TLNM model maps teeth from smartphone photos with 90.1% accuracy in external tests
A Mask R-CNN pipeline turns ordinary phone pics into a remote dental screening tool.
Arash Nedaei and colleagues from the University of Oulu and Finnish healthcare partners developed TLNM, a customized Mask R-CNN pipeline for tooth localization, numbering, and segmentation directly from smartphone images. Unlike prior systems that rely on clinical radiographs or intraoral cameras, this model targets consumer-grade photos, addressing a critical gap in accessible oral health screening. The team trained on 1,272 annotated smartphone images and added two domain-informed mechanisms: a masked gray-world white-balancing algorithm to reduce artificial color casts and an anatomically constrained detection layer that suppresses false positives by enforcing structural validity.
In four evaluation stages, TLNM achieved an instance-mask AP@50 of 0.818, class-aware PQ of 0.780, and operational F1 of 0.884 on internal held-out tests. Remarkably, on an external dataset with different populations, sensors, and acquisition protocols, the model improved to AP@50 of 0.901, PQ of 0.832, and F1 of 0.928. Training stability was high, with a standard deviation of only 0.009 across ten runs. The full inference pipeline is released as an open-source, containerized API, allowing developers to build remote screening applications on standard phone hardware.
- Trained on 1,272 annotated smartphone images with Mask R-CNN and domain-specific preprocessing
- External validation achieved 90.1% AP@50 and 0.928 F1, beating internal test performance
- Open-source containerized API enables low-cost, scalable tele-dentistry deployments
Why It Matters
Consumer phones could replace clinical imaging for dental triage, enabling affordable remote screening in underserved regions.