Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Haustral folds can serve as important landmarks to localize and navigate colonoscopes through the colon. Fold edges can be utilized for tracking in 3D reconstruction algorithms to generate colonoscopy coverage maps and ultimately reduce missed lesions. Current haustral fold detection models struggle with debris-filled colonoscopy videos and fail to maintain high temporal consistency due to their single-frame input. We introduce HalF-SAM, a Haustral Fold detection model utilizing the Segment Anything Model (SAM) image encoder, which suppresses edges from specular reflection and fecal debris. The SAM2-based memory module enhances temporal consistency, which is essential for tracking. Our experiments have shown significant improvements in haustral fold extraction accuracy and stability. We also release a training dataset with automatically annotated haustral fold edges in debris-filled high-fidelity colon phantom videos. The dataset and code will be available at: https://github.com/DurrLab/HalFSAM.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4509_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/DurrLab/HalFSAM

Link to the Dataset(s)

N/A

BibTex

@InProceedings{GolMay_HalFSAM_MICCAI2025,
        author = { Golhar, Mayank and Huang, Luojie and Durr, Nicholas J.},
        title = { { HalF-SAM: SAM-based Haustral Fold Detection In Colonoscopy with Debris Suppression and Temporal Consistency } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15968},
        month = {September},

}

Reviews

Review #1

  • Please describe the contribution of the paper

    Haustral folds are the colonic structures protruding from the colon wall. Its irregular shape and its orientation makes it more complicated to identify it accurately in colonoscopy. This paper describes a method to detect such folds from the colonoscopy video using DL approach.

  • 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.

    From technical point of view the paper is well written. Reasonable amount of literatures have been considered. The presence of the image noise due to fecal traces, floating liquid, undigested food problems are well addressed, but not the other clinical cases w.r.t the morphology and the shape of the haustral folds.

  • 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.

    Comments to address

    1. Please use the references in order (serially)

    2. As per the Heuristic approach, colonic structures whose height is greater than or equal to 6mm is considered as Haustral folds and those with less than 8mm are considered as the suspicious smaller polyps. Did your model learn this domain knowledge? any cases and output?

    3. What is the result of the model in case of incomplete colon distension? Thickened Haustral fold?

    4. Normally, the Haustral fold in the transverse colon shows triangular appearance when compared to circular appearance in the ascending colon. How are those detected?

    5. The images used in figure 2 has the camera position normal to the centerline. What if the camera position is changed w.r.t x and y axis in the coordinate system in the colonoscopy video? Do you have any such case to show in the output of the model?

    6. Pedunculated polyps resemble to Haustral folds, how are these identified?

    7. When there is a uniform thickness in the Haustral fold, it is called the polyp on Haustral fold. Whether model detects this?

    8. Haustral folds can also show polyp. The irregular shape of the Haustral folds and its orientation makes it more complicated sometimes. Do you have any such cases and what is the output of the model?

    9. Even though the method looks promising, there are only few clinical cases considered in the model training and testing which limits the acceptance of this as a novel solution. Authors may try these cases mentioned above and may re-train and re-test the model to cross check whether the model addresses these rare cases.

    10. If you can provide the colonoscopy video as supplementary material, it becomes easy to understand the output of the DL model.

  • 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.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    NA

  • 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 — could be rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Even though the method and approach looks promising, only 1-2 clinical cases out of many are discussed in the paper. I feel its a borderline paper and needs a thorough rework and paper update.

  • 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 clarifications are provided for the review comments. Carefully went through all rebuttals. Now the paper can be accepted



Review #2

  • Please describe the contribution of the paper

    The paper presents a new deep-learning model to detect haustral folds in a colonoscopy video and a new annotated video dataset of haustral folds from a silicone colon phantom. The proposed model comprises a frozen Segment Anything Model 2 (SAM2) encoder, four trainable adaptors, and a trainable decoder to produce fold edges. Automated haustral fold detection is very important since polyps behind thick haustral folds are missed if the polyps are not seen in the endoscope field of view in the first place. The problem has not received as much attention as polyp detection or segmentation. The lack of ground truth exacerbates the problem.

  • 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 dataset and code will be released.
    2. The proposed HalF-SAM model improves performance on two F1-based metrics, ODS and OIS.
    3. The unique contribution of the model architecture is the extension of N memory banks, one for each of the N prior frames, to 4*N memory banks for per-level memory attention across the previous N frames. The new loss accounts for all BDCN losses at all four levels.
  • 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.

    Related Work

    • Missing early related works on non-deep model methods for haustral fold detection from a single image and across video frames.

    Hong et al. Colon Fold Contour Estimation for 3D Visualization of Colon Structure from 2D Colonoscopy Images. In Proc. of IEEE Int’l Symp. on Biomedical Imaging: From Nano to Macro (ISBI 2011)

    Hong et al. 3D Reconstruction of Virtual Colon Structures from Colonoscopy Images. Computerized Medical Imaging and Graphics. 38(1):22-33, Jan. 2014

    Implementation details

    • The original BDCN loss has w_side and w_fuse which are weights for the side loss and fusion loss. These values are not specified in the performance study.

    • Proposed Dataset: Please add the descriptive statistics about the dataset such as the mean and std. deviation per video, the number of videos without fecal debris, and the number of videos with fecal debris.

    Experimental results

    • Since there is no cleaning involved in the videos in this study, all frames for each video should have fecal debris in them. The memory attention should still find the colon fold edges with the debris. Or is it because the test dataset includes videos without debris as well. If so, the descriptive statistics on the proposed dataset is important to shed light into the results.

