MetaDent: Labeling Clinical Images for Vision-Language Models in Dentistry.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Vision-language models (VLMs) have advanced medical image analysis.
• Intraoral photography analysis by VLMs is limited by data and benchmarks.
• Existing datasets lack the fine-grained annotations required for dental applications.

Purpose of the Study:

• Introduce MetaDent, a comprehensive resource for dental vision-language research.
• Establish benchmarks for evaluating VLMs on intraoral image understanding.
• Highlight the need for domain-adapted VLMs in dentistry.

Main Methods:

• Curated a large-scale dentistry image dataset (60,669 images).
• Developed a semi-structured meta-labeling framework for detailed annotations.
• Generated benchmark suites including visual question answering (VQA) and multilabel classification using LLMs.

Main Results:

• State-of-the-art VLMs achieved moderate accuracy (<70% in VQA) on intraoral images.
• VLMs demonstrated inconsistent performance in image captioning tasks.
• A significant gap exists between general VLMs and specialized dental image analysis requirements.

Conclusions:

• MetaDent provides a valuable resource for advancing dental AI.
• Current VLMs require domain-specific adaptation for clinical intraoral photography.
• Further research is needed for robust VLM development in dental practice and public health.