Detecting Dental Caries Using General-Purpose Large Multimodal Models From Oral Photographs

Area of Science:

• Artificial Intelligence in Dentistry
• Medical Image Analysis
• Teledentistry Applications

Background:

• Dental caries detection is crucial for oral health.
• Automated tools can enhance screening efficiency.
• Large multimodal models (LMMs) offer potential for image analysis without task-specific training.

Purpose of the Study:

• To evaluate a general-purpose LMM's ability to detect dental caries from intraoral photographs.
• To assess performance under zero-shot and few-shot learning conditions.
• To determine the feasibility of using LMMs for automated caries screening without fine-tuning.

Main Methods:

• Diagnostic accuracy study using 1255 public intraoral photographs.
• Gemini LMM queried via Vertex AI API with structured prompts.
• Evaluation of image-level classification and tooth-level localization.
• Performance metrics included sensitivity, precision, F1-score, and mAP@50.

Main Results:

• Performance varied by image view and model type, with best results on occlusal images using reasoning models.
• Zero-shot prompting achieved high sensitivity (0.95) but lower precision (0.80) at the image level.
• Five-shot prompting improved precision (0.87) and localization balance at the tooth level.
• Five-shot prompting yielded sensitivity, precision, F1-score, and mAP@50 of 0.77, 0.57, 0.64, and 0.63, respectively, for tooth-level detection.

Conclusions:

• General-purpose LMMs demonstrate potential as scalable, automated caries screening tools for teledentistry.
• Prompt engineering can modulate the sensitivity-precision trade-off.
• Clinical validation and governance are essential before deployment due to over-detection.