Towards clinical-level interpretation of dental panoramic radiography using an instance-guided vision-language model

Panoramic radiography is indispensable for dental diagnosis; however, the high demand for interpretation and limited radiologist availability often lead to reports that are incomplete, focused on the chief complaint or unavailable, thereby limiting comprehensive clinical assessment. We present DentFound, a vision-language model (VLM) for panoramic radiography diagnosis and report generation. To support its development, we assembled a large-scale dataset comprising more than 101,000 patients aged 2-98 years, spanning all dentition stages and covering 98 diseases and 11 post-treatment categories. Across multi-centre cohorts, DentFound outperformed state-of-the-art VLMs in both report generation and diagnosis. Expert evaluation by 12 dentists and radiologists showed that DentFound-generated reports were superior or comparable to human-written reports, owing to their broader diagnostic coverage and greater support for comprehensive clinical assessment. These results underscore the gap between existing VLMs and the demands of real-world dental practice, while highlighting the potential of DentFound in clinical practice.