Artificial intelligent recognition for multiple supernumerary teeth in periapical radiographs based on faster R-CNN and YOLOv8

Abstract

Objectives

The aim of this study was to compare the effectiveness of automated supernumerary tooth (ST) detection systems on periapical radiographs using Faster R-CNN and YOLOv8 with detection by 8 dental residents.

Methods

This was a diagnostic accuracy study of 469 periapical radiographs (419 training vs. 50 test datasets). The primary predictor variables were detectors (dental residents/Faster R-CNN/YOLOv8). The main outcome variables included the diagnostic performance of the model's using precision, recall and intersection over union (IoU). Appropriate statistics were calculated.

Results

In the test dataset, the precision of Faster R-CNN and YOLOv8 was 0.95 and 0.99, and their average precision was 0.90 and 0.97, respectively. A significant difference was observed between the two models in these metrics, with YOLOv8 outperforming Faster R-CNN in both precision and average precision (P<0.05). Both AI systems outperformed human subjects.

Conclusions

Based on our findings, both YOLOv8 and Faster R-CNN are highly effective in the automated detection of ST in periapical radiographs and could, for example, assist humans in resource-limited situations.