toothprint 1.1.1

Certified dental-imaging intelligence: identification, change, and surface certificates from teeth.

identity · change · surface — three reads of one durable signal, each returning a certificate instead of a guess.

---

A face can be lost; the teeth remain. ToothPrint turns dental scans, radiographs, and visit-to-visit records into evidence that can be checked: a match, a change finding, or a surface map, each with an explicit uncertainty certificate.

🔖 Preprint + DOI · 🔁 Reproduce · 🔬 Full write-up · ▶ 30-second reel · 🖥️ Desktop app

Start here

Need	ToothPrint returns	Primary viewer
Is this the same person?	a ranked dental match plus accept / abstain certificate	forensic examiner, dentist, researcher
Did bone level change between visits?	changed / stable / uncertain with a false-progression bound	dentist, periodontist, patient-facing report
Where did the 3D surface change?	regional surface-change map with a false-change bound	clinician, lab, reconstruction researcher
Can I inspect files locally?	safe ingest, local preview, API response, and PDF case report	hospital IT, engineer, reviewer

Every verdict is conformal: it fires only when the interval around the measurement lies entirely past the threshold, so the false-alarm rate is bounded by α in finite samples. The certification core depends only on numpy, scipy, opencv, and open3d; learned front-ends (tooth detection, point-correspondence, Gaussian-splatting reconstruction) are pluggable and optional, so the guarantees run without a GPU.

Current status

What is ready	What is not claimed
Research code, API, Studio app, safe file ingest, report export, tests, smoke reproduction, and committed visual evidence.	Not an autonomous clinical or forensic decision system. Not cleared by any regulator. Not validated on real multi-visit deployment data yet.
Validated on public dental data, synthetic re-scans/crops, fixture-sized reproducibility tests, and selected cross-dataset probes.	Headline numbers are best read as in-simulation ceilings until a clinical/forensic partner supplies real cross-session longitudinal data.
Designed for expert-in-the-loop review: show the evidence, report uncertainty, abstain when outside the validated regime.	No patient-affecting use without site recalibration, prospective validation, governance, and regulatory clearance.

For different readers

Reader	Start with	What matters
Patient	What a result looks like, Model card	local analysis, explicit uncertainty, no autonomous decisions
Dentist / clinician	Change, ToothPrint Studio	visit-to-visit bone-level tracking, PDF report, limitations
Hospital / lab	Current status, Security, Clinical readiness	local files, audit trail, parser hardening, validation gap
Researcher	Results, vs SOTA, Reproduce	conformal certificates, benchmarks, ablations, datasets
Engineer	Quick start, Formats, How it works	install, API endpoints, file loaders, tests

For patients

ToothPrint is designed to make dental evidence easier to review, not to make unattended decisions. A result should show what was compared, what changed or matched, how uncertain the finding is, and when the system abstained. Files can be processed locally; ToothPrint itself does not store a patient database.

For dentists and clinical reviewers

The practical workflow is longitudinal: compare a current radiograph, scan, or reconstructed surface against a prior visit, then export a report that separates changed, stable, and uncertain findings. The important promise is not "AI says so" — it is a measurement with an uncertainty interval and a stated false-alarm bound.

For hospitals, labs, and IT

The relevant surface is governance: local deployment, hardened parsers for untrusted medical files, explicit no-fallback behavior, provenance hashes, audit metadata, and a clear line between research validation and lawful clinical use. The code is ready to inspect; real deployment still requires site calibration, prospective validation, access control, and regulatory work.

For researchers

The contribution is a dental-imaging benchmark scaffold around conformal certificates: 3D identity, partial-overlap correspondence, radiograph identity, dental-work identification, change certification, surface-change certification, and photo-to-mesh reconstruction. The README keeps the strong numbers and the negative results together so the limits are visible.

For engineers

Use the package as a library, FastAPI service, or desktop Studio shell. The ingest layer detects files by content, streams uploads under caps, rejects hostile files cleanly, and exposes the same identity/change/surface primitives used by the tests and evaluation scripts.

What a result looks like

1. Upload an intraoral scan, radiograph, CBCT-derived mesh, or patient video.
2. ToothPrint extracts the relevant dental geometry: landmarks, arch surface, restorations, or reconstructed mesh.
3. It compares against a reference, a gallery, or a prior visit.
4. It returns one of three clinical-style outcomes: match / no confident match / abstain, changed / stable / uncertain, or a regional surface-change map.
5. ToothPrint Studio exports a PDF case report with inputs, findings, method, provenance, and limitations.

