List of Papers Browse by Subject Areas Author List
Abstract
Accurate polyp size stratification guides surveillance decisions, with lesions larger than 5 mm typically requiring closer follow-up. However, monocular colonoscopy lacks a reliable metric reference. We present a diagnostic audit of binary polyp size classification (≤ 5 mm vs. > 5 mm) across multiple public multi-center datasets, model families, and polyp-stratified cross-validation. Across architectures and input modalities, including RGB appearance, relative depth, and photometry, model performance is moderately consistent, suggesting reliance on cues correlated with examination behavior rather than true metric scales. By providing ground-truth scale at varying granularities, we quantify the potential improvement from perfect scale information and show that current depth estimation and global calibration offer limited gains. We further demonstrate that segmentation errors under distribution shift eliminate most of this potential, with oracle scale under predicted masks recovering only baseline performance. These results highlight metric scale and mask robustness as two independent bottlenecks and provide reusable evaluation tools—such as oracle scale ladders, shortcut partitions, and mask substitution—for auditing future polyp sizing pipelines. Our code is publicly accessible at https://github.com/anaxqx/polyp-sizing-audit.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5951_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/anaxqx/polyp-sizing-audit
Link to the Dataset(s)
SUN Colonoscopy Video Database: https://www.nature.com/articles/s41597-024-03359-0
REAL-Colon dataset: https://www.nature.com/articles/s41597-024-03359-0
BibTex
@InProceedings{XioXin_Understanding_MICCAI2026,
author = { Xiong, Xinqi AND Dunn Beltran, Andrea AND Choi, Junmyeong AND McGill, Sarah K. AND Niethammer, Marc AND Sengupta, Roni},
title = { { Understanding Model Behavior in Monocular Polyp Sizing } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16892},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper evaluates how well existing models estimate polyp diameter (size) from colonoscopy videos for two polyp-diameter categories (at least 5 mm or above 5 mm). The challenges in automated polyp size estimation stem from (1) the lack of depth information in monocular video cameras used for colonoscopy and (2) small polyps are inspected close-up, while large polyps are viewed from a distance.
The paper reports room for improvement if an appropriate scale factor is multiplied by the estimated size, as follows. Metric size S = camera_distance*estimated pixel-wise polyp diameter/focal length. The scale factor is camera_distance/focal length that needs to be estimated. In this work, the estimation was somehow done. The paper does not describe the method used to estimate the scale factors.
Given an oracle scale factor per frame, per polyp, and per procedure, improvement was shown on their benchmark dataset curated from two public multi-center datasets with endoscopist-reported visual size estimates, Real-Colon [4] and SUN-SEG [23], totaling 232 unique polyps from 159 patients. Five-fold cross-validation was performed to obtain Macro-F1 and Recall for (size > 5mm) as performance metrics. Models with metric depth [22] and relative depth [24], and models without any depth information (ResNet18, ViT-B, and a 2-layer MLP with 51 handcrafted geometric and photometric features) were evaluated. The authors trained all these models using the same hyperparameter values.
- 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 is well-motivated and underexplored. -The paper shows that there is room for improvement in automated polyp metric size estimation with the scale factor per frame or per polyp. -Table 1 shows that the per-frame oracle scale factor offers the greatest improvement of 0.16 macro-F1 and 0.14 >5mm recall when using CNN3 with relative depth. Nevertheless, even with the scale factor, Macro-F1 and recall remain below 0.90 on this test dataset.
- 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.
- Reproducibility is the major issue. No distribution of the number of unique polyps with > 5mm and <= 5mm. No citations on existing works of the 2-layer MLP on 51 handcrafted features or on how these features were computed.
- How to estimate the scale factor, per-frame (Oracle-frame), per-polyp (Oracle-Polyp), and per clip (Global)? Eq. 1 was used, but it is unclear how these scaling factors were determined and the experience of the experts who determined them. Or whether it was a trial and error. It was not clear how to generate the 20% bounding box perturbation.
- No mentioned of code availability to reproduce the experiments.
- No report of scale factor per clip, average per polyp, or average per frame.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper investigates an interesting, underexplored problem that, if satisfactorily solved, would have a significant clinical impact. However, reproducibility and technical novelty are the major issues. Given that the paper does not propose new methods to address the issue raised, some artifacts, such as code or a description of the method they used to estimate scale factors at different levels, and the actual scale factors should be available for other researchers to explore more on other datasets. The paper described the lack of accessible code for recent polyp size estimation methods, but this paper has the same issue.
