Can Large Language Models Detect Periapical Lesions in Anterior Teeth? A Comparative Study.

Area of Science:

• Artificial Intelligence in Dentistry
• Radiographic Diagnostics
• Medical Imaging Analysis

Background:

• Large language models (LLMs) are increasingly explored for medical applications.
• Periapical lesions require accurate radiographic detection for timely endodontic treatment.
• Evaluating the diagnostic capabilities of LLMs in dentistry is crucial.

Purpose of the Study:

• To assess the diagnostic performance of ChatGPT 5.0 and Gemini Flash 2.0 in identifying periapical lesions on periapical radiographs.
• To compare the accuracy, sensitivity, and specificity of these two LLMs.
• To determine the suitability of general-purpose LLMs for radiographic diagnostic tasks in endodontics.

Main Methods:

• Seventy-five periapical radiographs (with and without lesions) were analyzed by ChatGPT 5.0 and Gemini Flash 2.0 using a standardized prompt.
• A calibrated endodontic specialist provided the reference diagnosis.
• Performance metrics including balanced accuracy, sensitivity, specificity, and F1-score were calculated and compared using the McNemar test.

Main Results:

• ChatGPT 5.0 demonstrated higher balanced accuracy (0.54) than Gemini Flash 2.0 (0.48), with superior sensitivity (0.97 vs. 0.84).
• Both models exhibited poor specificity (0.11 for both) and high false-positive rates.
• Significant differences were found between the models (p < 0.05), favoring ChatGPT 5.0, but both struggled with tooth localization.

Conclusions:

• While LLMs show high sensitivity for detecting periapical lesions, their poor specificity limits their diagnostic utility.
• General-purpose LLMs currently exhibit a bias toward positive classifications and are not reliable for radiographic diagnosis in endodontics.
• Further development is needed to improve the accuracy and specificity of AI models for dental radiographic interpretation.