Visual tour

The top-level story is dental first, chart second:

Quick start

pip install -e ".[dev,io,api,desktop]"
python -m pytest -q
TOOTHPRINT_FIXTURES=1 PYTHONPATH=. python evaluation/scripts/smoke_test.py
uvicorn api.main:app

python -m desktop.app           # native Studio window; falls back to browser if headless

Contents — Start here · Status · Readers · Visual tour · Results · Identity · Change · Surface · Reconstruction · Studio · Formats · How it works · Clinical readiness · Security · Real-data gate · Reproduce

Results

The verdict first — not just the data. Every mechanism is held to the bar that actually matters (recognise the right person, flag real change without crying wolf, rebuild a usable mesh), and here is whether we clear it:

Capability	What "good" requires	Result
Identity — 3D scans	a stranger never outranks you	Rank-1 0.995 (N=200, EER 0.005, AUC 0.997), conformal FMR ≤ α, open-set FNIR 0.032, fidelity 0.05 mm	✅
Identity — partial overlap	survive missing teeth	Rank-1 0.87 at 50% tooth loss (learned correspondence; ~3.8× rigid GICP's 0.23)	✅
Identity — 2D radiographs	pick the right person out of hundreds	Rank-1 1.000 (N=400, EER 0), robust to jitter & magnification	✅
Identity — certified decision	one accept/abstain verdict, bounded FMR	unified FNIR@FMR=1% 0.00 full-coverage; abstains under heavy tooth loss	✅
Identity — dental work	identify by the restoration pattern	0.93 CBCT · 0.91–0.99 radiograph (forensic chart; n=55 / 165)	✅
Multimodal fusion	combine independent biometrics	quality-weighted fusion improves Rank-1 in the hard regime (oracle bound 1.0, complementary)	✅
Change — measurement	flag a real shift while avoiding false progression	recall 0.98 @ 0% false-progression	✅
Change — fully automatic	detector finds the teeth	0.91 end-to-end (YOLO26-pose; 0.98 ceiling)	🔄
Surface certificate	catch a lesion a global average misses	0.99 localized vs 0.00 naive (n=8), to 0.4 mm noise, 0% false-change	✅
Reconstruction	sharp enough for certificate experiments (≈0.5 mm target)	~0.3 mm median 2DGS mesh, 38% better than 3DGS (n=5)	✅
Cross-dataset generalization	works on a second real dataset	Teeth3DS Rank-1 1.0 (registration); learned descriptors carry an honest domain gap (0.87→0.42)	🔄

Specificity is the design target: the conformal false-positive rate is bounded by α in finite samples, and held a true 0 in most tests. All identity numbers are measured on public single-timepoint scans with synthetic re-scans/crops, so read them as in-simulation ceilings — the one binding gate is real cross-session data.

---

🪪 Identity — recognise a person by their teeth

ToothPrint identifies a person from the shape of their dental arch — at full coverage, and when half the teeth are missing, the case where every prior rigid method collapses. The four results that matter:

Full coverage: a stranger never wins

A query re-scan is given its best rigid alignment to each gallery arch (PCA principal-axis init + multi-scale Generalized-ICP — a global init the self-similar palate can't fool, rigid so no scale collapse), then scored by distance to the arch surface. For the same person the cloud sits on the surface (~0.05 mm, below sensor noise); for a stranger the best fit still floats ~4 mm off.

The discrimination holds across people (genuine ≤ 0.05 mm vs impostor ≥ 1.6 mm in every case), and the alignment is truly exact — a swept cross-section shows the genuine cloud on the surface at every depth:

At scale — all 200 subjects: Rank-1 0.995, EER 0.005, AUC 0.997 (95% CI 0.989–0.998). And a first for dental identity — full Detection-Error-Tradeoff curves (the literature reports only Rank-N): full-coverage 3D sits at EER 0.5%, learned correspondence holds a real curve even at 50% tooth loss (EER 10%), and the degraded multimodal regime is drawn honestly (EER 13%).

The certificate — bounded false-match rate, hardened against look-alikes

Two results go beyond the usual table. The decision is conformal — empirical false-match rate tracks the target α; no learned dental-ID method reports a finite-sample FMR bound — and it works open-set (a non-enrolled query is rejected, FNIR 0.032 at 1% false-positive identification). The bound even survives hard negatives: re-calibrated against each subject's nearest impostor (the ~0.8 mm look-alike, not a random arch), empirical FMR still tracks α (0.016 at α=0.01) for a 1–4% genuine-accept cost. A literature FMR on random arches is easy; holding it against look-alikes is the honest test.

