Lesion-centered 3D mapping of colonoscopy procedures: validation of a hierarchical ensemble pipeline on public benchmark videos
Hyunjun Kim, Hyeonwoo Na, Jaewoo Lee
cs.CV
Quick verdict
An honest systems-integration study assembling published components into a lesion-centered colonoscopy spatial record; results are internally consistent but directional, with no ground truth behind revisit detection and clinical validation still pending.
Summary
This validation study assembles four published components into a lesion-centered spatial record of colonoscopy procedures, explicitly claiming no new base algorithms: (1) topological mapping with entry-map Bayesian relocalization for revisit detection, (2) SAM 2 lesion tracking, (3) selective local 3D reconstruction with Endo3R over tracked intervals only, and (4) pairwise identity scoring combining EndoFM appearance embeddings with node and temporal features. On 40,245 frames from four public videos (two C3VDv2 phantom sequences with ground-truth depth, two full REAL-Colon procedures), the entry-map protocol detected 5,614 and 4,043 revisit events where mapping alone yields zero by construction; auto-merge at threshold 0.5 merged 20 of 231 track pairs at ground-truth purity 1.0 (same-lesion recall 0.19); and Endo3R beat VGGT on phantom depth (AbsRel 0.2276 vs 0.3523). Limitations: four-video sample, revisit events lack anatomical ground truth, identity recall is capped at 0.19 by the structurally disconnected identity score, and 31 of 33 real-procedure tracks show mask instability. Results are directional, as the authors themselves state; clinical validation remains the open task.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Entry-map Bayesian localization detects thousands of revisit events (5,614 and 4,043 across two REAL-Colon procedures) that are structurally impossible under forward-only topological mapping alone.
- Pairwise lesion-identity scoring auto-merges 20 of 231 candidate track pairs at threshold 0.5 while keeping ground-truth purity at 1.0, though same-lesion recall is capped at 0.19 by the structurally zero node component of the score.
- The endoscopy-specific engine Endo3R outperforms the general-purpose VGGT on phantom depth accuracy (overall AbsRel 0.2276 vs 0.3523) with per-frame median scale alignment.
- A lesion-centered spatial record (revisits, lesion tracks, local 3D, identity proposals) can be assembled from published components and validated end to end on public data, without full-colon 3D reconstruction; clinical validation remains future work.
Strengths
End-to-end reproducibility is unusually thorough
Appendix A: full pipeline code, adapters, configs and test suite on GitHub with Zenodo DOI (10.5281/zenodo.22136766), run snapshots (kci-4v-s03-001 / s04-001), all quantitative figures auto-generated from released result JSONs; datasets are public (C3VDv2 CC BY 4.0, REAL-Colon CC BY-NC-SA 4.0).
Variant disclosure instead of cherry-picking
Section 2.3/3.1 report the laxer gate 0.33 (localized 0.94-0.95) and full-map variants alongside the adopted gate 0.5 (Table 2, Fig. 4), and Table 3 reports the full threshold grid 0.75/0.5/0.45/0.40 including purity collapse at 0.40.
Structural analysis converts negative results into design insight
Leakage discipline and calibrated claim scoping
Weaknesses
Revisit detection has no ground truth
Identity layer performance is weak and threshold selection is in-sample
Track quality on real procedures undermines downstream layers
Local 3D evidence is phantom-only, tiny, and statistically untested
Several headline quantities are definitions, not validated measurements
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- 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 does this colonoscopy 3D mapping paper actually do?
It assembles four already-published AI components — topological mapping with relocalization, lesion tracking with SAM 2, local 3D reconstruction with Endo3R, and lesion identity scoring — into one pipeline that builds a lesion-centered spatial record of a colonoscopy (where the scope traveled, where lesions were seen, and whether the same lesion reappeared), without reconstructing the whole colon in 3D. The contribution is the integration and its evaluation, not new algorithms.
What were the main results?
On four public videos (~40,000 frames), the pipeline detected thousands of revisit events in two full procedures (5,614 and 4,043) where mapping alone detects zero; it auto-merged 20 of 231 candidate lesion-track pairs without ever wrongly merging different lesions (purity 1.0, but only ~19% of true same-lesion pairs found); and the endoscopy-specific depth model (Endo3R) beat the general-purpose VGGT on phantom data (absolute relative error 0.23 vs 0.35).
What are the main limitations?
Validation covers only four videos; revisit detections have no anatomical ground truth and are presented as plausible rather than verified; lesion-identity recall is low (~19%); and in the real procedures 31 of 33 lesion tracks suffered mask instability. All outputs are meant as proposals for physician confirmation, and the authors state no generalization claim is made.
Is the code and data available?
Yes. The pipeline code, adapters, configurations, and test suite are on GitHub (hyunjun1121/endovision-pipeline) with a Zenodo DOI, runs are reproducible from a single configuration file, and the evaluated datasets (C3VDv2 and REAL-Colon) are publicly available benchmarks.
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.
- Text-only extraction from a single-column PDF: figure graphics are not analyzable; all figure-derived statements rely on captions and in-text numbers.
- Fig. 6's axis labels appear interleaved with in-panel text in the extraction (reject/auto-merge threshold lines), so exact distribution shapes were not verified.
- The abstract's Conclusions sentence is truncated mid-word in the provided text; conclusions were reconstructed from Section 5.
- Tables contain no inferential statistics, so 'statsNotes' reflects the absence of tests rather than a failure to extract them.
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.