Automatic caries detection in bitewing radiographs: part I-deep learning.

Area of Science:

• Artificial Intelligence in Dentistry
• Medical Imaging Analysis
• Machine Learning for Healthcare

Background:

• Dental caries detection is crucial for timely intervention.
• Manual interpretation of dental radiographs can be time-consuming and subjective.
• Automating caries detection can improve efficiency and accuracy in dental diagnostics.

Purpose of the Study:

• To develop and evaluate convolutional neural networks (CNNs) for automated dental caries detection in bitewing radiographs.
• To create a large, annotated dataset of bitewing radiographs for training and testing AI models.
• To achieve human-level performance in detecting carious lesions using deep learning.

Main Methods:

• A dataset of 3989 bitewing radiographs was annotated with 7257 carious lesions.
• Multiple CNN architectures (YOLOv5, Faster R-CNN, RetinaNet, EfficientDet) were trained and tested.
• Model ensembling and post-processing techniques were employed to enhance detection accuracy.

Main Results:

• Tested CNN architectures achieved F1 scores between 0.72-0.76.
• Model ensembling improved the F1 score to 0.79-0.80.
• The best ensemble model achieved a precision of 0.83, recall of 0.77, and an average precision (AP) of 0.86 at IoU=0.5, with slightly lower accuracy for small lesions (AP 0.82).

Conclusions:

• Ensemble of object detection CNNs demonstrates satisfactory accuracy for caries detection, comparable to experienced dentists.
• Automated detection of dental caries using CNNs is feasible.
• Detecting incipient carious lesions remains a challenge, potentially due to dataset inconsistencies.