The partial-overlap breakthrough — learned point correspondence

A query missing half its teeth is the genuine open problem in dental identity: rigid GICP collapses to Rank-1 0.23, because the PCA-init can't align a missing-teeth query. We solve it in two layers, ruling out the cheap fix first:

1. Not the scoring — robust statistics (trimmed mean, median, fitness) of a wrong alignment don't help (they tie or trail the plain mean). The bottleneck is the descriptor, not the score.
2. Crop-hardening the embedding — a DGCNN + sub-centre ArcFace encoder, re-trained on aggressive partial crops (keep ≥ 0.35), lifts keep-0.5 from 0.46 → 0.64 at no full-coverage cost. This raised the floor.
3. CorrNet breaks the ceiling — instead of one pooled descriptor, CorrNet emits a unit descriptor per point (InfoNCE correspondence loss; crops give ground-truth matches for free). A partial query is matched point-to-point against each gallery arch (mutual NN → weighted Procrustes), scored by the residual over all query points.

Under realistic discrete whole-tooth dropout (random, non-contiguous teeth removed — not a clean slice), on held-out unseen subjects:

method	keep-0.5	keep-0.3 (70% gone)
rigid GICP	0.23	0.10
crop-hardened embedding	0.64	0.26
CorrNet (learned correspondence)	0.87 (AUC 0.98)	0.57 (AUC 0.94)

The honesty checks that shaped these numbers — both kept, both moved numbers down: a clean planar half-cut is easier (0.91 / 0.80), so the extreme figure was partly riding on crop geometry — the realistic dropout above is what we report. And an untrained CorrNet already scores 0.72 at keep-0.5, so the win is ~half the correspondence-plus-rigid-verification architecture (the right inductive bias) and ~half the learned descriptors.

Cross-dataset honesty. The learned descriptors are partly dataset-specific: CorrNet trained only on Poseidon3D, run on Teeth3DS+ (a different real dataset, all unseen), drops keep-0.5 0.87 → 0.42 — still 34× chance and still above GICP, but a real domain gap that needs multi-dataset training to close. Recorded, not hidden.

The partial-overlap picture on real arches — one chart

Beyond the bespoke Poseidon3D pipeline, two off-the-shelf point-cloud methods are part of the evaluation toolkit and were run on real Teeth3DS+ upper arches (N=40; BUFFER-X keep-0.5/keep-0.3 over 3 crop-seed reps so the numbers carry error bars) at the identical whole-tooth-dropout crop protocol — one a decisive positive, one an honest negative:

method (dataset · provenance)	keep-1.0	keep-0.5	keep-0.3
rigid GICP (Poseidon3D · recorded)	—	0.23	0.10
DGCNN embedding (Poseidon3D · recorded)	0.96	—	—
CorrNet (Poseidon3D · recorded)	—	0.87	0.57
BUFFER-X zero-shot (Teeth3DS+ · measured)	1.00	1.00	0.95 (0.90–0.98)
Sonata frozen head (Teeth3DS+ · measured)	0.28	0.13	0.03

BUFFER-X keep-0.5/keep-0.3 are the mean over 3 crop-seed reps: keep-0.5 held 1.00 in all 3 reps, keep-0.3 was 0.95 ± 0.04 (min–max 0.90–0.98) — the headline is stable, now with error bars (whiskers in the chart). Its keep-1.0 cell is a measured full-coverage number, not a placeholder: the same BUFFER-X backend on the full-coverage identity protocol (the one the GICP smoke uses) scores Rank-1 1.000 · Rank-5 1.000 · EER 0.000 · AUC 1.000 (N=40, bufferx_identity_full.json).

Cross-dataset by construction: BUFFER-X and Sonata are measured on Teeth3DS+; the recorded CorrNet / GICP / DGCNN rows are on Poseidon3D (not on this machine). The data is single-timepoint with synthetic genuine re-scans — still short of the longitudinal gate #7. Blank cells are honestly missing, not zero.

