List of Papers Browse by Subject Areas Author List
Abstract
Temporal segmentation of colonoscopy withdrawal videos into anatomical regions is a key component for automated post-procedure reporting, enabling coverage assessment and structured documentation. While recent temporal segmentation models show promising accuracy, they typically treat colon segments as unordered categories, ignoring the inherent anatomical ordinality of the colon. This often yields temporally inconsistent predictions with frequent flickering and anatomically implausible jumps between non-adjacent regions.
We propose an ordinal-aware framework that injects anatomical priors into both training and inference. During training, we add an auxiliary ordinal regression objective that penalizes errors in proportion to anatomical distance. At inference, we apply an ordinal-constrained decoder that enforces anatomically plausible transitions.
On the REAL-Colon dataset, our method achieves 0.7621 weighted F1 and 0.6286 weighted Jaccard while substantially reducing label transitions to 15.84 per video. On the CAS-Colon dataset, it attains 0.7300 weighted F1 and 0.5935 weighted Jaccard without target-domain fine-tuning, indicating improved cross-dataset generalization. Overall, incorporating anatomical ordinality improves temporal consistency and supports clinically applicable, anatomically grounded segmentation. Code is available at [URL].
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3291_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/jangbi1/OColT
Link to the Dataset(s)
N/A
BibTex
@InProceedings{JanSeu_Ordinal_MICCAI2026,
author = { Jang, Seunghyun AND Lee, Dongheon AND Jung, Yoonsung AND Kim, Eue Young AND Shin, Na Mi AND Kim, Byeong Gwan AND Im, Jong Pil AND Kim, Sihyun AND Park, Chang Min},
title = { { Ordinal Priors for Colonoscopy Temporal Segmentation } },
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 introduces an ordinal-aware framework for temporal segmentation of colonoscopy withdrawal videos into seven anatomical regions. It adds an ordinal regression head (ORH) and boundary smooth loss during training, plus ordinal-constrained Viterbi decoding (OCD) at inference.
Results on REAL-Colon and cross-dataset CAS-Colon show improved wF1/wJacc and reduced label transitions over ColonTCN, ASFormer, and TeCNO
- 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.
- clinical motivation is reasonable; anatomical ordering is indeed a natural and sensable prior for temporal segmentation
- combination of ordinal regression objective and transistion constrianed decoding is clean
- this task-speicfic integration is suitable for MICCAI short paper
- 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.
-
both ORH and OCD are well-established and authors should not claim contribution on these two
-
Test sets of 12 (REAL-Colon) and 20 (CAS-Colon) videos are small. No confidence intervals or significance tests are reported, making it difficult to judge whether the improvements are robust or within noise.
-
ASFormer (BMVC 2021) and TeCNO (MICCAI 2020) are no longer state-of-the-art for temporal action segmentation. Comparisons with more recent methods would be substantially more convincing, e. g. DiffAct++ (Liu et al. , TPAMI 2025), sequence-to-sequence translation approaches (Behrmann et al. , ECCV 2022), multi-scale transformer models for surgical phase recognition (Zhang et al. , EMBC 2024), or action-aware segment modelling (He et al. , CVPR 2022).
-
The wF1 and WMAPE values reported here for ColonTCN on REAL-Colon differ noticeably from those in the original ColonTCN paper on the same dataset. Not sure why?
-
Also, the OCD enforces near-monotonic transitions, but endoscope movement does go back and forth during withdrawal; particularly around the cecum and ascending colon, as visible in the ground truth of Fig. 3(A). Will there be a over-smooth risk?
-
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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?
In sum, the paper addresses a real clinical need with a sensible approach, but the experimental evaluation needs strengthening and thus the weak reject recommandation.
I am open to upgrade my recommandation if the authors could provide satisfactory rebuttal addressing weakness especially the following:
(1) Please explain the discrepancy between your ColonTCN numbers and those in the original paper. (2) Can you provide per-class results, particularly for cecum/ascending where back-and-forth transitions are common? (3) Can you add comparisons with at least one recent temporal segmentation method (e.g. DiffAct++)?
- 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.
As noted in my review, the clinical motivation is strong and the proposed method is well grounded in clinical procedure. These are both genuine strengths.
