💻 computer science

Vendor-Aware Three-Dimensional Missing-Tooth Localization in Partially Annotated Multi-Vendor Cone-Beam Computed Tomography: A Leakage-Controlled Evaluation of Semi-Supervision

This study evaluates semi-supervised and vendor-adversarial learning strategies for 3D missing-tooth localization in multi-vendor CBCT scans, finding that while vendor-adversarial training achieved the best locked test performance, semi-supervised approaches did not demonstrate significant superiority over pure supervision.

Original authors: Hana Asgari, Faraneh Zarafshan, Ahmadreza Talaeipoor, Elham Mazinan

Published 2026-09-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hana Asgari, Faraneh Zarafshan, Ahmadreza Talaeipoor, Elham Mazinan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of modern dentistry, three-dimensional imaging has become a standard tool for planning complex treatments like dental implants. These scans, known as cone-beam computed tomography, create a detailed volumetric map of a patient's jaw, showing the position of every tooth, the density of the bone, and the path of important nerves. While computers have become quite good at finding teeth that are present in these scans, identifying where a tooth is missing is a much harder puzzle. A missing tooth is not a physical object the computer can see; it is an empty space defined by the surrounding anatomy. To teach a computer to find these empty spaces, researchers need to show it thousands of examples where the missing spots have been carefully marked by human experts. However, creating these detailed maps is slow and expensive, leaving many scans without the necessary labels. This has led to a common hope in the field of artificial intelligence: that computers could learn from the vast number of unlabeled scans by guessing the answers and then teaching themselves, a process known as semi-supervised learning.

A team of researchers set out to test whether this self-teaching approach actually works for finding missing teeth across different types of scanning machines. They gathered a collection of jaw scans from three different manufacturers, each producing images with slightly different technical characteristics. The dataset included scans where missing teeth had been carefully marked by humans, as well as a large number of scans where no such marks existed. Before training any computer models, the team performed a rigorous audit of their data. They checked for duplicate files, ensured the images were ordered correctly to match the patient's actual anatomy, and removed any inconsistencies that could trick the computer. This careful preparation was crucial because computers are notorious for finding shortcuts; they might learn to recognize the specific brand of the scanner rather than the actual anatomy, leading to results that look good in the lab but fail in the real world.

The researchers then built several different computer systems to solve the problem. One system relied only on the scans with human labels. Another system tried to use the unlabeled scans to teach itself, using a method where a "teacher" model makes guesses and a "student" model tries to learn from them, but only if the teacher is very confident in its answer. A third system attempted to force the computer to ignore the brand of the scanner entirely, focusing only on the shape of the jaw. They tested these systems repeatedly, changing the starting conditions slightly each time to see if the results were stable or just a lucky fluke. They also tested how well the systems performed when faced with a brand of scanner they had never seen before, a scenario that mimics a dentist using a new machine in their clinic.

The results were surprisingly clear and challenged the prevailing optimism about self-teaching computers. The system that tried to learn from the unlabeled scans did not perform better than the one that used only the labeled data. In fact, the self-teaching approach sometimes made the computer worse, a phenomenon the researchers call negative transfer, where the guesses from the unlabeled data confused the model rather than helping it. The most successful approach was the one that specifically taught the computer to recognize and ignore the unique signatures of the different scanner brands. This system achieved the best balance of finding the missing teeth while avoiding false alarms, correctly identifying missing sites in about 79 percent of cases while producing fewer than half a false alarm per scan.

When the researchers tested the systems on scanners they had never seen during training, the results were mixed. The system designed to ignore scanner brands worked well for two of the three new brands but struggled with the third. This suggests that while teaching a computer to be aware of the equipment it uses is helpful, it does not guarantee it will work perfectly on every machine it encounters. The study concludes that simply adding more unlabeled data does not automatically improve performance in this specific medical task. Instead, the most reliable path forward involves carefully controlling for the differences between machines and being cautious about assuming that computers can learn effectively from data they have not been explicitly taught to understand. The work serves as a careful, auditable benchmark that prioritizes data integrity and honest evaluation over the promise of quick technological fixes.

Technical Summary: Vendor-Aware Three-Dimensional Missing-Tooth Localization in Partially Annotated Multi-Vendor CBCT

Problem Definition
The study addresses the challenge of localizing missing teeth (edentulous sites) in three-dimensional Cone-Beam Computed Tomography (CBCT) volumes. This task is complicated by two primary factors:

Sparse and Partial Annotations: Public datasets often provide 3D bounding boxes for only a subset of missing sites. Crucially, scans without released boxes cannot be treated as negative examples (i.e., scans with no missing teeth); they simply lack localization labels. Standard semi-supervised approaches risk amplifying teacher errors or misinterpreting unlabeled data as negative.
Vendor-Dependent Geometry: CBCT scanners (specifically Kavo, Meyer, and Newtom in this study) differ significantly in matrix size, voxel spacing, field of view, and reconstruction artifacts. Deep learning models often exploit these vendor-specific signatures ("shortcuts") to achieve high internal performance that fails to generalize across different acquisition systems.

The core research question is whether unlabeled CBCT volumes improve 3D missing-tooth localization when data integrity, scanner geometry, and vendor shortcuts are rigorously controlled, or if semi-supervised learning (SSL) introduces negative transfer.