Zero-shot registration transfers to teeth — the positive. BUFFER-X (Kim et al., ICCV 2025) is a generalist point-cloud registrar pretrained on indoor RGB-D scans with no dental training. On the real arches it registers dental micro-geometry cleanly: realistic whole-tooth-dropout Rank-1 1.00 @ keep-0.5 (all 3 reps) and 0.95 @ keep-0.3 (min–max 0.90–0.98 over 3 crop-seed reps; clean planar cut 1.00 / 1.00), and a complete full-coverage column (Rank-1/Rank-5/EER/AUC all 1.000). Read honestly: BUFFER-X gets a denser point budget as an indoor-scan method, and the genuine query is a crop of the same arch; the repo's own rigid GICP also reaches full-coverage Rank-1 1.0 on Teeth3DS+. So rigid + zero-shot registration is already strong on real dental arches — CorrNet's contribution is on the harder Poseidon3D distribution and as a certifiable, retrieve-then-verify component, not a claim that no generalist can register teeth. BUFFER-X is offered as an optional zero-shot registrar for the partial-overlap regime — it does not change the certified pipeline's defaults (PCA-init + Generalized-ICP for registration, CorrNet for the learned partial-overlap descriptor, conformal accept/abstain on top). (eval_bufferx_baseline.py → bufferx_baseline.json; full-coverage eval_bufferx_identity_full.py → bufferx_identity_full.json)

What you are watching — one real arch (014996KX), 50% of its teeth dropped, both registrars run live on this pair. Rigid GICP's PCA-init flips the missing-teeth half-arch into the wrong basin (RRE 110°, residual 5.2× the genuine-match floor); BUFFER-X locks on with no dental training (RRE 0.6°, at the floor). Honesty: GICP is not uniformly bad on Teeth3DS+ — it succeeds on most arches here, so this is one real failure case; the aggregate numbers (GICP keep-0.5 Rank-1 0.23 on Poseidon3D vs BUFFER-X 1.00) carry the story. Seeded generator + measured RRE/RTE/residual: evaluation/viz/bufferx_arch_registration.py → registration_demo_numbers.json.

A frozen foundation-model head does not (yet) transfer — the honest negative. The identity embedding ships a backbone="sonata" option: a PTv3 + Sonata self-supervised encoder behind the same ArcFace head and descriptor contract as the from-scratch DGCNN — to our knowledge the first application of a point-cloud foundation model to dental identity. The first measured result is a negative: a frozen Sonata encoder + ArcFace head, trained on 110 real arches, reaches only Rank-1 0.28 / 0.13 / 0.03 at keep 1.0 / 0.5 / 0.3 on 40 unseen subjects — well below the from-scratch DGCNN embedding (0.96 full-coverage) and the rigid pipeline (1.0 on Teeth3DS+). Frozen indoor-SSL features don't transfer to dental identity in this low-data, head-only recipe; the full fine-tune (TP_FREEZE=0) is untested and is the open next step. (sonata_identity.json; install + run recipe in evaluation/scripts/RUN.md.)

The registration identity pipeline reproduces on real arches, too. Run on real Teeth3DS+ upper arches, the PCA-init + multi-scale Generalized-ICP pipeline scores Rank-1 1.000 · EER 0.000 · AUC 1.000 · d′ 6.07 (N=40, freshly re-downloaded md5-verified OSF data; teeth3ds_identity_smoke_n40.json). Swapping the backend to BUFFER-X on the same full-coverage protocol and 40 subjects gives the same complete column — Rank-1 1.000 · Rank-5 1.000 · EER 0.000 · AUC 1.000 (bufferx_identity_full.json; its lower d′ 1.24 only reflects the different residual scale of the two backends — AUC 1.0 / EER 0 confirm every genuine still outranks every impostor). Both are real arches, but single-timepoint with synthetic genuine re-scans, so still not the longitudinal gate #7.

Reproduce: train_correspondence.py → eval_correspondence.py → correspondence_identity.json; cross-dataset in eval_correspondence_teeth3ds.py; the real-arch baselines in eval_bufferx_baseline.py and train_sonata_embedding.py.

Two more modalities, and a forensic signal

2D radiographs — same rigor. A per-tooth landmark constellation (scale-normalised so magnification cancels) recognises a person from a single radiograph: across N=400 DenPAR images, Rank-1 1.000, EER 0, d′ 4.01, robust to 20 px jitter and 50% magnification.

Multimodal fusion on real paired data. On the Figshare CBCT+IOS set, every patient has three real biometrics — IOS crowns, CBCT bone/root geometry, CBCT dental-work pattern. Scored separately and fused: IOS 1.000, bone 0.945, dental-work 0.927. Does fusion beat the best single? Degraded into a hard regime to find out: the oracle bound is 1.000 (modalities genuinely complementary), naive equal-weight fusion hurts (dilution — a real negative), and quality-weighted fusion edges past the best single (0.867 vs 0.833). A real but Rank-1-only gain (its AUC regresses) — honest about the size.

