Reproduce “AI-Dentify: Deep learning for proximal caries detection on bitewing x-ray -- HUNT4 Oral Health Study” on OpenResearch

About this paper

Javier Pérez de Frutos, Ragnhild Holden Helland, Shreya Desai, Line Cathrine Nymoen, Thomas Langø, Theodor Remman, Abhijit Sen

Deep learning models were developed to detect proximal caries on bitewing X-rays using a large dataset from the HUNT4 Oral Health Study, demonstrating superior performance compared to individual human experts. The top-performing YOLOv5 model achieved a mean Average Precision of 0.633 and a mean False Negative Rate of 0.176 on a challenging consensus test set.

Key insights

• Deep learning object detection models, particularly YOLOv5, can effectively detect and classify proximal caries on bitewing X-rays, often outperforming individual human dental experts.
• The use of a large, prospectively collected, multi-expert annotated dataset is crucial for training robust and clinically relevant AI models in dental diagnostics.
• Off-the-shelf deep learning architectures, without extensive customization, demonstrate strong performance for complex dental diagnostic tasks, suggesting practical feasibility for clinical integration.

Summary by alphaXiv.