List of Papers Browse by Subject Areas Author List
Abstract
Learning robust representations of polyp tracklets is key to enabling multiple AI-assisted colonoscopy applications, from polyp characterization to automated reporting and retrieval. Supervised contrastive learning is an effective approach for learning such representations, but it typically relies on correct positive and negative definitions. Collecting these labels requires linking tracklets that depict the same underlying polyp entity throughout the video, which is costly and demands specialized clinical expertise. In this work, we leverage the sequential workflow of colonoscopy procedures to derive self-supervised associations from temporal structure. Since temporally derived associations are not guaranteed to be correct, we introduce a noise-aware contrastive loss to account for noisy associations. We demonstrate the effectiveness of the learned representations across multiple downstream tasks, including polyp retrieval and re-identification, size estimation, and histology classification. Our method outperforms prior self-supervised and supervised baselines, and matches or exceeds recent foundation models across all tasks, using a lightweight encoder trained on only 27 videos. Code is available at https://github.com/lparolari/ntssl.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0772_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/lparolari/ntssl
Link to the Dataset(s)
REAL-Colon: https://doi.org/10.25452/figshare.plus.22202866
SUN database: http://amed8k.sundatabase.org/
PolypSize: https://doi.org/10.6084/m9.figshare.28030115
PolypsSet: https://dataverse.harvard.edu/citation?persistentId=doi:10.7910/DVN/FCBUOR
BibTex
@InProceedings{ParLuc_Contrastive_MICCAI2026,
author = { Parolari, Luca AND Gori, Pietro AND Ballan, Lamberto AND Biffi, Carlo AND Le Folgoc, Loïc},
title = { { Contrastive Learning under Noisy Temporal Self-Supervision for Colonoscopy Videos } },
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 manuscript proposes a self-supervised framework for learning robust polyp tracklet representations in colonoscopy videos. The method exploits the sequential nature of colonoscopy procedures to derive positive associations from temporal proximity between tracklets. Extensive evaluation across four downstream tasks (retrieval, re-identification, size estimation, histology classification) on four public datasets, demonstrating superiority over prior self-supervised/supervised baselines and competitiveness with large-scale foundation models using a lightweight encoder trained on only 27 videos.
- 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 observation that colonoscopy workflows are inherently sequential provides an intuitive and practically grounded source of self-supervision. This is particularly valuable in medical imaging where annotation costs are prohibitive. 2.The max-based contrastive loss (Eq. 1) is an elegant solution to the label noise problem inherent in temporal associations. By allowing the model to “ignore” noisy positives rather than forcing all bag members to align, the formulation better reflects the uncertainty of the self-supervision signal. This is more principled than heuristic re-weighting schemes. 3.The method achieves state-of-the-art performance across all four downstream tasks. 4.The analysis of exponential sampling, curriculum learning, and multi-level objectives provides clear evidence for each component’s contribution. 5.The lightweight architecture and minimal data requirements address deployment constraints in clinical settings, where computational resources and annotated data are often limited.
- 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.Novelty of temporal self-supervision requires clearer positioning: While applying temporal proximity as a supervisory signal to colonoscopy is novel, the broader concept is well-established in video representation learning. Works such as MoCo-v3 and temporal instance discrimination methods have long exploited temporal continuity. The paper should more explicitly differentiate its contribution from these prior approaches, particularly regarding domain-specific adaptations for endoscopic imagery (e. g. , handling deformation, occlusion, instrument interference). 2.The max-based objective resembles strategies in Miech et al. [21] for learning from uncurated instructional videos, where not all temporally co-occurring pairs are reliable. 3.All experiments use tracklets constructed from ground-truth detections, completely decoupling representation learning from detection and tracking errors. In real colonoscopy pipelines, detection/tracking failures are frequent due to motion blur, occlusion, and instrument interference. 4.The condition “at least one tracklet in Bi should be more similar to ai than any tracklet in Bi is to any other anchor aj” is relaxed to the point that it may provide insufficient gradient signal, especially when bags are highly impure. The paper does not analyze the gradient dynamics or provide theoretical justification for why this objective leads to better representations than simpler alternatives (e. g. , label smoothing, robust aggregation). 5.The comparison with foundation models focuses on parameter count and FLOPs but does not account for differences in pretraining data scale, augmentation strategies, or task-specific fine-tuning protocols. Claiming superiority over DinoV2-giant based solely on downstream metrics while ignoring the massive pretraining gap is misleading.
- 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.
(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?
While the paper addresses a relevant problem and presents a coherent framework, the novelty of the core contributions is limited relative to existing literature in temporal contrastive learning and robust optimization. The evaluation, though broad, relies on idealized oracle tracklets that do not reflect real-world deployment conditions, and statistical reporting is insufficient to support the claimed improvements.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The main contribution of the paper is the development of a self-supervised model for the representation of polyp tracklets, and in particular the use of “noisy temporal structure” (closeness in time but allowing for some mismatches) as a supervision signal.
- 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 use of a noisy supervision signal (and the modification of the loss function to encode mismatches between anchors) is novel and particularly well-suited for this data. The method deals well with realistic colonoscopy video data, as shown through its performance on downstream evaluation tasks. In terms of broader impact, the main idea of the paper is likely to apply to other kinds of sequential data as well.
- 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 several empirical choices in the paper which work well in the given setting but may not transfer to other contexts, for example, for the bag construction, the use of the exponential distribution and the choice of the temperature parameter. This is only a minor criticism.
