YOLOv8 was trained and validated on 1,180 bitewing radiographs and tested on 247 images annotated by two experienced dentists across six lesion depths. It exceeded general dentists on recall, precision, F1 and mAP, but both struggled with initial lesions.
Key findings
- The model outperformed dentists across reported metrics, yet recall of 0.51 and precision of 0.41 remain modest. It tended to under-stage lesions, whereas dentists tended to over-stage them.
Why this matters globally
Localized, severity-aware AI could reduce reading variability and support screening where expertise is scarce, but errors could drive under- or overtreatment.
Thai researcher contribution
Chiang Mai and University of Phayao researchers developed the dataset, reference process and clinical evaluation in primary teeth, with Onnida Wattanarat as corresponding author.
Limitations to consider
This retrospective single-dataset evaluation used a two-expert reference and lacks external prospective validation or evidence on treatment decisions, workflow time and safety.