Advances in Periodontal Diagnostics: Application of MultiModal Language Models in Visual Interpretation of Panoramic...

Area of Science:

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

Background:

• Periodontitis diagnosis relies on clinical exams and radiographic evaluation.
• Panoramic radiographs are cost-effective for assessing alveolar bone loss.
• Large language models (LLMs) are increasingly used in medical sciences, with some models now processing visual data.

Purpose of the Study:

• To evaluate the effectiveness of ChatGPT models (4.5, o1, o3, o4-mini-high) in measuring remaining bone height (RBH) and counting teeth from panoramic radiographs.
• To compare AI-driven measurements with evaluations by dental professionals.

Main Methods:

• Analysis of 10 panoramic radiographs, assessing 246-271 approximal sites per ChatGPT model for RBH measurement.
• Independent evaluation of RBH by three dental examiners.
• Consensus-based tooth count by examiners.
• Comparison of ChatGPT outputs with human evaluations.

Main Results:

• ChatGPT 4.5, o3, and o4-mini-high demonstrated substantial agreement (κ = 0.65-0.69) with clinicians for tooth counts.
• ChatGPT o1 showed moderate agreement (κ = 0.52) for tooth counts.
• All ChatGPT models consistently overestimated RBH values, with ChatGPT 4.5 showing the lowest mean bias (+11-12 pp).

Conclusions:

• ChatGPT 4.5 and o3 show promise for tooth counting on panoramic radiographs but lack clinical accuracy.
• Current ChatGPT models significantly overestimate RBH, rendering them unsuitable for periodontitis classification.
• Further development is needed before AI models can be reliably used in periodontal diagnostics.