Developing Synthetic Orthopantomogram Datasets Through Generative Models.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Generative Models for Healthcare Data

Background:

• Generating high-quality synthetic medical images is crucial for AI model training.
• Orthopantomograms (OPGs) are essential for dental diagnostics but require diverse datasets.

Purpose of the Study:

• To generate synthetic OPGs using generative AI.
• To evaluate the realism of synthetic OPGs using human and AI assessments.

Main Methods:

• Trained a Denoising Diffusion Probabilistic Model (DDPM) on 5,383 OPGs.
• Evaluated model performance using Fréchet Inception Distance (FID).
• Assessed image realism with dentists and AI models (MesoNet, ViT).

Main Results:

• DDPM achieved a superior FID score (26.90) compared to GANs (118.49).
• Dentists had low accuracy in distinguishing real from synthetic images (low AUC).
• AI models achieved perfect classification accuracy in differentiating images.

Conclusions:

• Generated synthetic OPGs exhibit high realism, indistinguishable from real images by experts.
• Diffusion models are promising for creating annotated synthetic datasets for AI in diagnostics.