    • What were the values tried for the layer weights w_0 to w_32 before settling on the values indicated in the paper? Please comment on how sensitive are these weight values. The weights for w_0 and w_8 are the highest.

    • Table 1 shows that HalF-SAM does not give the best precision. An explanation 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.

  • 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/2025/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 — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper addresses a very important problem in this area. The dataset will be very valuable for the community if it is released. However, there are several questions as listed under the weaknesses that hinder the acceptance of the paper.

  • Reviewer confidence

    Very confident (4)

  • [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’s contributions are two-fold:

    1. It proposes HalF-SAM, an architecture based on pre-trained SAM2 model, custom adapters and custom memory attention modules, for edge detection of haustral folds during colonoscopy
    2. It features a new dataset of colonoscopy videos from high-fidelity silicone colon phantoms, with challenging conditions, for model training and evaluation. The paper presents very strong results in terms of edge detection metrics, compared to baselines, which include HalF-SAM without memory modules.
  • 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.

    High-quality combination of pre-existing work and novel techniques: utilization of pre-trained SAM2 with custom adapters and memory modules, which are demonstrated to actually improve edge detection performance, serves as a good example of combination of prior work and novel art in a meaningful way.

  • 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.

    Lack of comparison against any other methods for the same task, except DexiNed. If I understand correctly, edge detection can be re-formulated as semantic segmentation, where one tries to detect pixels belonging to the edges (albeit, this is a highly class-imbalanced case), and existing semantic segmentation architectures, such as SegFormer, can be used for the same task. It would be insightful to show how well such methods fare on this task.

  • 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.

  • 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/2025/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?

    While comparisons against competing methods may be lacking, strong ablation studies indicate that every addition to the basic pre-trained SAM2 model is meaningful and advances state-of-the-art, especially given how strong SAM2 is known to be in the first place.

  • 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 the reviewers for their insightful comments & constructive feedback. Dataset Details(R2): Our dataset includes paired clean(C) and debris-filled(D) colonoscopy videos as mentioned in Sec. 4.1, and is split as Train: 37C+37D; Validation: 8C+7D; Test: 12C+12D. Each debris-filled video contains fecal debris across all frames. Our videos have an average of 424 (SD: 175) frames. We will revise the dataset description with the above details. Evaluation Details(R2): The test set comprises 5,019 debris-filled frames and 5,019 clean frames, representing varied bowel preparation quality in real-world colonoscopies. For a better insight into detection in different cases, we will amend the presentation of Table 1 by breaking down the current experiment results into: HalF-SAM on Clean: ODS: 0.883, OIS: 0.883, AP: 0.397 HalF-SAM on Debris: ODS: 0.852, OIS: 0.852, AP: 0.374 Dataset Diversity(R3): Our dataset includes diverse haustral fold shapes (~30% triangular, ~45% circular, 25% irregular) and camera poses (parallel, normal, oblique, laterally shifted to centerline). The model is robust to such varied morphologies and viewpoints. The colon models include 12 polyps-1 pedunculated, 9 subtle near folds, and 2 large on folds. As our goal is to improve downstream feature tracking, detecting polyp edges as landmarks is beneficial. Since the model learns depth discontinuities, it detects prominent polyp edges (e.g., pedunculated, large polyp on-fold), while subtle flat polyps are not detected. We plan to add diverse video examples as a supplement. Additionally, we respectfully clarify that our paper focuses on methodological contribution, introducing the novel HalF-SAM and a preclinical dataset for robust haustral fold detection with the challenge of debris. With additional post-processing or ground truth modification, HalF-SAM could potentially adapt to various polyp analyses as suggested by R3. Rigorous validation on clinical conditions, such as colon distension, is planned for future studies. Hyperparameter(R2): We made a common modification to the BDCN Loss. Sec. 4.3 describes that w_0=1.5 and is equivalent to w_fuse in BDCN. The single w_side is replaced with varied loss weights from 4x to 32x layers: 0.7, 1.1, 0.7, 0.3, respectively. We ran a grid search for the weights. The highest w_0 is set to highlight the fused prediction, directly affecting output quality. W_8 is also high since haustral folds are most salient at 8x level, whereas 4x captures fine-grained but often artifacts, and 16x and 32x are mainly for coarse localization. HalF-SAM is more sensitive to the loss weights in the first 5 epochs, as it first optimizes the higher-weighted layers. The current setting helps to speed up convergence. However, the network converges to similar optima once all layers catch up, resulting in marginal performance differences in the final model regarding different weight choices. We clarify that w_0-w_32 refer to loss weights, and not layer weights. Within the model, we use a channel concatenation layer rather than fixed weights (Sec 3.1), which lets the model learn the optimal layer fusion from training. Precision & Recall Trade-off(R2): We have discussed the cause of the lower AP from HalF-SAM in 4.4, due to the thicker edge predictions. While baseline yields higher AP, its incomplete edges (Fig. 2) will hinder downstream tracking for 3D reconstruction. Comparative Methods(R4): We agree with R4 that general semantic segmentation models might work for haustral folds. Our model is built upon SAM, one of the strongest segmentation models. HalF-SAM further strengthens the segmentation backbone for haustral fold detection by integrating task-specific modules, including the multi-layer edge decoder & memory module, which are demonstrated in the ablation study (Table 1). We also compared HalF-SAM with existing models with proven SOTA performance (row 2). Related Work(R2): We appreciate the suggested papers & will cite them in the revision.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top