Certified Dental Biometric Verification Under Partial Overlap

Abstract

Dental identification is usually evaluated as closed-set retrieval: given a dental scan, rank the enrolled gallery and report whether the correct subject appears first. This framing hides two operational requirements. First, forensic and clinical deployments need an accept/abstain decision with a controlled false-match rate (FMR), not only a Rank-1 score. Second, practical queries are often partial: field-of-view limits, trauma, extractions, and missing teeth break the rigid alignment assumption used by most 3D dental matching pipelines. We present ToothPrint, a dental biometric verification pipeline for 3D intraoral scans and 2D radiographs that combines rigid surface verification, learned partial-overlap correspondence, and split-conformal calibration. On 200 public 3D dental arches, rigid surface verification achieves Rank-1 0.995 and EER 0.005 under full coverage. Under realistic whole-tooth dropout, however, rigid GICP falls to Rank-1 0.23 at 50% tooth retention. A learned point-correspondence verifier lifts this regime to Rank-1 0.867 (AUC 0.984), while a crop-hardened embedding baseline reaches 0.635. The same score family supports a conformal open-set decision: full-coverage FNIR at FMR=1% is 0.00, while heavy tooth loss is correctly exposed as an abstention regime. We then move from synthetic scans to real intraoral arches: on 150 Teeth3DS+ upper arches acquired ungated and md5-verified, full-coverage identity reproduces Rank-1 1.000 (N=40), and an off-the-shelf zero-shot registrar (BUFFER-X) with no dental training recovers partial overlap to Rank-1 1.00±0.00 at 50% tooth retention and 0.95 at 30% over three crop-seed repetitions—evidence that the partial-overlap leg may be better served by a general-purpose zero-shot registrar than by a bespoke correspondence network, without touching the certified pipeline. A negative control is equally informative: a frozen indoor self-supervised encoder (Sonata/PTv3) with an ArcFace head does not transfer to dental identity (Rank-1 0.275 at full coverage). We report DET curves, full per-condition partial-overlap tables, a sensor-perturbation robustness sweep, hard-negative calibration, an untrained-correspondence control, and a cross-dataset transfer test, together with auxiliary results that carry the same split-conformal machinery to 2D radiographs, restoration patterns, and longitudinal bone-level and 3D surface change—each with a finite-sample false-alarm bound, and each on single-session synthetic data. The central limitation remains explicit: the new real-arch results are still single-session, so identity queries are synthetic reacquisitions and the real cross-session gate stays open. The method is therefore evidence for a research direction, not a clinical claim.