Dental work as a forensic identifier — on CBCT and radiographs. The restoration pattern (fillings, crowns, implants) is a classic identifier. On CBCT it's directly observable (metal/ceramic HU > 2500): the restoration cloud alone identifies people at Rank-1 0.927. And it extends to 2D radiographs — global thresholding fails (over-saturated JPEGs grab anatomy), but a per-tooth local-contrast extractor (a restoration is the patch far brighter than its own tooth's median) recovers a constellation that identifies 165 restoration-bearing DenPAR subjects at Rank-1 0.91–0.99 (robust to jitter + a dropped restoration; chance 0.006). A classic forensic cue now works on the modality most clinics already have.

The unified certified decision — one verdict from the parts

The full matcher assembles into a single accept/abstain certificate: retrieve by the embedding (recall, partial-robust) → verify the shortlist by CorrNet correspondence (precision) → accept only above a conformal threshold (FMR ≤ α), else abstain. At full coverage FNIR@FMR=1% = 0.00 — every genuine accepted and correctly identified, impostors rejected at the bound. Under partial overlap it degrades to 0.74 — best of all methods, but open-set rejection is a near-full-coverage property, so the unified decision abstains under heavy tooth loss rather than risk a false accept. One pipeline, one conformal verdict, honest about when to decline. (eval_unified.py)

Robustness ablations + the partial-overlap miss, shown not hidden

Rank-1 holds at 1.0 through 0.4 mm sensor noise and 4× voxel coarsening, every query already repositioned (rotation + translation). The one rigid-method degradation is tooth loss (keep 0.5 → 0.23 for plain GICP) — which the learned correspondence above recovers to 0.87.

---

📐 Change — certify a bone-level shift

On a real DenPAR tooth the bone margin recedes between visits and the certificate's sub-pixel registration tracks it live (green = baseline, red = now), flipping to changed only once it clears the clinical threshold — false progression bounded by α:

Measured differentially (sub-pixel registration of the margin between timepoints, not by re-detecting landmarks), recall is 0.98 even when the threshold is set so false-progression is a true 0. The fully-automatic pipeline reaches 0.91 with a fine-tuned YOLO26-pose detector localizing the CEJ/bone-crest to a median 18 px (vs ViTPose's ~38 px / 0.81).

Why 0.91 is the honest ceiling, not laziness. We trained a 2× larger detector at 1280 px to push further — it made localization worse (20 px) and recall lower (0.87). The ~18 px floor is DenPAR label noise, not model capacity: only better labels or real longitudinal pairs close the gap to the 0.98 measurement ceiling. The negative result is kept so nobody re-runs it.

Robust to repositioning. Between visits a patient is re-seated at a different angle and distance. A multi-anchor affine model cancels the full motion, dropping spurious "change" ~8× vs a single-crown reference:

---

🗺️ Surface — certify 3D change

The displacement is measured differentially and de-biased (subtracting the reconstruction-noise power, which a naive mean-of-distances would rectify into a false signal), extending usable noise from 0.1 mm to 0.4 mm. It's also regional: a real lesion moves a patch that a whole-surface average dilutes to nothing (recall 0.00), which a per-region max statistic recovers (0.99, n=8) and localizes — with the conformal false-change rate still 0:

Benchmarked against M3C2 (the geomorphology-standard change distance) on a 0.5 mm lesion over 2% of the arch: both localize what the whole-surface average dilutes away. M3C2 edges us on raw recall at extreme noise — we don't claim to beat it there; our complementary edge is the finite-sample conformal false-change bound it lacks.

---

🧱 Reconstruction — photos to a dentist-usable mesh

No scanner? 2D Gaussian Splatting (oriented surfels) + multi-view TSDF fusion rebuilds a real arch from shaded photos into a watertight ~1 M-triangle mesh. 2DGS disks lie on the surface, so meshing from the median depth (the first-surface crossing, not the alpha-weighted mean that averages an arch's front and back walls) is markedly sharper than 3DGS: ~0.3 mm median, 38% better (2.4× on the hardest arch).

Candidate front-ends to push past the current ~0.264 mm median (MeshSplat, SurfelSplat, Gaussian-Voxel Duet) — and why the dental-splatting systems that report no mm-level ground-truth accuracy and do not certify aren't a fit — are catalogued in docs/RECON_UPGRADES_2026.md.

---

ToothPrint Studio

ToothPrint Studio (Linux · Windows · macOS) presents the three certificates as a forensic Certificate of Dental Analysis: drop in any scan, radiograph, or patient video — videos play in-place with full play / pause / seek — run an examination, and every finding is logged with its conformal interval. Export the whole case as a self-contained PDF report (inputs, findings, method, provenance, and an explicit limitations statement). Files never leave the machine.

pip install -e ".[api,io,desktop]"
python -m desktop.app           # native window; falls back to the browser if headless

Build native installers per OS with the bundled PyInstaller spec — see desktop/README.md.

