Diagnostic Accuracy of Two Multimodal Artificial Intelligence Systems Versus Dental Students in Endodontic Radiographic Classification

Objectives: Multimodal artificial intelligence (AI) systems are increasingly used to support image interpretation in dentistry, but evidence regarding their diagnostic performance on endodontic radiographs remains limited. This study compared the diagnostic accuracy of two general-purpose multimodal AI systems with that of undergraduate and postgraduate dental students and evaluated whether image cropping influenced AI performance. Methods: Fifty anonymized periapical radiographs were retrospectively selected, with 10 cases in each of five predefined endodontic diagnostic categories. Reference diagnoses were established by consensus between two experienced endodontists. ChatGPT version 5.2 and Gemini version 3 independently classified each case using both full-size and tooth-cropped images and a standardized prompt. Twenty-five fifth-year undergraduate dental students and 20 postgraduate students in clinical endodontics classified the full-size images. Overall and category-specific accuracies were calculated, and agreement with the reference standard was assessed. Results: Overall accuracy was highest among postgraduate students (84.9%), followed by undergraduate students (78.7%). Gemini achieved accuracies of 72.0% with full-size images and 82.0% with cropped images, whereas ChatGPT achieved 68.0% and 50.0%, respectively. Cropping significantly reduced ChatGPT accuracy (p = 0.0117) but did not significantly affect Gemini performance (p = 0.3018). Diagnostic accuracy varied substantially across categories. Agreement with the reference standard for the AI systems ranged from 50.0% to 82.0%. Conclusions: Diagnostic performance depended on the evaluator, AI system, image format, and diagnostic category. Postgraduate students showed the highest overall accuracy, while AI performance was heterogeneous and did not consistently exceed that of trained human observers. General-purpose multimodal AI may support supervised radiographic interpretation and education, but it should not replace comprehensive clinical assessment or professional judgment.