- 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 addressed the critical point of how to obtain the deterministic scaling factor from PPSNet’s estimated depth (d_bar_i) for a frame-specific correction factor c_i = (S_p/A_i) / d_bar_i, Oracle-frame, and Oracle-polyp. The final version needs to add these clarifying answers. The public availability of the code, datasets, and tools addresses concerns about the reproducibility of the proposed work and enhances its impact. Statistical tests are important and must be included in the camera-ready version.
Review #2
- Please describe the contribution of the paper
The primary contribution of this paper is a systematic diagnostic audit of the fundamental limitations inherent in monocular polyp size classification. Rather than proposing a new performance-driven model, the authors evaluate multiple model families and input modalities—including RGB, relative depth, and photometry—to demonstrate that current methods converge to a common performance ceiling. This ceiling is shown to be driven by shortcut learning tied to examination behavior rather than true metric scale reasoning. A key strength of the work is the introduction of a reusable diagnostic framework, featuring oracle scale ladders, shortcut-consistent partitions, and segmentation mask substitution. Through these controlled interventions, the study successfully disentangles metric scale ambiguity from segmentation robustness, identifying them as two independent bottlenecks that must be addressed in future monocular sizing research.
- 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.Clear and Principled Diagnostic Framing of a Clinically Critical Problem Instead of proposing yet another end-to-end model, the paper reframes monocular polyp sizing as a fundamental diagnostic challenge. By shifting the focus from “performance chasing” to a rigorous causal analysis of why current models plateau, the authors address a critical gap in clinical AI. This approach is particularly valuable in tasks where shortcut learning can lead to erroneous patient management decisions.
2.Systematic Disentanglement of Confounders via Controlled Interventions The study meticulously isolates three primary bottlenecks—metric scale ambiguity, segmentation robustness, and dataset bias—through targeted probes and oracle interventions. The introduction of the “Oracle Scale Ladder” (global, per-polyp, and per-frame) is a particularly effective methodological contribution. It provides a quantifiable measure of the potential headroom attributable to scale information, revealing insights that are typically obscured in standard evaluation pipelines.
3.Compelling Empirical Evidence of Behavior-Driven Shortcut Learning By partitioning test frames into shortcut-consistent and shortcut-inconsistent subsets, the authors provide strong evidence that models rely on endoscopist behavior (e.g., closer proximity for larger lesions) rather than true geometric reasoning. The significant performance gain observed in shortcut-inconsistent frames (+40.1 pp) under oracle-scale conditions is a particularly convincing demonstration of this phenomenon, exposing the limitations of current “black-box” approaches.
4.Robust and Clinically Realistic Evaluation Protocol The use of multi-center public datasets, patient-stratified cross-validation, and frame-level analysis reflects a deep awareness of real-world clinical variability. Notably, the paper goes beyond in-domain benchmarks by evaluating segmentation under distribution shift, revealing that high segmentation scores on curated datasets often fail to translate into robust downstream sizing performance.
5.Reusable Diagnostic Tools with Broad Applicability The proposed diagnostic probes—including oracle scale ladders, shortcut partitions, and mask substitution tests—represent a significant contribution to the broader medical imaging community. These tools provide a concrete, reproducible framework that researchers can adopt to audit model behavior and ensure clinical readiness in other tasks affected by scale ambiguity.
6.Transparent and Well-Supported Conclusions The manuscript is characterized by its scientific honesty, clearly distinguishing between theoretical potential (under oracle conditions) and practical achievability. This transparency, backed by rigorous data, strengthens the credibility of the findings and establishes the work as a reliable and high-quality reference point for future research in monocular depth and scale recovery.
- 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.Limited Methodological Novelty in Model Architecture and Paradigms While the paper provides a compelling diagnostic analysis, the underlying model architectures and learning paradigms are largely based on existing approaches. The RGB and depth-based probes closely follow previous work on monocular sizing and depth-assisted classification (e.g., Itoh et al., 2018, 2021), and the photometry-based features build on established illumination cues (e.g., Paruchuri et al., 2024). Consequently, the core novelty of this work resides in its evaluative and analytical framework rather than in the introduction of fundamentally new modeling techniques or learning objectives.
2.Binary Size Formulation and Limited Clinical Granularity The study focuses exclusively on a binary classification at the 5 mm threshold. While this is a clinically meaningful cut-off, it does not capture the full spectrum of polyp size stratification required in practice (e.g., 6–9 mm vs. $\ge10$ mm). Although the authors justify this choice based on class imbalance, the current findings may not directly generalize to more fine-grained or continuous size estimation tasks, which have been explored in more complex reference-based or reconstruction-based methods (e.g., Azagra et al., 2023; Du et al., 2024).
