A Novel Hybrid Large Language Model Approach for Reporting Panoramic Radiographs and Performance Comparison with...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Dental Radiology
• Machine Learning Applications

Background:

• Current multimodal AI systems struggle with panoramic radiograph interpretation due to low diagnostic accuracy and high hallucination rates.
• Large language models (LLMs) show promise for clinical reporting but require integration with robust image analysis for reliability.

Purpose of the Study:

• To develop and evaluate a hybrid AI framework integrating deep learning image analysis with LLM-driven reporting for enhanced reliability in dental radiology.
• To assess the performance of this framework in terms of reporting accuracy, structural validity, consistency, and hallucination rates.

Main Methods:

• A YOLOv12 model was trained on 30,954 panoramic radiographs for tooth detection and 14-category segmentation.
• Detection and segmentation outputs were converted to structured JSON data for processing by locally hosted LLMs (DeepSeek R1, Mistral, Llama 3.2, Gemma 3, Qwen3, SmolLM3).
• Reporting accuracy was evaluated on 50 unseen radiographs, comparing AI outputs against expert assessments.

Main Results:

• The segmentation model achieved an overall F1-score of 0.708, with high performance for ectopic/supernumerary teeth (0.994), impacted teeth (0.990), and implants (0.984).
• All hybrid LLMs generated structurally valid JSON outputs (100%). DeepSeek R1 demonstrated the highest reporting accuracy (466 true findings) and lowest hallucinations (30).
• Commercial LLMs exhibited hallucinations in 100% of reports, whereas the hybrid framework significantly minimized these errors.

Conclusions:

• Integrating structured image-derived findings with LLM reasoning markedly improves reporting accuracy and minimizes hallucinations in dental radiology.
• The developed hybrid framework offers a promising and reliable solution for AI-assisted interpretation of dental radiographs, outperforming current commercial LLMs.