Efficient and Explainable Deep Learning Approach for Dental Caries Detection and Segmentation

Walid Brahmi, Imen JDEY, Fadoua Drira

Abstract:

Early and accurate detection of dental caries from panoramic X-rays is fundamental to
preventive dentistry, yet manual analysis is time-consuming and prone to subjectivity. This
study presents a novel multi-level framework for automated caries detection and segmentation,
designed to evaluate model robustness across varying levels of diagnostic complexity. We
utilize three datasets of increasing specificity: Full Panoramic Radiographs (FPR), Cropped
Panoramic Regions (CPR), and Single Tooth patches (ST). A rigorous comparative analysis is
conducted between the established You Only Look Once (YOLOv8) model and the (YOLOv11)
architecture, assessing their efficacy in this domain. To bridge the gap between model output and
clinical trust, Explainable Artificial Intelligence (XAI) techniques like Gradient-weighted Class
Activation Mapping (Grad-CAM) and Local Interpretable Model-agnostic Explanations (LIME)
are integrated to provide transparent visual and analytical explanations. Evaluation extends
beyond standard metrics (mean Average Precision (mAP), precision, recall) to encompass
computational efficiency and environmental impact, quantified via CO2 emissions. Results
demonstrate that YOLOv8 consistently and significantly outperforms YOLOv11 across all
dataset levels, achieving peak mean Average Precision (mAP@50) scores of 47.0% (FPR),
79.7% (CPR), and 90.0% (ST). YOLOv8 also exhibited superior training efficiency (2.06 h) and
a lower environmental cost (0.041 kg Carbon Dioxide (CO2) per 100 epochs). The integrated
XAI successfully illuminated model reasoning, validating its focus on clinically critical diag-
nostic regions. In conclusion, YOLOv8 is the preferred architecture for accurate, efficient, and
environmentally conscious caries detection. Its integration with XAI provides a transparent and
reliable tool, poised for effective incorporation into clinical dental workflows.