Leakage-controlled evaluation of vendor-aware three-dimensional missing-tooth localization in partially annotated multi-vendor cone-beam computed tomography

Abstract Three-dimensional missing-tooth localization in cone-beam computed tomography (CBCT) is challenged by partial annotation and scanner-dependent acquisition differences, while the benefit of unlabeled scans remains uncertain. We therefore conducted a leakage-controlled evaluation of physical-space reconstruction, vendor-adversarial learning, dense teacher–student consistency, and reliability-gated semi-supervision under common case-level partitions. After excluding two discordantly annotated exact duplicates, 83 labeled and 73 unlabeled scans with 111 objects were eligible. Across 75 repeated-seed runs, mean pFROC was 0.890 for pure supervision, 0.899 for vendor-adversarial training, and 0.872 for gated semi-supervision. In 50 post-lock exploratory benchmark runs, conventional 3D detection achieved 0.921 versus 0.886 for confidence-only Mean-Teacher-3D; the primary fold-blocked test was not significant ( $$p=0.25$$ ). A post-lock exploratory cumulative gate ablation showed no pFROC change from uncertainty alone; full gating increased mean pseudo-label precision at confidence 0.70 to 0.826 while reducing coverage to 0.680, and modestly increased pFROC by 0.007 ( $$p=0.50$$ ). After fold-specific calibration, vendor-adversarial training achieved the strongest locked outer-test operating point: sensitivity 0.793 (95% CI 0.697–0.880) at 0.458 false positives per volume (0.301–0.639). Post-lock leave-one-vendor-out results were heterogeneous: vendor-adversarial training was numerically strongest for Meyer and Newtom but not Kavo. Semi-supervised superiority was not demonstrated. The study is therefore positioned as an auditable evaluation of acquisition control and unlabeled-data use in multi-vendor CBCT rather than as evidence for a fundamentally novel or universally superior semi-supervised algorithm. These conclusions are specific to the investigated cohort and acquisition profiles and do not establish scanner-independent generalization.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-16
DOI
https://doi.org/10.1038/s41598-026-79069-x
Primary Topic
Dental Radiography and Imaging
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Leakage-controlled evaluation of vendor-aware three-dimensional missing-tooth localization in partially annotated multi-vendor cone-beam computed tomography

Faraneh Zarafshan, ahmadreza talaeipor, Elham Mazinan, Hana Asgari
Scientific Reports
Dental Radiography and Imaging
article

Leakage-controlled evaluation of vendor-aware three-dimensional missing-tooth localization in partially annotated multi-vendor cone-beam computed tomography

Faraneh Zarafshan, ahmadreza talaeipor, Elham Mazinan, Hana Asgari
article en

Abstract

Abstract Three-dimensional missing-tooth localization in cone-beam computed tomography (CBCT) is challenged by partial annotation and scanner-dependent acquisition differences, while the benefit of unlabeled scans remains uncertain. We therefore conducted a leakage-controlled evaluation of physical-space reconstruction, vendor-adversarial learning, dense teacher–student consistency, and reliability-gated semi-supervision under common case-level partitions. After excluding two discordantly annotated exact duplicates, 83 labeled and 73 unlabeled scans with 111 objects were eligible. Across 75 repeated-seed runs, mean pFROC was 0.890 for pure supervision, 0.899 for vendor-adversarial training, and 0.872 for gated semi-supervision. In 50 post-lock exploratory benchmark runs, conventional 3D detection achieved 0.921 versus 0.886 for confidence-only Mean-Teacher-3D; the primary fold-blocked test was not significant ( $$p=0.25$$ ). A post-lock exploratory cumulative gate ablation showed no pFROC change from uncertainty alone; full gating increased mean pseudo-label precision at confidence 0.70 to 0.826 while reducing coverage to 0.680, and modestly increased pFROC by 0.007 ( $$p=0.50$$ ). After fold-specific calibration, vendor-adversarial training achieved the strongest locked outer-test operating point: sensitivity 0.793 (95% CI 0.697–0.880) at 0.458 false positives per volume (0.301–0.639). Post-lock leave-one-vendor-out results were heterogeneous: vendor-adversarial training was numerically strongest for Meyer and Newtom but not Kavo. Semi-supervised superiority was not demonstrated. The study is therefore positioned as an auditable evaluation of acquisition control and unlabeled-data use in multi-vendor CBCT rather than as evidence for a fundamentally novel or universally superior semi-supervised algorithm. These conclusions are specific to the investigated cohort and acquisition profiles and do not establish scanner-independent generalization.

Scientific Reports
Islamic Azad University, Karaj (IR), Islamic Azad University, Arak (IR), Tehran University of Medical Sciences (IR)
Good health and well-being
Dental Radiography and Imaging
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon