Caries Detection with Near-Infrared Transillumination Using Deep Learning.

Area of Science:

• Dentistry
• Medical Imaging
• Artificial Intelligence

Background:

• Dental caries is the most common chronic disease globally.
• Early detection of caries is crucial for effective treatment and preventing invasive procedures.
• Near-infrared transillumination (TI) imaging shows potential for identifying early-stage dental lesions.

Purpose of the Study:

• To develop and evaluate a deep learning model for automated detection and localization of dental lesions in TI images.
• To address challenges in training data scarcity, class imbalance, and overfitting for dental image analysis.
• To assess the model's performance in both a 5-class semantic segmentation task and a binary classification task for caries presence.

Main Methods:

• Utilized a convolutional neural network (CNN) for a semantic segmentation task on TI images.
• Implemented strategies to overcome data limitations, including scarcity, imbalance, and overfitting.
• Evaluated the model on a 5-class segmentation task and a binary classification task for occlusal and proximal lesions.

Main Results:

• Achieved a 72.7% mean Intersection-over-Union (IOU) on the 5-class segmentation task.
• Obtained IOU scores of 49.5% for proximal and 49.0% for occlusal caries.
• Reached an Area Under the Receiver Operating Characteristic Curve (AUC) of 83.6% for occlusal and 85.6% for proximal lesions in binary classification.

Conclusions:

• Deep learning analysis of dental TI images can significantly enhance the speed and accuracy of caries detection.
• The developed model shows potential to support dental practitioners in diagnosis.
• This AI-driven approach may lead to improved patient outcomes by facilitating earlier and more precise interventions.