Reads every format a dentist has

toothprint.io ingests every common dental format behind one safe loader — radiographs (DICOM, PNG/JPG/TIFF/BMP), intraoral scans (STL/PLY/OBJ/OFF/GLB/3MF), CBCT volumes (NIfTI, DICOM series) — detected by content (magic bytes), normalized to known units, and hardened against hostile files (decompression bombs, billion-element headers, external-reference smuggling).

from toothprint.io import load
scan = load("patient_upper.stl")     # -> Scan (vertices in mm)
xray = load("bitewing.dcm")          # -> Radiograph (MONOCHROME1-corrected, pixel spacing)

How it works

scan / radiograph ─▶ detect ─▶ register ─▶ certify
                     teeth +    rigid /     conformal interval ─▶ identity
                     landmarks  learned     (FMR ≤ α)           ─▶ change
                     or cloud   correspond.                     ─▶ surface

• Identity (3D): best rigid alignment to each gallery arch (PCA-init → Generalized-ICP), smallest mean surface distance wins; CorrNet learned point-correspondence for the partial-overlap regime; embedding-retrieve → verify → conformal accept/abstain for the unified decision.
• Identity (2D): per-tooth landmark constellation, scale-normalised, rigid-ICP aligned.
• Change: bone-level shift measured differentially (sub-pixel template matching, multi-anchor affine motion model), certified conformally.
• Surface: de-biased, regional differential displacement, max region certified conformally.

import numpy as np
from toothprint.identity import identify_surface
from toothprint.surface import certify_surface_change

distances = identify_surface(query_points, gallery_scans, voxel_size=0.5)
person = labels[int(np.argmin(distances))]
verdict = certify_surface_change(measured_mm=1.2, certifier=certifier)   # -> "changed"

A FastAPI service exposes the same logic plus safe file ingest (pip install -e ".[api]", then uvicorn api.main:app): /api/inspect, /api/identify/scan, /api/identify/radiograph, /api/certify/change, /api/certify/surface.

---

vs the state of the art

Honest placement against the published literature (per mechanism)

ToothPrint is in the registration / conformal family; its defensible edge is certification (finite-sample bounded error) + learned partial-overlap robustness, a lane the SOTA literature leaves almost empty — not a higher saturated accuracy.

Identity. The only dedicated 3D-IOS identity pipeline with proper Rank-1 is Zhou et al. 2024 (Bioengineering) — FPFH + SAC-IA + ICP + RMSE, the same family as ours: Rank-1 100% on 160 real adults with ~1-yr re-scans (closed-set, private). We match it (Rank-1 0.995 / AUC 0.997 on 200 arches) and lead on a certified, bounded-FMR decision + open-set rejection that no prior dental work reports, and on learned partial-overlap correspondence (0.87 at 50% tooth loss vs rigid 0.23) which the dental literature lacks. We trail on real cross-session data (ours is synthetic re-scan).

Change. Cross-sectional periodontal staging tops out at ~84% (meta-analysis); we don't compete there. Differential longitudinal change + a conformal false-progression certificate is named THE open problem by three systematic reviews and is unoccupied. Our detector front-end (~18 px) is the honest weak point.

Surface. Benchmarked vs M3C2 (Lague 2013): both localize a lesion the whole-surface average dilutes; M3C2 edges us at extreme noise, we add the conformal false-change bound it lacks.

Reconstruction. SOTA GS→mesh reports sub-mm Chamfer on DTU/T&T; ours reaches ~0.3 mm median at arch scale (38% better than our 3DGS baseline) — usable for the surface certificate's edge.

Bottom line: competitive on saturated metrics, ahead on the axis nobody occupies — a distribution-free certificate on every verdict, and the only learned partial-overlap matcher for this domain. The honest gaps (real longitudinal data, the radiograph detector) are data/label limits, not method.

Clinical readiness

What stands between this and lawful hospital use

Status: NOT clinically deployable. The engineering is built with deployment concerns in mind; the gate is real-world validation and regulatory clearance, neither producible from code or synthetic data. ✅ done · ⬜ requires real data / an institution / a regulator.

