Deep learning-based fractured tooth detection in occlusal radiographs

This study highlights the importance of accurate and quick diagnosis of detection, and classification of fractured teeth and the potential benefits of applying deep learning (DL) techniques to solve this problem.

In the study, occlusal radiography (OR) of the premaxilla is used for tooth fracture detection of teeth 11 and 21. For that, a dataset that contains 200 ORs of various tooth conditions was constructed. In the proposed method, teeth 11 and 21 are automatically detected in OR images using a YOLOv9-based machine learning framework as the first step. Then, images of these two teeth are obtained by cropping OR images using the bounding box coordinates of teeth 11 and 21, obtained by a YOLO-based detector. Finally, these cropped images are provided as input to a pre-trained CNN-based network to classify between the "fracture" or "non-fractured tooth". For this purpose, VGG19, EfficientNetB0, InceptionResNetV2, and InceptionV3 nets are employed, and the obtained classification results are fused by applying a majority voting step to further improve the performance.

The experimental studies obtained a 99.5% mean average precision (mAP50) score and 85.6% for mAP50-95. On the other hand, percent accuracy rates in the range from 84.67% to 87.94% were observed using five pre-trained networks, and the percent accuracy metric was improved to 91.94% using the majority voting-based fusion approach.

The findings indicate that the proposed method effectively detects fractured teeth by leveraging machine learning techniques. Furthermore, this approach provides a pioneering framework for integrating artificial intelligence (AI) methodologies into dental diagnostics, offering clinicians a reliable decision-support tool for improved diagnostic accuracy.