3.Potential Confounding from Subjective Visual Estimates in Ground Truth Labels The size annotations in the evaluated datasets are derived from endoscopists’ visual estimates rather than tool-verified, objective measurements. As documented in clinical literature (e.g., Chaptini et al., 2014), such estimates are prone to systematic bias and rounding errors, particularly near clinical thresholds. This inherent label noise may act as a confounding factor, potentially contributing to the observed “performance ceiling” and complicating the interpretation of oracle-scale gains around the 5 mm decision boundary.
4.Underutilization of Temporal Dynamics in Video Data Despite the datasets consisting of video sequences, the analysis remains largely frame-based, leaving temporal consistency and long-range dynamics unexplored. Prior research (e.g., Li et al., 2025) has demonstrated that temporal cues can stabilize depth and size estimation in endoscopy. Incorporating temporal aggregation or sequence-level reasoning could potentially mitigate the “flickering” of predictions or some of the identified bottlenecks, yet this direction is not addressed in the current study.
5.Indirect Implications for Real-World Clinical Deployment While the paper convincingly diagnoses the failure modes of current monocular sizing, it stops short of providing a practical roadmap toward real-time clinical deployment. The oracle experiments, while theoretically illuminating, rely on information that is fundamentally unavailable in routine workflows. Furthermore, the study lacks an evaluation of latency, computational overhead, or integration constraints, which are critical factors for determining the bedside feasibility of such diagnostic tools.
- 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
This is a highly thoughtful and methodologically rigorous study that provides a much-needed shift in focus: from incremental performance chasing to a principled diagnosis of why monocular polyp sizing remains fundamentally challenging. The paper excels in its experimental design, particularly in its ability to disentangle the confounding effects of metric scale ambiguity, segmentation robustness, and procedure-induced shortcuts. By reframing the problem as a diagnostic audit, the authors offer a refreshingly honest and scientifically grounded perspective on the current state of the field.
The primary strength of this work lies in the introduction of reusable diagnostic tools—such as the oracle scale ladders and shortcut partitions—which are likely to have a lasting impact beyond this specific application. The clarity with which the authors isolate independent bottlenecks (scale vs. mask quality) provides a solid foundation for future research. This “Audit-first” approach sets an important benchmark for how monocular sizing methods should be critically evaluated before any claims of clinical readiness are made.
To further strengthen the impact of the work, future iterations could benefit from a more detailed discussion on how this diagnostic framework can inform the design of next-generation practical systems. For instance, the authors might outline concrete sensing or annotation strategies that could realistically recover per-frame scale information, or explore segmentation-robust architectures that reduce the heavy reliance on precise masks. Additionally, while the 5 mm binary threshold is clinically well-motivated, a brief discussion on how these findings might generalize to multi-class or continuous sizing scenarios would help readers better contextualize the broader applicability of the proposed audit.
Overall, the paper is exceptionally well-written, methodologically sound, and exemplary in its transparency regarding current limitations. It provides a crucial reality check for the community and offers a clear, principled roadmap for moving toward more reliable and clinically grounded monocular polyp sizing.
- 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?
The primary factor behind my recommendation is the paper’s exceptional diagnostic contribution to a critical clinical problem. Rather than proposing yet another incremental model, the authors provide a principled and rigorous analysis of why performance in monocular polyp sizing has plateaued across diverse architectures and input modalities. The experimental design is particularly compelling; by utilizing multi-center datasets, patient-stratified cross-validation, and carefully controlled oracle interventions, the authors successfully disentangle metric scale ambiguity, segmentation robustness, and shortcut learning in a highly convincing and reproducible manner.
A major strength influencing my score is the clarity with which the paper exposes shortcut reliance induced by examination behavior, specifically through the ingenious “shortcut-consistent” versus “inconsistent” frame analysis. The Oracle Scale Ladder experiments provide profound quantitative insights into the theoretical performance headroom, clearly identifying the specific points of failure for current state-of-the-art methods. These findings transcend the task of polyp sizing, offering broader implications for other medical imaging domains plagued by scale ambiguity and procedural bias.
At the same time, the paper’s contributions are primarily analytical rather than algorithmic, as it does not introduce fundamentally new modeling techniques or learning paradigms. While this “Audit-first” approach is highly valuable and appropriate for the study’s goals, it places the work slightly below the threshold for a “Strong Accept (6),” which typically requires a novel methodological framework with immediate, tangible clinical impact. Furthermore, the reliance on subjective visual estimates for ground-truth labels and the focus on binary classification somewhat limit the direct generalizability of the findings to more granular clinical scenarios.