Item	Status
Sound, current methods (GICP, learned correspondence, conformal prediction, 2DGS)	✅
Finite-sample false-alarm guarantee, verified ≤ α across ablations	✅
No fallbacks — failures raise, not degrade	✅
First-class abstention + input quality gates	✅
Site recalibration + provenance hashing + audit trail · covered core test suite	✅
Reproducible from committed fixtures (no off-machine data)	✅
Real multi-session / longitudinal data; cross-device, cross-operator	⬜
Expert ground truth; prospective pre-registered study; demographic diversity	⬜
ISO 13485 / 14971; FDA 510(k)/De Novo or CE-MDR; post-market surveillance	⬜

The path: partner with a clinical/forensic institution (IRB) → acquire real longitudinal data → recalibrate and re-measure (expect lower numbers) → prospective study → QMS → regulatory clearance → supervised pilot. No step from acquisition onward can be completed in code.

Model card

Intended use, out-of-scope use, performance, ethics (Mitchell et al. 2019)

Intended use: research and method development; and decision-support, expert-in-the-loop, after site-specific validation — a ranked candidate list a forensic examiner confirms (never an automated identity decision), and longitudinal change flagging for clinician review. The system outputs a certificate with an explicit abstention; it defers, not over-claims.

Out-of-scope / prohibited: autonomous clinical or forensic decisions; any patient-affecting use without recalibration, prospective validation, and regulatory clearance; populations/devices/pathologies outside the validated distribution.

Performance: identity Rank-1 0.995 / EER 0.005 and partial-overlap 0.87 @ 50% tooth loss (N=200/50 synthetic re-scans/crops) with a conformal FMR bound; change/surface recall strong only in good-quality regimes and degrade under noise (quantified in the report and PAPER.md).

Limitations: validated on synthetic perturbations of single-timepoint data (optimistic ceilings); the learned correspondence carries a cross-dataset domain gap (0.87 → 0.42); open-set rejection collapses under heavy partial overlap (→ abstain); tooth-detection caps end-to-end change recall at ~0.91; no demographic/device/pathology diversity.

Ethics: dental biometrics is sensitive personal data — galleries must be consented, encrypted, access-controlled, retention-limited (ToothPrint stores none; it is a library). Bias risk is unverified across age/ethnicity/dentition. A false identification has severe consequences; identifications must stay expert-confirmed with the candidate evidence shown.

Risk analysis

ISO 14971-style hazard register

#	Hazard	Severity	Mitigation (implemented)	Residual
H1	False "no change" — progression missed	Critical	conformal abstention; noise/repositioning-robust registration; regional detection	High until real longitudinal validation; never a sole screen
H2	False "change" — needless intervention	Serious	conformal FPR ≤ α, verified in all ablations	Holds only if calibrated on the deployment distribution
H3	Mis-identification	Critical	smallest-distance + separation reported; conformal open-set; abstains under partial overlap; expert confirmation mandated	Cross-session unvalidated; show candidates, never auto-decide
H4	Garbage-in verdict	Serious	quality gates → abstain ("recapture")	Thresholds need per-site tuning
H5	Untraceable decision	Serious	append-only audit log (input hash, calibration id, operator, time)	Must wire into the deploying record system
H6	Guarantee voided by distribution shift	Critical	site recalibration + provenance hash + min-sample floor	Requires the site to recalibrate
H7	Silent failure	Serious	no fallbacks — bad input / missing dep / unmeasurable patch raise	—
H8	Biased accuracy	Serious	documented out-of-scope until validated	High — needs diverse prospective data
H9	Privacy breach	Serious	library stores no patient data; integrator owns consent/encryption	Deployment-dependent

Principles built in: abstain over guess · no fallbacks · provenance on everything · expert-in-the-loop.

Security

Threat model for untrusted medical files + the API

Medical parsers are a documented attack surface. Controls (toothprint/io/_limits.py):

• Decompression / element bombs → hard caps before decode: file 1 GiB, decoded pixels 120 MP, mesh 25 M/50 M verts/faces, volume 1.5 G voxels; gzip ISIZE + zip directory checked before inflation.
• File-type confusion → detection by magic bytes, not extension.
• Path traversal / SSRF → materials/textures not loaded (OBJ mtllib, GLB URIs skipped); no loader follows an external reference.
• Process integrity → every loader raises a ValueError subclass; the API maps these to a clean 422.
• API availability → Pydantic bounds every list/scalar, rejects NaN/Inf; middleware caps bodies (16 MiB), streams uploads under a 1 GiB cap; security headers on every response.

Not covered (deployment responsibilities): authn/authz, rate limiting, TLS, HIPAA/GDPR retention.

The real-data gate (gate #7)

Every headline here is an optimistic ceiling — all of it is measured on single-timepoint public data with synthetic re-scans and crops. The one binding constraint between ToothPrint and a real-world-validated claim is real same-patient data across two timepoints, locked behind academic data-use agreements I can't sign for a personal project. It is tracked openly as gate #7 in evaluation/DATA_GATE.md — a Zenodo 11392406 DUA application checklist plus the credentialed PhysioNet path — and it is open progress: I'm looking for help.

