CNN-based remote dental diagnosis model for caries detection with grad-CAM

Area of Science:

• Artificial Intelligence in Dentistry
• Medical Imaging Analysis
• Oral Health Diagnostics

Background:

• Dental caries is a widespread oral health issue requiring professional diagnosis.
• Access to traditional dental care is often hindered by cost, availability, and patient apprehension.
• Remote diagnostic tools are needed to improve dental care accessibility.

Purpose of the Study:

• To develop and evaluate a remote dental caries detection model.
• To leverage a ResBlock-AutoEncoder with domain-specific pre-trained weights for enhanced performance.
• To improve the interpretability and reliability of AI-driven caries diagnosis.

Main Methods:

• Implementation of a ResBlock-AutoEncoder model for caries detection.
• Utilizing domain-specific pre-trained weights to optimize model performance.
• Employing Grad-CAM for visual interpretability and localization of detected caries.
• Analysis of frontal oral images for caries identification.

Main Results:

• The model achieved high accuracy (0.9989), F1-score (0.9979), and precision (1.0).
• The average inference time was notably low at 5.7939 seconds.
• Grad-CAM effectively visualized caries, confirming model reliability.
• High precision was linked to the clarity of frontal oral images.

Conclusions:

• The developed remote caries detection model demonstrates exceptional accuracy and efficiency.
• The AI model shows promise in improving access to dental diagnostics.
• Future work will focus on incorporating multi-angle images to enhance model generalization across diverse oral imaging perspectives.