AI suggests an orthodontic or prosthetic route from an occlusal photograph (AUC 0.896)
On 2,962 occlusal photographs, models sorted single-tooth edentulous spaces towards an orthodontic or a prosthetic solution: logistic regression reached an AUC of 0.896 and XGBoost an accuracy of 0.869.
Triage from a single occlusal photograph
When one tooth is missing and a solution other than an implant is sought, the choice comes down to orthodontic space closure or a prosthetic replacement. The team around Shirdel tested whether that first dividing line can be drawn from an ordinary occlusal photograph.
A total of 2,962 occlusal photographs taken under routine conditions, with smartphones and digital cameras, were collected, and two groups of specialists annotated the images independently. A YOLOv8m model localised single-tooth edentulous spaces and their mesial and distal neighbours, while ResNet-50/101 and VGG-16/19 networks classified the clinical condition and the anatomical category of the adjacent teeth. The mesiodistal width of the space was converted from pixels to millimetres by a deterministic function anchored to the average width of the central incisors.
In classifying the clinical condition of the teeth, VGG-19 had the highest macro F1 (0.928), while ResNet-101 had the highest macro AUC (0.961) and weighted kappa (0.927), with the narrowest confidence intervals. For the triage recommendation itself, two classical models were compared: logistic regression gave the highest AUC (0.896) with a calibration error of 0.041, and XGBoost the highest accuracy (0.869) with a sensitivity of 0.874 and a specificity of 0.862.
Limitations
The authors explicitly call the work a proof of concept and state that external validation and better generalisability are prerequisites for clinical use. All the figures were obtained on a single dataset.
For your practice
- This is not a decision tool but a sorting tool: even when it matures, the output is a suggested direction, not a treatment plan.
- The methodological message already holds today, because photographs taken with a phone during routine work were good enough to train the models.
- If you already photograph occlusal surfaces, archive them systematically and at a consistent angle; such series become usable material as soon as tools like this are validated.