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

Area of Science:

• Artificial Intelligence in Medicine
• Dental Radiology
• Computer Vision

Background:

• Panoramic radiography is crucial for dental diagnosis but faces challenges with interpretation.
• Limited radiologist availability leads to incomplete or delayed reports, hindering comprehensive patient assessment.

Purpose of the Study:

• To develop and evaluate DentFound, a novel vision-language model (VLM) for automated panoramic radiography diagnosis and report generation.
• To improve the completeness and accuracy of dental diagnostic reports.

Main Methods:

• Assembled a large-scale dataset of over 101,000 panoramic radiographs from diverse age groups and dentition stages.
• Trained and evaluated DentFound VLM on a dataset covering 98 diseases and 11 post-treatment categories.
• Conducted multi-center cohort studies and expert evaluations comparing DentFound to state-of-the-art VLMs and human experts.

Main Results:

• DentFound demonstrated superior performance in both diagnosis and report generation compared to existing VLMs across multi-center cohorts.
• Expert evaluation revealed DentFound-generated reports were comparable or superior to human-written reports.
• The model showed broader diagnostic coverage and enhanced support for comprehensive clinical assessment.

Conclusions:

• DentFound shows significant potential to address the limitations of current panoramic radiography interpretation in clinical practice.
• The VLM can bridge the gap between AI capabilities and the demands of real-world dental diagnostics.
• DentFound offers a promising tool for improving diagnostic accuracy and efficiency in dentistry.