AI Triaged Single-Tooth Gaps From Phone Photos in Proof-of-Concept Study

Edited by Dentaltown staff

An artificial intelligence system reading ordinary occlusal photographs identified single-tooth edentulous spaces and sorted cases toward orthodontic or prosthodontic treatment with accuracy the authors describe as acceptable for preliminary triage, according to a proof-of-concept study published Aug. 17 in Scientific Reports.

The system was built for cases requiring alternatives to implant-based treatment, not as a general treatment planner. Researchers collected 2,962 occlusal photographs under routine conditions using smartphones and digital cameras, with two independent groups of dental specialists annotating the dataset.

A YOLOv8m model located single-tooth edentulous spaces, defined as gaps within the arch where one tooth is absent and both neighbors remain, along with the mesial and distal adjacent teeth. ResNet-50, ResNet-101, VGG-16, and VGG-19 models then classified the clinical condition and anatomical category of the adjacent teeth. A deterministic function converted the gap’s mesiodistal width from pixels to millimeters using mean central incisor widths.

On the clinical condition task, VGG-19 posted the highest macro F1 score at 0.928, while ResNet-101 returned the highest macro area under the curve at 0.961 and the highest weighted kappa at 0.927, with the narrowest confidence intervals on both. ResNet-101 also led the anatomical categorization task on F1 score, recall, and precision.

Two models handled the triage output itself. Logistic regression reached the highest area under the curve at 0.896 with an expected calibration error of 0.041, and XGBoost delivered the highest accuracy at 0.869, with sensitivity of 0.874 and specificity of 0.862.

The authors concluded that combining deep learning and machine learning can achieve acceptable performance for preliminary triage from occlusal images within acknowledged limitations, and that external validation and improved generalizability are necessary before clinical application.

Ehsan Shirdel, S. Azizipour, M. Magdalyanova, and Iliya Ashurko conducted the study at Sechenov First Moscow State Medical University and Pirogov Russian National Research Medical University.

The models were trained and tested on a single annotated dataset without external validation, so the reported figures describe internal performance rather than results in an independent clinical population.

Sources:
Scientific Reports, “A novel AI system for preliminary triage support in single-tooth edentulous spaces using intraoral images,” published Aug. 17, 2026: nature.com/articles/s41598-026-58704-7
PubMed, PMID 42608436: pubmed.ncbi.nlm.nih.gov/42608436