Overall, the manuscript is exceptionally well-written and methodologically rigorous. By setting a high standard for how monocular polyp sizing should be audited and evaluated, it makes a significant and necessary contribution to the field. These strengths firmly justify acceptance independent of the rebuttal process.
- 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.
I agree.
Review #3
- Please describe the contribution of the paper
Insightful Experiments: The paper conducts an interesting set of experiments using an Oracle scale probe. By injecting ground truth metric scales at the frame level, the authors effectively quantify the theoretical performance ceiling, providing a clear picture of how much perfect scale information could actually improve sizing accuracy.
Identifying the Source of Noise in Monocular Sizing: The study goes beyond simple performance metrics to explain why monocular polyp sizing is inherently noisy. It pinpointed that current models often rely on shortcuts rather than true geometry, and identified bounding box instability as a critical bottleneck that undermines even perfect scale data.
- 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.
Originality of the Oracle Scale Probe: The paper introduces an interesting experimental setup using the Oracle scale probe. By injecting ground-truth metric scales at the frame level, the authors successfully quantify the theoretical performance ceiling. This is a clever way to isolate how much of the sizing error stems from missing scale information versus other factors.
Clear Identification of Noise and Bottlenecks: The study provides a solid explanation for why monocular polyp sizing is inherently noisy. It effectively demonstrates that current models rely on shortcuts rather than true geometric scales. Most importantly, it identifies region instability as a critical, independent bottleneck that can undermine even perfect scale data.
- 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 Actionable Solutions: It stops short of proposing concrete methodological improvements or a clear roadmap for future work. A diagnostic study in a venue like MICCAI would be significantly more impactful if it suggested specific ways to overcome the identified bottlenecks, such as scale-invariant feature learning or uncertainty-aware mask refinement strategies.
Insufficient Analysis of High-Risk Lesions (≥10mm): The primary clinical motivation for polyp sizing is identifying high-risk lesions (≥10mm) that require immediate intervention. However, this study focuses almost exclusively on the 5mm threshold and excludes larger polyps due to data imbalance. This focus limits the clinical relevance of the findings, as the management of large polyps is a critical requirement for any real-world sizing pipeline.
Limited Technical Novelty of Oracle Probes: The Oracle scale probe relies on back-calculating scaling factors using ground-truth sizes and pixel dimensions. While this is an informative auditing tool, it lacks technical novelty as it is a straightforward application of the pinhole camera model. Furthermore, because it requires prior knowledge of the target size, it does not offer a generalizable or deployable solution for actual monocular colonoscopy.
Absence of Robustness Analysis for BBox/Mask: As highlighted in Figure 2, the system is highly sensitive to mask errors. The paper would be much stronger if it explored how to achieve mask robustness or provided a more granular statistical analysis (e.g., sensitivity analysis) regarding how different types of segmentation errors—such as over-segmentation versus under-segmentation—specifically impact the final size estimation.
Weak Statistical Significance in Scale Analysis: In Figure 1 (“Does scale help?”), the analysis remains largely qualitative. A more rigorous statistical interpretation, including p-values for performance gains or correlation coefficients between depth estimation errors and sizing errors, is necessary to fully justify the authors’ claims regarding the utility of scale information.
- 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 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?
The paper remains largely descriptive. I would have preferred to see at least one concrete methodological proposal or a pilot solution aimed at addressing one of the identified bottlenecks such as a technique to improve mask robustness under distribution shifts or a way to incorporate metric cues more effectively.
- 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
Author Feedback
We thank reviewers for the detailed feedback.
Actionable steps (R1,R2,R3,Meta): Our work is the first controlled study to show why improvements in monocular depth and segmentation do not automatically yield reliable polyp sizing. Existing model designs and oracle assumptions are deliberately used to identify the following actionable directions for future research. 1) Metric Scale: Relative depth, global scale, and colonoscopy-trained metric depth (MetricCol) fall short of the oracle frame ceiling. Future work needs robust endoscopy-specific metric scale estimation or frame-level calibration anchors. 2) Segmentation Robustness: Despite high in-domain Dice/IoU, SoTA segmentation degrades under distribution shift, hurting sizing. Models need evaluation on downstream sizing robustness; future work should explore uncertainty-aware mask refinement (R3). 3) Data Limitations: Current datasets rely on visual estimates and lack ≥10 mm polyps. We need large tool-verified datasets or realistic simulations supporting scale recovery, depth, and segmentation despite tool occlusions/glare.
