Detection of caries around restorations on bitewings using deep learning

Area of Science:

• Dentistry
• Artificial Intelligence
• Medical Imaging

Background:

• Secondary caries adjacent to restorations are a primary cause of restoration failure.
• Accurate diagnosis of caries is crucial for effective treatment and optimal outcomes.
• Traditional methods include visual inspection and radiographs, with potential for AI enhancement.

Purpose of the Study:

• To develop a convolutional neural network (CNN)-based algorithm for detecting primary and secondary caries lesions.
• To utilize deep learning for improved caries detection on bitewing radiographs.
• To create an automated system for aiding clinicians in daily practice.

Main Methods:

• A Mask-RCNN architecture with a Swin Transformer backbone was employed.
• Data augmentation and ten-fold cross-validation were used for model training.
• Diagnostic accuracy was assessed using Free-Response Receiver Operating Characteristics (FROC) curves, sensitivity, precision, and F1 scores.

Main Results:

• The CNN model achieved an area under the FROC curve of 0.806 for primary caries and 0.804 for secondary caries.
• F1 scores for primary and secondary caries detection were 0.689 and 0.719, respectively.
• The developed algorithm demonstrated accurate detection of both caries types.

Conclusions:

• An accurate CNN-based automated system for detecting primary and secondary caries on bitewings has been successfully developed.
• This represents a significant advancement in automated caries diagnostics.
• Integrating primary and secondary caries detection can lead to automated systems supporting clinical practice.