loop annotation pipeline involving expert supervision and verification. This makes it a valuable resource for advancing surgical vision tasks. 2.Inclusion of both laparoscopic and robotic surgery is a meaningful strength, this allows for robustness, generalizability, and domain-adaptation experiments of surgical vision models. 3.The proposed ATLAS architecture is well formulated from clinical and technical perspectives. By explicitly modeling procedure and phase queries, the design reflects the idea that anatomical visibility and interpretation depend strongly on the context of the current procedure, and benefits surgical segmentations.
- 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.While the experiments show that ATLAS outperforms the compared baselines, the absolute performance remains modest, with the best reported results at AP 0.64 and mDice 0.49.Although authors state that the low mDice score proves the challenge of the dataset, this counters the technical contributions of the proposed method in the class-imbalanced and long-tail distribution setting. The paper does not report results of experiments conducted on other surgical segmentation benchmarks, so it is difficult to determine that ATLAS has extended capabilities in anatomy recognition and scene understanding. 2.The dataset construction and annotation workflow require further clarification. It remains unclear how extensive the human supervision was beyond the first frame and how propagation errors were identified and corrected in practice. Additionally, the object tracking model was iteratively fine-tuned in the annotation process, but the metrics are not reported for the performance of initially annotated frames. Since the tracking model is not designed specifically for the medical domain, errors in the initial propagation could potentially bias later stages unless corrected carefully. Therefore, further explanation of the annotation quality control would be helpful.
- 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?
The authors introduce a large-scale dataset for surgical anatomy segmentation, which addresses an important limitation in existing surgical datasets, while the context-aware ATLAS model provides a domain-specific baseline for future research.
- 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 all reviewers (R1–R3) for their careful feedback. All three recognize ATLAS-120k as a meaningful contribution and appreciate the context-aware design of ATLAS. We have grouped the reviewer feedback into the following sections.
Reproducibility and Real-Time Claim [R1 (general), R2 (1)]
Two points may not have been clear: (1) R1’s reproducibility concern: the code, models, annotations, and annotation platform will be publicly released upon acceptance (Abstract, final sentence). (2) R2’s real-time concern: ATLAS ViT-L achieves 64 FPS on an H100 GPU (Section 4, page 7), supporting our real-time claim.
Annotation Quality and Pipeline Reliability [R1 (1), R3 (2)]
The annotation pipeline was human-in-the-loop throughout the entire process, not only at initialization. Propagated annotations for all frames were systematically inspected and corrected by surgical research fellows before being accepted. The tracking model was iteratively refined at 10k, 25k, and 50k frames, each followed by additional correction rounds to mitigate drift and ensure consistency. Statistics on correction frequency were not recorded during dataset construction. However, annotation quality was ensured through a final surgeon review of the first frame, followed by a final inspection of all frames by a research fellow.
Class Exclusion and Per-Class Reporting [R1 (2)]
Classes appearing in fewer than three videos were excluded because, under the patient-level split (70/10/20 videos), such classes would not reliably appear across all splits, making evaluation unstable or infeasible. Fine-grained subclasses with insufficient coverage were merged into parent categories (e.g., specific arteries → “artery”, specific ducts → “duct”). Per-class results are available but omitted due to space constraints; they will be released on GitHub upon acceptance.
Ablation Inconsistency: Temporal Propagation and mDice [R1 (3)]
R1 rightly flags this as counterintuitive. While we cannot draw hard conclusions, this may be explained by how the metrics differ. AP rewards consistent instance detection, so a propagated query that reliably localizes an instance can improve AP even with imprecise boundaries. mDice, however, is a pixel-level metric aggregated over the full frame, making it more sensitive to boundary drift in propagated frames. The full ATLAS model recovers and surpasses all metrics (AP: 0.64, mDice: 0.49), suggesting context queries counteract this effect. We will clarify this in the final version.
Presentation and clarification [R2 (2-8)]
We acknowledge these presentation issues, which require only minor textual and structural revisions. We will move in-domain pretraining and the training loss to the Methods section, formalize the loss as an equation, clarify procedure balance across splits, improve the EoMT→VidEoMT→ATLAS description, and balance tables and text. Figure 1 will be revised into a clearer pipeline diagram. The DINO ablation results will be retained as they show meaningful differences in mDice and mVC. On table formatting: color gradients were used to convey nuanced relative performance; we will consider revising toward standard best/second-best highlighting.
Generalization and Benchmark Scope [R3 (1)]
Our primary contribution is demonstrating consistent improvements over strong baselines in the challenging multi-procedure setting of ATLAS-120k, the only available dataset of this scope. We did not specifically target long-tail class imbalance and did not intend to imply this. We do hypothesize that context queries may help identify underrepresented classes, as the additional context could guide the model toward anatomy that is rare overall but expected within a specific procedure or phase. However, we did not investigate this explicitly and consider it an interesting direction for future work. We performed a zero-shot evaluation on other datasets, but it was omitted due to page limits and will be considered in future work.
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 introduces ATLAS‑120k, a large‑scale surgical anatomy segmentation dataset covering 14 procedures and multiple modalities, along with a video segmentation model that leverages procedure‑ and phase‑aware context queries. Reviewers consistently recognize the dataset as a major and valuable contribution, filling an important gap beyond single‑procedure benchmarks, and agree that incorporating procedural context is well motivated. Nonetheless, key concerns remain regarding limited validation of annotation quality, incomplete reporting of class exclusions and imbalance handling, and an unexplained ablation behaviour where temporal propagation improves AP while degrading segmentation metrics. Further issues relate to methodological clarity and presentation, including incomplete loss specification and claims that are not fully supported by the reported results. Clearer documentation of annotation reliability, improved methodological transparency, and ensuring the claims are fully supported by the evaluation would strengthen the overall contribution. Authors should address each concern raised by the reviewers. Moreover, authors should refine their technical writing to address the concerns about lack of experimental details, non standard notations, and figures.
- 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 most key concerns, especially by clarifying the annotation pipeline (human-in-the-loop corrections), explaining class exclusion and imbalance handling, and providing a reasonable explanation for the AP vs mDice inconsistency. It also commits to improving presentation issues (loss definition, figures, writing clarity), which were a major source of concern. Some limitations remain, particularly lack of quantitative annotation quality metrics and incomplete reporting, but these are mitigated by the detailed clarification and the strong dataset contribution. The reviewers agree on 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.
Accept
- Please justify your recommendation.
The significant data contribution of ATLAS-120k combined with the methodological contribution of the ATLAS model has been recognized by reviewers, and all post-rebuttal reviews 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.
After the rebuttal, all reviewers uniformly recommend acceptance.