Lightweight Deep Learning for Automated Dental Caries Screening from Pediatric Oral Photographs

Area of Science:

• Artificial Intelligence in Dentistry
• Medical Imaging Analysis
• Public Health Informatics

Background:

• Early childhood caries (ECC) is a global health issue requiring early detection and intervention.
• Current deep learning models for dental diagnostics are often computationally intensive, hindering real-world deployment.
• There is a need for efficient, deployable AI solutions for ECC screening in community and mobile settings.

Purpose of the Study:

• To evaluate the performance of compact convolutional neural networks (CNNs) for detecting dental caries from oral photographs.
• To compare the diagnostic accuracy of different lightweight CNN architectures (ResNet-18, MobileNetV3-Small, EfficientNet-B0).
• To assess the clinical relevance and interpretability of AI-driven caries detection.

Main Methods:

• A dataset of 435 intraoral images from children (aged 3-14) was curated and annotated by dentists.
• Three CNNs were fine-tuned and evaluated using patient-level stratified splitting to ensure data integrity.
• Performance metrics included sensitivity, specificity, balanced accuracy, ROC-AUC, and PR-AUC, with interpretability analysis via Grad-CAM.

Main Results:

• ResNet-18 demonstrated high balanced accuracy (0.929) and perfect sensitivity (1.00).
• EfficientNet-B0 achieved superior threshold-independent performance with the highest ROC-AUC (0.978) and PR-AUC (0.990).
• MobileNetV3-Small offered competitive results with significantly reduced computational requirements.

Conclusions:

• Compact CNN architectures can achieve clinically meaningful performance for ECC detection from oral photographs.
• AI models demonstrated reliance on relevant dental features, validated by Grad-CAM interpretability analysis.
• These efficient models are suitable for scalable, real-world applications in dental caries screening.