Methodology
The authors propose a Vendor-Aware Reliability-Gated Semi-Supervised 3D (VR-SSL3D) framework, evaluated under a strict "leakage-controlled" protocol.

Data Integrity and Eligibility: Before model training, a comprehensive audit was performed on the MORPH-CBCT dataset. This included pixel-level fingerprinting to identify exact duplicates (M38 and M61 were excluded due to discordant annotations), verification of physical slice ordering, and validation of bounding box geometry. The final eligible cohort consisted of 83 labeled scans (111 objects) and 73 unlabeled scans.
Geometry-Preserving Reconstruction: All volumes were reconstructed from DICOM physical geometry and resampled to 0.5-mm isotropic spacing. Bounding boxes were transformed analytically to maintain physical coordinates (millimeters) rather than voxel indices, ensuring consistency across different scanner resolutions.
Model Architecture: The system utilizes an anchor-free 3D detector with a Feature Pyramid Network (FPN).
- Teacher-Student Mechanism: An Exponential Moving Average (EMA) teacher generates pseudo-labels for unlabeled data.
- Reliability Gating: Pseudo-boxes are accepted only if they pass three gates: (1) Confidence (high objectness/class probability), (2) Uncertainty (low dispersion in center, extent, and class across stochastic views), and (3) Geometric Stability (consistency under paired augmentations). Rejected candidates are ignored, not treated as negatives.
- Vendor-Adversarial Learning: A gradient-reversal layer (GRL) branch is attached to the encoder to discourage the model from learning vendor-specific features, promoting domain invariance.
Evaluation Protocol: The study employs a strict 5-fold cross-validation where calibration and outer-test partitions are inaccessible during model development. Thresholds are selected on a disjoint calibration set to maximize sensitivity subject to a constraint of $\le$ 1 false positive (FP) per volume. Performance is measured using partial Free-Response Receiver Operating Characteristic (pFROC) scores, sensitivity, and median center error.

Key Contributions

Auditable Benchmark: A reproducible integrity lock that prevents data leakage (e.g., duplicate volumes, slice-wise splitting) and ensures case-level independence.
Falsification of SSL Superiority: A controlled evaluation demonstrating that semi-supervised consistency, even with reliability gating, does not necessarily improve performance over strong pure supervision in this specific domain.
Vendor-Controlled Analysis: Evidence that acquisition-aware vendor control (adversarial learning) yields more favorable operating points than semi-supervision in multi-vendor settings.
Gate Ablation: A cumulative analysis linking pseudo-label precision and coverage to downstream localization, showing that full gating improves pseudo-label quality (precision) at the cost of coverage without statistically significant gains in final localization metrics.
Heterogeneous Generalization: A strict Leave-One-Vendor-Out (LOVO) analysis revealing that generalization benefits are vendor-dependent, not universal.

Results

Developmental Validation: Across 75 repeated-seed runs, Vendor-Adversarial Training achieved the highest mean pFROC (0.899), slightly outperforming Pure Supervision (0.890). Reliability-Gated VR-SSL3D (0.872) and Dense Ungated SSL (0.874) performed worse than pure supervision, indicating negative transfer.
Locked Outer-Test Performance: At the calibrated operating point (max sensitivity with $\le$ $\leq$ 1 FP/volume):
- Vendor-Adversarial: Sensitivity 0.793, 0.458 FP/volume.
- Pure Supervision: Sensitivity 0.757, 0.602 FP/volume.
- Reliability-Gated VR-SSL3D: Sensitivity 0.766, 0.518 FP/volume.
- Dense SSL: Sensitivity 0.748, 0.530 FP/volume.
- Conclusion: Vendor-adversarial training provided the most favorable trade-off. Semi-supervised variants did not demonstrate superiority.
Post-Lock Benchmarks: A conventional 3D detector (trained only on labeled data) achieved a higher descriptive validation mean (0.921) than a confidence-only Mean-Teacher model (0.886), though the primary fold-blocked test was not statistically significant ( $p=0.25$ ).
LOVO Analysis: When a vendor was held out entirely:
- Meyer & Newtom: Vendor-adversarial training showed the highest pFROC.
- Kavo: Vendor-adversarial training performed worse than pure supervision (0.372 vs. 0.395), and reliability gating reduced false positives but also reduced sensitivity.
- Conclusion: Generalization is heterogeneous; no single method guarantees robustness across all unseen scanners.

Significance and Claims
The paper positions itself as an auditable benchmark and a falsification study rather than a demonstration of a superior SSL mechanism. Its primary significance lies in:

Cautious Interpretation of Unlabeled Data: It provides evidence that unlabeled data does not automatically improve 3D object detection in medical imaging, particularly when annotation status is ambiguous (missing vs. negative).
Importance of Acquisition Control: It demonstrates that explicit vendor control (adversarial learning) is more effective than semi-supervision for achieving robust operating points in multi-vendor CBCT datasets.
Methodological Rigor: It highlights the necessity of strict data integrity audits (removing duplicates, verifying geometry) and leakage-controlled evaluation to avoid inflated performance claims.

The authors explicitly state that the work does not establish scanner-independent or clinical superiority for the proposed SSL mechanism. Instead, it cautions against assuming that unlabeled localization targets improve performance and emphasizes that robustness must be tested via vendor-stratified evaluation rather than assumed.

Technical Summary: Vendor-Aware Three-Dimensional Missing-Tooth Localization in Partially Annotated Multi-Vendor CBCT

More like this