Detection is solved, delineation is not: what governs tooth segmentation on panoramic radiographs
Muhammad Rehan, Moaz Amjad, Syed Danial Ahmed, Mariam Adnan, Haider Ali
cs.CV · cs.LG
Quick verdict
Claims resolution, not architecture, governs OPG tooth segmentation; rigorous ablations support it, but the private corpus blocks independent verification.
Summary
This empirical study asks which factors govern automatic tooth segmentation and FDI numbering on panoramic radiographs. The authors assemble a private corpus of 1,422 radiographs with 42,142 expert-drawn tooth polygons (32 FDI classes), annotated by 30 dentists with two independent reviewers, then hold corpus, evaluation code and post-processing fixed while varying one factor at a time, scoring every difference with paired bootstrap intervals (2,000 resamples, Holm-corrected). Three findings emerge. Resolution dominates: raising input size from 640 to 1280 px lifts mask mAP50-95 from 0.656 to 0.717 while mAP50 stays flat near 0.982. Architecture is nearly irrelevant in-domain: a 47.4M-parameter Mask2Former statistically ties a 22.4M YOLO11m-seg (95% CI +/-0.0064), is only marginally better under domain shift (+0.0083), yet is 5.5x slower on CPU and not executable under ONNX Runtime. Three interventions fail: a LoRA-adapted DINOv3 encoder trails, SAM2 refinement degrades masks by 39% by truncating roots, and optimal Hungarian label assignment gains just +0.0007 despite fixing duplicates in 40% of out-of-domain images. Zero-shot transfer to DENTEX costs 62% of mAP50-95 but only 18% of mAP50, and error concentrates in the root apex. Main limitations: imagery and the annotation layer are not released, each image has a single annotator, and domain-shift results partly confound annotation-protocol differences.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
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
- Raising input resolution from 640 to 1280 px raises mask mAP50-95 from 0.656 to 0.717 while mAP50 stays flat near 0.982, so added resolution buys boundary precision, not detection.
- Mask2Former (47.4M parameters) is statistically equivalent in-domain to YOLO11m-seg (22.4M) within +/-0.0064 mAP50-95, only marginally better under domain shift (+0.0083), 5.5x slower on CPU and not executable under ONNX Runtime.
- Three targeted interventions fail: a LoRA-adapted DINOv3-B encoder trails (0.6470), SAM2 refinement degrades masks by 39% by truncating roots, and optimal Hungarian label assignment yields only +0.0007 despite correcting duplicates violated in 39.9% of zero-shot predictions.
- Zero-shot transfer to DENTEX costs 62% of mask mAP50-95 but only 18% of mAP50, and residual error concentrates in the apical third of the tooth (IoU 0.776 vs ~0.905 elsewhere).
- Perceptual hashing is unreliable as a contamination check on panoramic radiographs because all such images share the same global arch structure; exact hashing plus normalized pixel correlation is required.
Strengths
Statistical practice far above field norm
Paired bootstrap over images (2,000 resamples) preserving image-level correlation, Holm correction within metric families, exact COCO-mAP reconstruction on resamples verified to 5e-5 (Section 3.7), and equivalence framed via bounded intervals rather than failed significance tests (Table 2: +/-0.0064).
Clean one-factor-at-a-time experimental design
Corpus, evaluation code and post-processing held fixed across resolution (Table 1: 0.656/0.710/0.717 at 640/1024/1280 px with mAP50 flat ~0.982) and architecture comparisons; rectangular 1280x672 canvas, horizontal flip correctly disabled to protect FDI labels (Section 3.5).
Negative results come with identified mechanisms
Contamination audit is a genuine methodological contribution
Unusually candid limitations and provenance disclosure
Weaknesses
The stated contribution is unverifiable: corpus and annotation layer withheld
Apical-third finding may partly be a label-ambiguity ceiling
Domain-shift numbers confound model failure with annotation-protocol differences
Architecture table lacks run-to-run variance; in-domain split is small
Material conflict of interest with unaudited annotation quality
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper, written by five employees of AI Dentify (Pakistan), asks which factors actually govern tooth instance segmentation with FDI numbering on panoramic radiographs. Its stated contribution is an annotation layer — 42,142 expert-drawn polygons over 1,422 radiographs covering all 32 FDI classes — plus a one-factor-at-a-time ablation protocol with paired bootstrap inference. Central claims: (1) input resolution governs boundary precision but not detection (Table 1: mask mAP50-95 0.656 to 0.710 to 0.717 at 640/1024/1280 px, mAP50 flat near 0.982, both gains significant at p < 0.001 and p = 0.024 Holm-corrected); (2) architecture is nearly irrelevant in-domain (Table 2: Mask2Former 0.7163 vs YOLO11m-seg 0.7166, equivalent within +/-0.0064) but marginally helps under domain shift (+0.0083) at 5.5x CPU latency (4,105 ms vs 749 ms) and no ONNX deployability; (3) three interventions fail — LoRA-adapted DINOv3-B trails at 0.6470, SAM2 refinement cuts mAP50-95 by 39% (0.7166 to 0.4391), Hungarian label assignment gains only +0.0007 despite fixing duplicates violated in 39.9% of zero-shot predictions; (4) zero-shot transfer to 634 DENTEX images costs 62% of mAP50-95 (0.7166 to 0.2740) but only 18% of mAP50 (0.9815 to 0.8089), with error concentrated in the apical third (Table 3: IoU 0.776 vs 0.905/0.906 crown/middle).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 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 the main finding of this paper?
On panoramic dental X-rays, detecting and correctly numbering teeth is essentially solved (mAP50 about 0.982 in-domain), while performance is limited by boundary precision. Raising input resolution from 640 to 1280 pixels raises mask mAP50-95 from 0.656 to 0.717 with detection unchanged, and a 47.4M-parameter transformer merely ties a 22.4M-parameter YOLO detector.
Is the tooth segmentation dataset from this paper publicly available?
No. The 1,422-radiograph corpus with 42,142 tooth polygons is not released: the imagery comes from an undocumented public redistribution that cannot be re-shared, and the annotation layer is withheld. Training and evaluation code is on GitHub, and the DENTEX zero-shot evaluation can be reproduced from public data.
Why did SAM2 refinement make tooth segmentation worse?
SAM2 learned edges from natural images where objects have clear intensity boundaries. In panoramic radiographs the tooth root fades into trabecular bone with no visible edge, so SAM2 snapped to the high-contrast enamel and dropped the root entirely, keeping only 58-67% of the mask area and cutting mask mAP50-95 by 39%.
Where does the remaining segmentation error concentrate?
In the apical (root-tip) third of the tooth: IoU there is 0.776 versus about 0.905 for the crown and middle thirds, worst for thin-rooted lower incisors and multi-rooted molars. Part of this gap may be an annotation-ambiguity ceiling, since the apex has no clear radiographic boundary.
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.
- Figures 1-6 are only available as captions in the extracted text; visual content (e.g., SAM2 truncation patterns, apical disagreement maps) could not be independently inspected.
- The abstract in the extraction is truncated mid-sentence ('verif...'); the zero-shot cohort description was reconstructed from Section 4.3 of the body.
- Full training hyperparameters (epochs, batch size, tuning budget per architecture) are not specified in the extracted text, limiting assessment of tuning fairness.
- Only the highlights, main text and Appendix Table 5 were available; no supplementary material was reviewed.
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.