Automated Cavity Detection and Classification Using Deep Learning

Cavity detection in dental X-rays is essential for early diagnosis and treatment planning, yet traditional deep learning approaches often rely on binary classification of entire images, overlooking localized analysis. This study introduces a multi-scale AI-assisted approach for both classification and detection, leveraging the Ultralytics YOLO11 framework to compare image-level and tooth-level methodologies. The single-tooth classification model achieved a test accuracy of 0.854, while panoramic classification performed slightly better at 0.864. For cavity detection, the tooth-level model outperformed the panoramic approach, with the test set achieving mAP@50 of 0.845 compared to 0.669, with significantly higher recall (0.743 vs. 0.394), highlighting challenges in full-image localization. These findings emphasize the trade-offs between segmentation-based and direct image-based approaches, demonstrating the advantages of tooth-level analysis for improved detection accuracy. Future work will refine segmentation techniques, expand clinical datasets, and validate performance across varied imaging conditions.