The authors have also satisfactorily addressed my question on the reported discrepancy with prior work. My remaining concern is the reliance on a 5-year-old baseline; the paper has not fully convinced me that the method advances the segmentation state of the art.
Overall, I am laregely neutral and not against its acceptance.
Review #2
- Please describe the contribution of the paper
The paper introduces a methodological framework for temporal segmentation of colonoscopy videos across seven anatomical landmarks. It comprises (i) a distance-weighted ordinal regression loss that penalizes errors based on anatomical distance, and (ii) an ordinal-constrained decoding scheme that enforces anatomically plausible transitions. Evaluation on seven withdrawal-phase cases demonstrates improved performance compared to existing temporal segmentation methods.
- 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 methodological advance by explicitly modeling colon anatomy as an ordered sequence rather than as independent classes, addressing a key limitation of prior approaches. It introduces (i) a distance-weighted ordinal regression loss that incorporates anatomical relationships into learning (as clearly illustrated in Fig. 2); (ii) an ordinal-constrained decoding scheme that enforces anatomically plausible transitions; and (iii) a post-processing strategy that improves temporal consistency by reducing flickering and unrealistic label changes, as supported by the ablation study. The method is validated on the challenging REAL-Colon dataset and further evaluated on the CAS-Colon dataset to assess robustness without additional fine-tuning.
- 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.
Major concerns: (i) The primary limitation is the restricted clinical scope, as the method focuses only on the seven withdrawal-phase anatomical segments and does not extend to full-procedure analysis, thereby limiting its applicability for assessing key clinical indicators such as cecal intubation. Moderate concerns: (i) The ordinal-constrained decoding may suppress valid retrograde events that can naturally occur during colonoscopy (e.g., as illustrated in Fig. 3(A), showing the first appearance of the ascending colon segment). This limitation is not discussed; (ii) The framework is designed for offline, post-procedure analysis, with limited discussion of real-time applicability or deployment in clinical settings; (iii) The evaluation is limited to the withdrawal phase.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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
Minor observation: Add anatomical labels to Figure 3 to improve clarity.
- 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 paper presents a well-motivated methodological contribution by introducing ordinal priors into temporal segmentation, improving temporal consistency. The proposed distance-aware loss and constrained decoding are simple yet impactful additions that lead to stable and anatomically plausible predictions, as supported by experimental results and ablation analyses. However, the overall impact is somewhat limited by the restricted clinical scope, as the method focuses only on withdrawal-phase segmentation and does not address full-procedure analysis or key clinical endpoints.
- 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 rebuttal clarified several concerns regarding the evaluation protocol, the differences from prior ColonTCN settings, and the role of ordinal constraints as soft priors rather than strict monotonic assumptions. The authors also acknowledged the limitations associated with the withdrawal-only analysis and the offline nature of the framework. Although the clinical scope remains somewhat limited, I believe the paper makes a useful contribution to temporal offline colonoscopy analysis by integrating ordinal priors for anatomically consistent segmentation. However, the handling of clinically valid retrograde events remains an important limitation that warrants further discussion.
Review #3
- Please describe the contribution of the paper
This paper proposes distance-weighted ordinal regression losses and a transitionconstrained decoder that jointly enforce anatomical ordinality during training and inference.
- 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.
This paper studies a meaningful clinical problem, namely the temporal segmentation of colonoscopy withdrawal videos into anatomical regions. Its most interesting conceptual contribution is the way it reformulates the task: instead of treating the anatomical segments of the colon as independent, unordered categories, the authors model them as an ordinal label space with a natural proximal-to-distal ordering. This formulation is highly consistent with the anatomical structure of the colon and the actual withdrawal process, and it also provides a clear explanation for why conventional classification-based temporal segmentation methods often produce anatomically implausible jumps. Another strength of the paper is that the method is simple, targeted, and reasonably interpretable. Rather than introducing an overly complicated new architecture, the authors inject the ordinal prior into an existing temporal segmentation framework through three focused components: an auxiliary ordinal regression head, a boundary-aware smoothness loss, and an order-constrained decoder. As a result, the method is easy to follow and potentially easier to reproduce, adapt, and extend. In this sense, the work is a good example of improving a standard pipeline by explicitly leveraging domain structure. Importantly, the paper does not stop at merely modifying the loss function. Instead, it incorporates the ordinal prior at both the training and inference stages: during training, prediction errors are penalized according to anatomical distance, while during inference, the decoder suppresses implausible backward transitions and segment-skipping behavior. This makes the overall approach more than a simple auxiliary-loss trick, and turns it into a relatively complete ordinal modeling framework.
