woma: a real-time foundation model and its fine-tuned models for endoscopy
cs.SE · cs.CV · cs.LG
Quick verdict
Pre-registered endoscopy foundation model, honest misses, beats TensorRT at ~100 fps; strong engineering, but shipped weights chosen on the test record.
Summary
The paper presents woma, a 15.6-million-parameter self-supervised foundation model for gastrointestinal endoscopy, selected from eight candidate architectures under pre-registered pass marks and trained with masked-autoencoder objectives on roughly one million unlabelled frames. Two fine-tuned multi-task models are built on it: a colonoscopy model (polyp detection and outlining, colon-segment naming, polyp type, bowel-preparation grading) and a gastroscopy model (22-station naming, lesion flagging and outlining, seven findings). On held-out public data, the colon model finds 96% of PolypGen polyps at precision >=0.85 and 19/19 REAL-Colon polyps at 1.6 false alarms per procedure; the gastroscopy model names landmark regions on 92.3% of unseen-patient frames and flags 37/39 neoplasia frames at 0.912 specificity. Three pass marks are missed and reported: colon-segment recall (0.521 vs 0.70), polyp type (0.617-0.635 vs 0.80), and other-hospital outline Dice (0.807 vs 0.82). Every task exceeds ~100 fps at 1080p on one workstation GPU, where the authors' numbat runtime beats PyTorch, ONNX Runtime and TensorRT in all four tested precision regimes; a Vulkan build runs without vendor libraries. Key limitations: shipped weights were selected on the same held-out record used for reporting, decision sets are small (19 polyps, 39 neoplasia frames), there is no clinical validation, and code and weights are available only on request.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 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 pre-registered two-stage experiment over eight candidates selects ConvNeXt-V2-Nano, continued with masked autoencoding on ~1M endoscopy frames, as the best production trunk.
- The colonoscopy model meets both pre-registered detection marks on unseen-centre data: 96% of PolypGen polyps at precision >=0.85 and 19/19 REAL-Colon polyps at 1.6 false alarms per procedure.
- The gastroscopy model meets all five gated marks in one set of weights (region 0.923, gate 0.949 at specificity 0.912, findings 0.877, UGIAD 0.932, GastroVision 0.861).
- The numbat runtime is faster than PyTorch, ONNX Runtime and TensorRT on the same GPU in all tested precision regimes, and a Vulkan build needs no vendor libraries.
- Twelve-hour frozen-probe screening ordered five trunks exactly as two weeks of full fine-tuning did, with no inversions.
Strengths
Pre-registered, falsifiable evaluation design
Pass marks derived from published anchors and fixed before any run (Table 1); a two-gate project plan with written kill/fallback branches (Fig 3); screening used equal-clock runs, four comparison views and a pre-registered rank ladder (Sec 2.3).
Frozen probes predicted the final fine-tuned ranking exactly
Stage-1 station-probe order 3>2>5>7>1 (0.772/0.761/0.735/0.713/0.698, Table 3) matches post-fine-tune landmark-region accuracy 0.923/0.738/0.662/0.634/0.575 with no inversion (Sec 3.2) - compelling evidence that cheap screening suffices.
Unusually honest reporting of misses and operating points
Confound investigation beyond field norm
Rigorous, reproducible-in-structure speed protocol
Weaknesses
Shipped weights selected on the reporting data
Headline claims rest on tiny positives
Closed weights and a self-cited runtime
Single-machine speed evidence
Partly private data mixture
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This paper bundles three artefacts: woma, a 15.6M-parameter ConvNeXt-V2-Nano foundation model self-supervised on ~1M endoscopy frames; two fine-tuned multi-task models (colonoscopy: detection, outline, ten colon segments, polyp type, bowel preparation; gastroscopy: 22+1 stations, lesion gate/outline, seven findings); and numbat, the authors' self-contained training and inference runtime. The claims are unusually falsifiable: pass marks anchored to published results and fixed before any run (Table 1), every outcome reported met or missed, and a speed claim that numbat beats PyTorch, ONNX Runtime and TensorRT in every tested precision regime on one workstation GPU (Table 7). Headline numbers: 0.960 PolypGen sensitivity at precision >=0.85; 19/19 REAL-Colon polyps at 1.6 false alarms per procedure; 0.923 landmark-region accuracy on unseen patients; 37/39 neoplasia frames flagged at specificity 0.912; ~100 fps delivered at 1080p (Table S2).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 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 woma?
woma is a 15.6-million-parameter self-supervised foundation model for gastrointestinal endoscopy, trained without labels on about one million endoscopy frames. From it, the authors fine-tune two task models: a colonoscopy model (polyp detection and outlining, colon-segment naming, polyp type, bowel-preparation grading) and a gastroscopy model (naming 22 protocol stations, flagging and outlining lesions, and classifying seven findings).
How accurate is the woma endoscopy model?
On held-out public data it finds 96% of PolypGen polyps at precision of at least 0.85, and 19 of 19 REAL-Colon polyps at 1.6 false alarms per procedure; the gastroscopy model names the correct anatomical region on 92.3% of frames from unseen patients and flags 37 of 39 neoplasia frames at 0.912 specificity. Three pre-set targets were missed and openly reported: colon-segment recall (0.521 vs 0.70 target), polyp typing, and cross-hospital outline Dice (0.807 vs 0.82 target).
Is woma faster than TensorRT?
On the single workstation GPU tested, the authors' numbat runtime is faster than PyTorch, ONNX Runtime and TensorRT in every precision regime measured, delivering roughly 100+ frames per second end to end at 1080p. A second build uses Vulkan kernels and needs no NVIDIA libraries at all. These results come from one machine and one sitting, so they have not been shown across different hardware.
Can the woma models be used clinically?
No. The authors state the models are research artefacts, not medical devices; several heads are trained on research-use-only licensed data, no endoscopist has used the overlay during a real procedure, and the decision sets behind the headline numbers are small (19 polyps, 39 neoplasia frames).
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.
- All figures are available only as captions; pipeline diagrams and training-curve plots (Figs 1-8, S1-S2) could not be visually verified.
- The metadata abstract is truncated mid-sentence; the full-text abstract is complete and was used instead.
- All speed and accuracy records are author-generated; nothing in the text allows independent re-measurement.
- Code, model files and records are stated available only from the corresponding author, not from a public repository.
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.