Integrated deep learning and geometric reasoning for detection, segmentation, numbering and anatomical mapping of teeth in panoramic dental radiographs, plus a smartphone photo screening aid that tells people when an X-ray is worth getting.

Research prototype. Not a medical device and not for diagnostic use. All findings require review by a qualified dentist.

A collaboration between Bharati Vidyapeeth (Deemed to be University) College of Engineering and Bharati Vidyapeeth Dental College & Hospital, Pune, developed under the C-CAMP Inter-Institutional Biomedical Innovations Programme (AICTE–C-CAMP IBIP).

        

What it does

Page	Input	Output
Analyse	A panoramic radiograph (OPG)	Per-tooth masks, FDI numbers, findings mapped to teeth (caries, restorations, endodontic, prosthetic), a dental chart, PDF / Markdown / JSON reports
Photo screening	Up to five phone photos of the mouth	Possible decay, wear and restorations marked on each photo, a triage level ("routine", "book a check-up soon", "see a dentist promptly"), a screening index and a PDF. It is a screening aid, not a diagnosis, and recommends an OPG when something is seen
How it works, Performance, About		The method, the measured numbers (read live from paper_eval/results), credits and licences

Measured performance

Every number names the configuration it comes from. Nothing here is estimated.

OPG analysis (716 held-out radiographs, no train overlap, Tesla T4, bootstrap 95 % CIs; paper_eval/). Published engine (v1, full_fix_routing: original + CLAHE passes, routing fix; the configuration in the paper):

Measure	Value
Tooth detection recall	94.9 % [94.0, 95.8] (n = 117)
Jaw, given the tooth was detected	97.9 %
FDI numbering, end to end	61.6 %; complete arches 76.6 % [70.7, 82.2] (n = 67), partial arches 35.4 % [27.7, 42.8] (n = 50)

The partial-arch collapse is the headline limitation of the published engine: one missing tooth shifts every later number in that jaw. A single pass with the routing fix (single_pass_fix_routing, evaluation only) measured FDI 64.5 % [58.7, 70.1], +2.88 pp [+0.09, +5.90] paired.

Improved numbering engine (v2, the app's default; tooth_ai/engine_v2.py). Designed on the validation split, reported once on the same test images (paper_eval/results/engine_v2.csv):

Measure	Published (v1)	Improved (v2)	Paired difference
FDI, 117 canonical images	61.6 %	85.8 %	+24.2 pp [+18.6, +29.9]
FDI, complete arches (67)	76.6 %	90.6 %	+14.0 pp [+7.7, +20.3]
FDI, partial arches (50)	35.4 %	77.4 %	+42.0 pp [+33.5, +50.6]
Tooth detection recall, full arches (93)	94.3 %	95.5 %	+1.2 pp [+0.7, +1.7]

v2 runs one pass, removes duplicate boxes, splits the jaws along a fitted bite line, caps to 32 teeth before numbering (never deleting a detected one) and numbers each tooth by its position along the arch, using a template learned on the training split. The Analyse page uses v2 by default; the published engine stays selectable under Numbering for comparison. Finding (caries, restoration) accuracy on OPGs has not been validated.

Photo screening (AKU official test split, 628 smartphone photos; photo_screening/EVAL.md), service defaults: 90.7 % [86.2, 93.8] of photos with caries get a check-up or dentist referral; 5.6 % [3.7, 8.3] of caries-free photos are referred anyway. Wear, filling and crown marks are not validated on phone photos. See photo_screening/MODEL_CARD.md for the rules and limitations.

Run it locally

Python 3.11. The repository lives on any disk; run commands from the repository root.

# The app (Streamlit). Demo mode needs no Detectron2: bundled sample cases run through the real engine.
pip install -r requirements-app.txt
TOOTH_AI_DEMO=1 streamlit run app.py

# Full mode also needs torch + Detectron2 and the weights in output/ (see docker/ui.Dockerfile for pins)
streamlit run app.py

# Photo screening API (the Photo screening page talks to it at PHOTO_API_URL, default localhost:8001)
pip install -r photo_screening/requirements.txt
python photo_screening/scripts/fetch_weights.py
uvicorn photo_screening.app.main:app --port 8001

# Tests
python -m pytest tests/
cd photo_screening && python -m pytest

On an Apple Silicon Mac the photo API uses the GPU (MPS) when available: about 1 s per photo.

Deploy

The public demo runs on an Oracle Cloud Always Free ARM VM (2 OCPU, 12 GB) with Docker and Caddy (automatic HTTPS). The images are built on an arm64 Mac and streamed to the VM; nothing is compiled there. Step-by-step runbook, costs and limits: deploy/README.md.

deploy/push.sh <vm-ip> --weights     # first deploy; afterwards: deploy/push.sh <vm-ip>

Repository

Path	What
app.py	Streamlit entry point (navigation, theme)
tooth_ai/	OPG pipeline (pipeline.py), analysis and reliability, rendering, reports, UI theme and pages
photo_screening/	Photo screening API, rules, PDF, UI page, training notebooks, evaluation (README)
paper_eval/	Evaluation harness and audit: the source of every OPG number shown
docker/, deploy/	App image, production compose, Caddy, VM setup and deploy scripts
assets/, static/fonts/	Brand assets (generated by tools/build_brand.py) and self-hosted fonts
tools/	Brand builder, demo fixtures, report figures (report_media/) and one-off dataset scripts (dataset/)
docs/	Brand and UX spec, product overview, requirements, set-up guides for Mac and Windows (setup/)
tests/	Unit tests for the OPG side (no GPU needed)
workspace/	Earlier phases, the previous single-file app (app_legacy.py) and old scripts, kept for reference

Model weights are not in git: OPG checkpoints live in output/<run>/, photo models in photo_screening/weights/ (fetch_weights.py, or the training notebooks' artifact zips).

Privacy

Uploads are processed in memory and never stored. Photo metadata (EXIF location, device, time) and file names are removed before analysis, and there are no accounts, databases or analytics. Use de-identified images only: crop or cover any name, date or ID printed on a radiograph.

Data and licences

• OPG models: trained on the dental radiograph dataset of Adnan et al., Data in Brief (2024), CC BY-NC 4.0, so ToothAI and its OPG weights are for non-commercial use.
• Photo models: AlphaDent (Apache-2.0), the AKU intraoral caries dataset (CC BY 4.0) and DentalAI (CC BY 4.0); attributions in photo_screening/MODEL_CARD.md.
• Fonts: Plus Jakarta Sans and JetBrains Mono, SIL Open Font Licence (static/fonts/).