• Multimodal Dental Dataset (PhysioNet) — 169 patients, multi-visit timestamps + CBCT + 16k periapical radiographs. Credentialed. https://physionet.org/content/multimodal-dental-dataset/
• 3D pre/post-orthodontic models (Zenodo 11392406) — 1,060 real pre/post 3D intraoral pairs from 435 patients, our exact modality. Maintained by Prof. Liu Yongjin, Tsinghua University — [e-mail]. https://zenodo.org/records/11392406

How you can help: if you can legitimately access either dataset, or you know the maintainers — send or forward the note below to them, and point them back to me for the reply at [e-mail]. An introduction is worth as much as the data itself. Any contribution gratefully credited.

📧 The note to send / forward to the dataset maintainer

Subject: Introduction — dataset access for an independent dental-imaging researcher

Dear [maintainer],

I'd like to introduce Krishi Attri, a master's student building ToothPrint — an independent, non-commercial, conformal-certified system for dental identity and longitudinal change detection. The methods are currently validated only on synthetic perturbations of single-timepoint public scans, and your [dataset]'s real same-patient [pre/post · multi-visit] data is exactly what's needed to test them on genuine re-acquisitions.

Would you be open to granting access under whatever terms you require? Krishi will use it strictly for non-commercial research, will not redistribute it, and will acknowledge the dataset in any results. You can reach Krishi directly at [e-mail] to arrange access or ask anything about the project.

Thank you for considering.

Writing on your own behalf rather than forwarding? Just swap the opening for "I'm Krishi Attri, …" — either way, replies come to [e-mail].

Reproduce

Datasets are large, license-gated, and gitignored — so reproducibility is built in, not "trust the JSON":

# runs the 3D identity pipeline end-to-end on committed synthetic fixtures, NO off-machine data:
TOOTHPRINT_FIXTURES=1 PYTHONPATH=. python evaluation/scripts/smoke_test.py   # -> Rank-1 1.000, SMOKE OK

# point at your own data via env vars (TP_POSEIDON3D, TP_TEETH3DS, TP_CBCT_IOS, TP_DENPAR):
TP_POSEIDON3D=/data/poseidon3d PYTHONPATH=. python evaluation/scripts/eval_correspondence.py

Every result JSON maps to a committed script; reference baselines are read from committed artifacts, not pasted constants. Full guide: REPRODUCE.md. The Teeth3DS+ cross-dataset identity result runs on freshly re-downloaded, md5-verified ungated OSF data (smoke-scale, single-timepoint — still not the real-longitudinal gate); acquisition + numbers in evaluation/DATA_GATE.md.

Tests

python -m pytest -q → 183 passed across identity, change, surface, io, api, clinical, geometry, and embedding suites. Release history: CHANGELOG.md.

pip install -e ".[dev,io,api]"
pytest --cov=toothprint --cov=api      # core package/API coverage gate

Layout

toothprint/   the library — identity · change · surface · io · clinical
api/          FastAPI service (hardened; safe file ingest)
desktop/      ToothPrint Studio cross-platform app + PyInstaller spec
evaluation/   eval scripts · result JSONs · synthetic fixtures · smoke test
docs/         result figures (the write-up lives in PAPER.md + the engrXiv preprint)

Citation

ToothPrint is described in the preprint "Certified Dental Biometric Verification Under Partial Overlap" (Krishi Attri, Seoul National University), posted to engrXiv, 2026.

• 📄 Preprint: https://engrxiv.org/preprint/view/7403
• 🔖 DOI: 10.31224/7403

@article{attri2026toothprint,
  title   = {Certified Dental Biometric Verification Under Partial Overlap},
  author  = {Attri, Krishi},
  journal = {engrXiv preprint},
  year    = {2026},
  doi     = {10.31224/7403},
  url     = {https://doi.org/10.31224/7403}
}

License

PolyForm Noncommercial License 1.0.0 — free for everyone for any non-commercial purpose. That explicitly includes research, education, personal projects, and non-profit / public health & safety organizations — hospitals, clinics, forensic labs — regardless of how they are funded (the license's Noncommercial Organizations clause). Use it, modify it, share it, build on it.

What is not permitted: selling the software, folding it into a paid product or service, or any use for commercial advantage — so no one can fork ToothPrint and resell it (to hospitals or anyone). Need a commercial license? Ask: [e-mail].