Evaluation of text- and image-based generative artificial intelligence for simulating impacted third molar extraction

Recent advances in generative artificial intelligence (AI) have opened new opportunities in procedural simulation. However, its clinical accuracy for simulating impacted mandibular third molar extraction remains unclear.

Both text-based and image-based surgical simulations were generated using generative AI. Simulations were produced based on a detailed description of a horizontally impacted mandibular third molar. Six expert evaluators independently assessed the quality of the generated images and text using the Global Quality Scale (GQS) and Likert Scale (LS). Statistical comparisons were performed using the Kruskal-Wallis test.

Panoramic images appeared visually realistic but initially exhibited inaccurate tooth positioning. Iterative prompt adjustments improved anatomical representation, with panoramic image scores significantly increasing from the first to the final generation. No significant improvement was observed for intraoperative images. All text-based procedural outputs received generally high scores, with no significant differences among steps.

Generative AI can generate text-based guidance appropriate for surgical simulation. However, image-based outputs demonstrated inconsistent improvement and may not provide sufficient anatomical accuracy for reliable use. To enhance clinical relevance, the development of specialized dental AI trained on diverse data sources-such as images, clinical records, and procedural data-and capable of supporting diagnosis and treatment planning is needed.