- 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 main weakness of the paper is that the methodological contribution appears more incremental than fundamentally novel. In essence, the work adapts existing ideas on ordinal learning and constrained decoding to the task of colonoscopy temporal segmentation, rather than introducing a genuinely new learning framework. This does not mean the work lacks value, since the task is clinically meaningful and the adaptation is sensible, but it does weaken the originality of the paper. Another important weakness is the lack of clarity regarding several key implementation details. The method introduces multiple important hyperparameters, including the loss weights, the boundary dilation radius, the temporal pooling window, and the backward used in the decoder. However, in the current manuscript, these quantities are mostly introduced only symbolically, while their concrete values and model-selection strategy are not clearly described. Since the final performance may be sensitive to these design choices, the omission weakens the reproducibility of the work. It also makes it difficult to assess whether the reported gains reflect a robust modeling improvement or depend substantially on careful hyperparameter tuning. For a methods paper, these details are important for both fair evaluation and practical reuse by others. A further weakness is that the overall framework depends on a strong structural assumption: namely, that the withdrawal process follows a near-monotonic anatomical progression, such that backward transitions are rare and segment skipping is anatomically implausible. This prior is clinically reasonable on average, and it is exactly what motivates the proposed constrained decoding. However, in real procedures, this assumption may not always hold strictly. Local revisits, unstable camera views, transient ambiguity, or uncertain segment boundaries may all produce short-term reversals or irregular transitions that are not purely errors. In such cases, the constrained decoder may potentially suppress predictions that are locally valid but globally disfavored by the prior. The paper would therefore be more convincing if it provided a more detailed failure-case analysis, especially clarifying when the constrained decoding helps and when it may incorrectly suppress true local backward motion or uncertain transitions.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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?
I gave this paper a weak accept because, although the methodological novelty is moderate, I think the paper addresses a meaningful clinical problem and proposes a well-motivated solution that is well aligned with the anatomy and the procedure workflow. The main idea of modeling colon regions as an ordered label space is sensible, and the paper consistently incorporates this prior in both training and inference rather than only adding a minor auxiliary loss. I also appreciate that the overall method is relatively simple and interpretable, which makes the contribution practically useful. At the same time, my score is not higher because the technical contribution feels more like a thoughtful adaptation of existing ordinal learning and constrained decoding ideas than a fundamentally new framework. In addition, several implementation details and hyperparameter choices are not clearly specified, which weakens reproducibility, and the paper would benefit from a clearer discussion of cases where the monotonic anatomical prior may not strictly hold. Overall, I found the work useful and reasonably convincing, but its originality and methodological depth are somewhat limited, which is why I settled on weak accept.
- 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.
After the rebuttal, I maintain my weak accept recommendation. The authors have reasonably addressed my main concerns by clarifying the evaluation protocol, implementation details, clinical scope, and the soft nature of the ordinal constraint. The added video-level bootstrap analysis also improves confidence in the reported gains. While the novelty remains moderate, the paper is clinically relevant and sufficiently convincing for acceptance.