Code/data/tools will be public to benchmark future models under our protocol.
Scale-factor (R1,Meta): Oracle factors are deterministic diagnostic variables, not expert tuning or deployable estimators. They back-calculate the missing distance scale from the dataset-provided size label. Using the pinhole model (A = fS/Z ⇒ Z/f = S_p/A_i), and given the arbitrary scale of PPSNet’s estimated depth (d_bar_i), we compute a frame-specific correction factor c_i = (S_p/A_i) / d_bar_i. Oracle-frame applies c_i per frame, Oracle-polyp averages c_i across a polyp’s frames, and Global averages c_i across the training fold.
Label validity & threshold (R2,R3,Meta): While ≥10 mm polyps and finer strata are clinically vital, >85% of public-dataset polyps are <10 mm, making large-polyp evaluation underpowered. The 5 mm task is a balanced starting point to benchmark data-driven models, sharing the same scale ambiguity and robustness challenges. We used checked datasets to reduce label noise: SUN labels were checked by expert endoscopists (>5,000 procedures) and an expert committee; REAL-Colon sizes were eCRF-extracted and specialist-validated. We acknowledge residual visual-estimate noise and advocate for tool-verified datasets.
Mask robustness (R1,R3,Meta): Table 1 reports ±20% bbox perturbations (centered scaling clipped to image bounds) and PolypPVT mask substitution to test mask robustness. For CNN3+Oracle, 10% shrinkage/expansion changed Macro-F1 by -2.5/-3.1 pp, respectively, while 20% shrinkage and expansion changed Macro-F1 by -3.4pp and -2.7pp. The corresponding >5 mm recall changes were -3.4pp and -3.1pp at 10%, and -4.5pp and -2.9pp at 20%. Thus, both under- and over-estimated polyp mask detection reduces downstream sizing performance, with 20% under-segmentation causing the largest recall drop.
Statistical support for scale analysis (R3, Meta): To quantify Fig. 1, we summarized the no-scale vs. Oracle-frame comparison at the fold level. Oracle-frame gives a small Macro-F1 gain in shortcut-consistent frames (+3.6±2.0 pp), but a much larger gain in shortcut-inconsistent frames (+45.9±15.9 pp). A paired t-test over the 5 fold-level gain differences gives p=0.0059, supporting that the gains differ between groups.
Temporal dynamics (R2,Meta): Temporal cues stabilize masks and reduce prediction flicker, but monocular video alone cannot derive metric scale without absolute camera-polyp distance or a physical reference. We isolated this frame-level bottleneck intentionally before sequence modeling.
Further clarifications (R1,Meta): Sec.3.1 reports 147 polyps <=5 mm and 85 polyps >5 mm, trained with class-weighted cross-entropy. Our 2-layer MLP is an interpretable diagnostic baseline using basic geometric, photometric, and texture features. The no-geometry ablation removes apparent-size features (bbox dimensions, area, compactness) to test shortcut reliance.
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 reviewers commonly note that the paper is more of a diagnostic analysis than a technically novel or deployable method, relying on existing model designs and oracle-scale assumptions that are not available in real clinical workflows, while offering limited actionable solutions or a clear path toward real-time deployment. They also raise concerns about reproducibility and clinical validity, including unclear scale-factor estimation and bounding-box perturbation details, no code availability, subjective size labels, limited statistical and robustness analyses, underuse of temporal video information, and an overly narrow focus on the 5 mm binary threshold rather than more clinically important size strata such as ≥10 mm lesions.
- 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.
This paper clarifies the relationship between the estimation results of deep learning-based depth estimation and segmentation models and size estimation of colonic polyps. It provides important information for the clinical application of many studies related to endoscopic image analysis. The authors’ rebuttal has addressed many of the reviewers’ concerns, and I think this paper is acceptable.
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.
Clarifications in the rebuttal regarding the scaling factor and statistical analysis, among others, were acknowledged by the reviewers. Although the paper does not propose a new method, its contribution seems clear. The authors should include, as promised, the code and datasets, as well as the statistical tests.
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.
Initially, one reviewer voted to reject, whereas the two other reviewers proposed an accept (2-5-4). The concerns of the reviewers were adequately addressed, leading to two post-rebuttal accepts and the initial weak accept of reviewer 3.
Reviewers mentioned that the work can serve as an excellent starting point for a community-wide discussion on moving away from performance-driven benchmarks to rigorous diagnostic audit, which is seen as a very valuable contribution.