The comparison with other methods is good but not exhaustive (for example, no comparison with recent video SSL methods).
- 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
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.
(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?
Overall a very solid paper, with a clear methodological contribution that is backed up by good results.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes a self-supervised method for learning polyp tracklet representations in colonoscopy videos. The core idea is to avoid relying on costly manual annotations to determine which tracklets belong to the same polyp, and instead exploit the temporal structure inherent in colonoscopy procedures to construct a self-supervision signal: tracklets that are temporally close are more likely to correspond to the same polyp. Since such temporal associations are not always correct, the authors further design a noise-aware contrastive loss that enables the model to learn from noisy temporal positives while reducing the adverse impact of incorrect associations. During training, the method also incorporates a temporal rank-based sampling strategy and curriculum learning, gradually transitioning from “safer but less diverse” nearby associations to “more diverse but noisier” long-range associations.
Experimentally, the paper evaluates the method on four downstream tasks, including polyp retrieval, re-identification, size estimation, and histology classification. The results show that the proposed approach not only significantly outperforms existing self-supervised methods, but also overall surpasses the supervised methods compared in the paper, while matching or exceeding the performance of foundation models on multiple tasks. At the same time, it uses only a lightweight encoder and is trained on just 27 videos, indicating strong practical value and promising scalability for annotation-scarce colonoscopy scenarios.
- 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 problem setting is highly practical: it avoids relying on costly and highly specialized polyp entity annotations, and instead directly exploits the inherent temporal order in colonoscopy procedures to construct a self-supervision signal. This substantially lowers the barrier to training data acquisition and makes the method more suitable for scaling to real clinical scenarios.
The method design is well targeted: rather than naively treating temporally close tracklets as always-correct positive pairs, the authors explicitly recognize that such associations can be noisy, and therefore introduce a noise-aware contrastive loss. This allows the model to tolerate incorrect temporal associations and prevents the representation learning process from being misled.
The training strategy is well designed: through temporal rank-based sampling and curriculum learning, the model starts from nearby samples that are safer but less diverse in the early stage, and gradually incorporates more distant samples that are more diverse but also noisier. This design effectively balances stability and diversity.
The experimental results are convincing and efficient: the method performs well across four downstream tasks, including retrieval, re-identification, size estimation, and histology classification. It outperforms existing self-supervised and supervised baselines, and matches or exceeds foundation models on multiple tasks. At the same time, the model itself is lightweight and is trained on only 27 videos, indicating strong efficiency and a favorable performance-to-cost ratio.
- 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 temporal assumption is relatively strong: the method relies on the premise that tracklets that are temporally close are more likely to belong to the same polyp, but this assumption may not remain stable in scenarios with multiple polyps appearing consecutively, repeated revisits, or more complex procedural operations.
The experimental setting is somewhat idealized: the tracklets are constructed from ground-truth detections, and the paper does not yet sufficiently validate the end-to-end performance under realistic conditions such as detection errors, tracking drift, and missed detections.
The generalization evidence is still limited: although the method achieves promising results when trained on only 27 videos, the training scale remains relatively small, and more empirical evidence is needed to support its robustness across centers, devices, and more complex clinical distributions.
The mechanism analysis could be further strengthened: while the paper includes ablation studies, the stability of the noise-aware loss under different noise levels, bag purity conditions, and parameter settings is still not analyzed in sufficient depth.
- 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?
I recommend a Weak Accept because the paper addresses a practical and clinically relevant problem, and proposes a technically sound self-supervised framework that is well aligned with the characteristics of colonoscopy videos. In particular, leveraging temporal structure as supervision is a reasonable and useful idea, and the proposed noise-aware contrastive loss is a thoughtful design choice for handling imperfect temporal associations. The experimental results are also strong: the method performs consistently well across multiple downstream tasks, surpasses prior self-supervised baselines, and remains competitive with or better than stronger supervised and foundation-model baselines, while using a lightweight architecture and limited training data. At the same time, I do not think the paper is strong enough for a clear accept, mainly because the approach relies on a relatively strong temporal assumption, the experimental setup is somewhat idealized due to the use of ground-truth detections for tracklet construction, and the analysis of robustness and generalization is still not sufficiently thorough. Overall, I believe the strengths outweigh the weaknesses, but the paper would benefit from deeper validation and analysis.
- Reviewer confidence
Somewhat confident (2)
- [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 sincerely thank the reviewers and the meta-reviewer for the time dedicated to review our work and for the valuable feedback they provided. We are pleased by the early acceptance to MICCAI 2026.We will take the reviewers’ suggestions into consideration in our future work.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper proposes a temporal contrastive framework for colonoscopy videos that uses temporal closeness as a cheap, noisy supervisory signal and reshapes the contrastive loss to handle realistic violations of this signal, such as multi-polyp scenes and revisits.
The idea is that tracklets that are temporally close are likely to depict the same polyp, and the loss is modified to encode mismatches between anchors instead of treating positives as hard. The novelty over standard temporal contrastive learning is concrete.
Two reviewers endorse the work. R2 (Accept) calls it “a very solid paper, with a clear methodological contribution backed up by good results.” R3 (Weak Accept) emphasizes the practical value of avoiding polyp-entity annotation and judges the framework technically sound. R1’s Weak Reject says novelty is “limited relative to existing literature in temporal contrastive learning and robust optimization”, but does not cite any specific prior work the authors should have positioned against. Without that, the negative review does not outweigh the two positive ones.