WSPolypNet: Weakly Supervised Polyp Localization in Colonoscopy Videos
Giseong Hwang, Minjae Jo, Yeonghyeon Park, Kyeonghun Kim, Seoyeon Han, Donghoon Han, Haneul Kim, Yului Jeong, et al.
cs.CV · cs.AI
Quick verdict
Proposes video-label-only polyp localization via 3D-CNN CAM + MedSAM2; solid small-polyp gains but no external baselines and no test-set description.
Summary
WSPolypNet is a weakly supervised framework that localizes polyps in colonoscopy videos using only video-level labels, avoiding costly frame-level annotation. A 3D CNN (X3D, Kinetics-400 pretrained) is trained with binary clip labels to produce class activation maps (CAMs); CAM evidence from five views (a full 224×224 view plus four 144×144 corner crops resized back to 224) is fused by pixel-wise maximum, and the peak point of the seed frame is supplied to MedSAM2 as a point prompt, which propagates masks bidirectionally with a confidence-thresholded frame-wise fallback. Trained on 871 positive clips from LDPolypVideo and 615 negative clips from 60 polyp-free videos, the full framework reaches CorLoc of 47.80%, 43.68%, and 35.01% at IoU thresholds 0.3/0.5/0.7, versus 36.87%, 33.72%, and 27.94% for the single-view configuration, and recall of 94.51% versus 74.18%. The multi-view strategy nearly doubles small-polyp CorLoc@0.5 (16.01% to 30.97%) but reduces large-polyp CorLoc@0.5 (71.84% to 65.75%). Main limitations: no external baseline comparison, an undescribed test set, single-run results without variance, and localization of only one polyp per video.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A 3D CNN trained only with video-level binary labels yields CAMs that, after five-view fusion and MedSAM2 refinement, localize polyps at CorLoc@0.5 of 43.68% versus 33.72% for the single-view configuration.
- The multi-view corner-crop strategy nearly doubles small-polyp CorLoc@0.5 (16.01% to 30.97%) at the cost of large-polyp CorLoc@0.5 (71.84% to 65.75%).
- The framework achieves 94.51% recall in identifying positive clips, up from 74.18% in the single-view setting.
- The pipeline requires no frame-level bounding-box or pixel-mask annotation of the training videos.
Strengths
Honest negative results reported in full
Table III openly shows the multi-view strategy degrades large-polyp CorLoc@0.5 from 71.84% to 65.75% while helping small polyps (16.01% to 30.97%), and the conclusion states the single-polyp limitation — the size-agnostic headline is not the whole story and the authors say so.
Explicit defense against source-specific shortcut learning
Positives come from LDPolypVideo and negatives from 60 separate polyp-free videos with different resolution and borders (Sec. IV-A); the FOV-based ROI cropping plus training-time spatial randomization of scale (90–100%) and position (Sec. III-B) directly targets this confound.
Unusually complete hyperparameter disclosure
Meaningful small-polyp gain on a known CAM failure mode
Weaknesses
No external baseline comparison
The evaluation set is never described
Lopsided contribution attribution
Single-run results with no uncertainty
Tuned-looking thresholds with no tuning protocol
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
WSPolypNet targets weakly supervised polyp localization in colonoscopy video, motivated by the 20–26% polyp miss rate of routine colonoscopy (Sec. I) and the cost of dense frame-level annotation. Using only video-level binary labels, the system must decide whether a clip contains a polyp and where it lies in each frame. The pipeline chains three components: an X3D 3D CNN (Kinetics-400 pretrained) trained with BCE clip-level loss that yields class activation maps (Sec. III-C); a five-view CAM strategy fusing the full 224×224 view with four 144×144 corner crops re-scaled to 224 (≈1.56× magnification, 64-px overlap) via pixel-wise maximum (Sec. III-D, Eq. 2–3); and MedSAM2 prompted with up to five CAM-derived points, propagating masks bidirectionally, re-scoring candidate tracks against CAM responses, and falling back to frame-wise prompting when a propagated mask is empty or its confidence drops below τ_conf = 0.755. Headline claims: CorLoc of 47.80%/43.68%/35.01% at IoU 0.3/0.5/0.7 versus 36.87%/33.72%/27.94% single-view (Table II); recall 94.51% versus 74.18%; small-polyp CorLoc@0.5 up from 16.01% to 30.97% (Table III) — all without frame-level spatial annotation of training videos.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is WSPolypNet and how does it work?
WSPolypNet is a weakly supervised method for finding polyps in colonoscopy videos using only video-level labels (does/doesn't contain a polyp), with no frame-level boxes or masks for training. A 3D CNN (X3D) trained for binary clip classification produces class activation maps; CAM evidence from five views (full frame plus four magnifying corner crops) is fused, and the strongest point of the best frame is given to MedSAM2 as a point prompt, which segments the polyp and propagates the mask across the clip.
How accurate is WSPolypNet?
In the paper's experiments it reaches CorLoc (correct localization rate) of 47.80%, 43.68%, and 35.01% at IoU thresholds 0.3, 0.5, and 0.7, versus 36.87%, 33.72%, and 27.94% without the multi-view strategy; recall is 94.51% versus 74.18%. For small polyps, CorLoc@0.5 improves from 16.01% to 30.97%, though large-polyp CorLoc@0.5 drops from 71.84% to 65.75%.
Does WSPolypNet remove the need for annotations entirely?
No. It removes frame-level spatial annotations (boxes/masks) for training — only a video-level 'polyp present/absent' label is needed. Ground-truth bounding boxes are still required to evaluate the method (CorLoc), and the authors position the tool as annotation assistance rather than a full replacement.
What are the main limitations of the paper?
Per the authors, only one polyp per video can be localized, multi-view fusion hurts large polyps, and clips cover at most 30 frames at 1 FPS. Reviewer-level concerns add: no comparison with any prior published weakly supervised method, the test set is never described, results come from single runs without variance, and no code release is mentioned.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure contents (Figs. 1–3) are images not present in the extracted text; only captions were available, so qualitative localization claims could not be assessed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.