Author Feedback
We thank the reviewers for recognizing the clinical motivation (R1, R2, R3) and the consistent integration of ordinal priors across training and inference (R2, R3). Our novelty is not in introducing new standalone components, but in a task-specific formulation of withdrawal-phase anatomical segmentation as an ordinal label space, implemented via ordinal supervision and ordinal-constrained decoding. We address the major concerns below. (A) ColonTCN discrepancy (R1). We thank the reviewer for flagging this point. The discrepancy with the original ColonTCN paper stems from different task definitions and evaluation targets. The original benchmark addresses full-procedure segmentation with a 9-class label space (including outside/insertion), uses ResNet50 features, and reports WMAPE as withdrawal-time estimation error. Our submission evaluates 7-class withdrawal-phase anatomical segmentation: non-withdrawal frames are excluded, GastroNet embeddings are used, and WMAPE measures anatomical segment-duration error. All baselines in Table 1 are retrained under the same split/feature/exclusion protocol for a fair comparison. (B) Requests for newer baselines (R1). We acknowledge the requests for comparisons to newer temporal segmentation methods. However, rebuttal rules do not allow new experiments, so we cannot add new baselines here. We instead clarify the existing protocol and note that robustness is already assessed via CAS-Colon evaluation and ablations in the manuscript. (C) Monotonicity and oversmoothing risk (R1, R2, R3). We agree that withdrawal is not strictly monotonic and that short local revisits can occur, particularly near cecum–ascending (Fig. 3(A) GT). Both components of our framework are designed as soft priors rather than hard constraints: the training-time ordinal objective penalizes anatomically distant errors without preventing local deviations, and ordinal-constrained decoding (OCD) applies tunable backward/skip penalties (Eq. (4)–(5)) so that strong evidence can override the prior. As shown in Fig. 3, OCD suppresses unsupported flickering and non-`adjacent jumps relative to both the ColonTCN baseline and our raw predictions. In the cecum and ascending, we consistently observe improved F1 over the ColonTCN baseline on both datasets (REAL-Colon: cecum +4.3pp, ascending +2.4pp; CAS-Colon: cecum +12.4pp, ascending +4.2pp). Nevertheless, we acknowledge a trade-off: overly aggressive decoding penalties could smooth brief valid retrograde events. We will discuss this risk and its mitigation in the final version. (D) Clinical scope and deployment (R2). We thank the reviewer for raising this important question regarding clinical applicability. Our framework targets withdrawal-phase region tracking for post-procedure reporting (documenting examined regions and lesion locations from the completed video), matched by the withdrawal-only CAS-Colon benchmark used for external validation. We acknowledge that this scope does not cover full-procedure endpoints such as cecal intubation assessment or real-time deployment during the procedure. We will clarify these scope limitations and discuss potential extensions as future work. (E) Reproducibility and implementation details (R1, R3). We agree that hyperparameters should be stated concretely. The submitted experiments used: CE (0.5), TMSE (0.15; threshold 4.0), boundary-smooth loss (0.5; radius r=5; KL), and ORH (λ_reg=0.3); OCD used α=10, β=5, free start, and pooling window w=3.Model selection used early stopping on validation wJacc with no test-set tuning. We will consolidate these into a concise table and, upon acceptance, release the code (links omitted per policy). For statistical robustness, paired, video-level bootstrap resampling (B=2,000) on the existing test outputs confirmed statistically significant improvements over ColonTCN in wF1, wJacc, and transitions (p<0.001). (F) Figure clarity (R2). We will add anatomical-segment labels to Fig. 3 in the final version.
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.
The paper has received 3 borderline reviews. All reviewers agree on the approach of anatomical ordering being interesting and clinically intuitive. Reviewers are also in agreement on the paper’s weaknesses criticising
1) The limited novelty of the method which is mostly an aggregation of existing techniques
2) The limited comparison to SOTA
3) Key information on methodological design and choices are unclear.
My recommendation is for the paper to be invited for a rebuttal
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
All 3 reviewers are overall satisfied with the rebuttal and recommend acceptance. In the final paper, the authors are strongly advised to:
1) Explicitly state the innovation of the “task-specific formulation of withdrawal-phase anatomical segmentation” and adjust the claims for novelty of the ORH and OCD modules
2) Acknowledge the requirement for a more comprehensive comparison to more recent baselines
3) Clearly state and discuss the trade-off: overly aggressive decoding penalties could smooth brief valid retrograde events.
4) Integrate the rest of their answers in the rebuttal
I recommend 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 paper addresses a clinically meaningful problem in colonoscopy video analysis (temporal segmentation of withdrawal videos into anatomical regions). The main merit is the clinically intuitive reformulation of colon regions as an ordered anatomical sequence rather than independent classes. The rebuttal effectively clarifies several concerns regarding the evaluation protocol, implementation details, and the soft nature of the ordinal constraint. Overall, the paper makes meaningful contribution to anatomically consistent temporal segmentation in colonoscopy videos and is suitable for 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.
All reviewers unanimously accepted this work post-rebuttal.