Developing Synthetic Orthopantomogram Datasets Through Generative Models

OBJECTIVE: To generate synthetic Orthopantomograms (OPGs) using generative Artificial Intelligence (AI) models and to evaluate the realism of synthetic images through assessments by humans as well as AI models. METHODOLOGY: This study conducted at Aga Khan University Hospital (AKUH) aimed to generate synthetic OPGs using anonymized images from the existing database at AKUH dental clinics. Additionally, OPGs from the Tufts Dental Dataset and an open-sourced dataset from Kaggle were incorporated to enhance image diversity. After resizing all images to 512 × 512 pixels, a total of 5,383 OPGs were utilized for AI model training. Generative Adversarial Networks (GANs) were initially employed but yielded poor results. Subsequently, a Denoising Diffusion Probabilistic Model (DDPM) was trained on Google Colab, generating 2,500 synthetic images. The model's performance was evaluated using the Fréchet Inception Distance (FID). To assess image quality, a set of 20 synthetic and 20 original images was examined by dentists and two AI models; MesoNet (MN) and Vision Transformer (ViT). RESULTS: The DDPM achieved an FID score of 26.90, markedly superior to 118.49 achieved by GANs. Dental experts exhibited suboptimal performance in distinguishing real from synthetic images, as evidenced by low Area Under the Curve (AUC) scores, indicating the high realism of the DDPM generated images. In contrast, both MN and ViT models achieved perfect classification accuracy with high AUC scores. GradCAM was the explainable AI technique applied to MN to elucidate AI performance. CONCLUSIONS: Inability to differentiate synthetic OPGs highlights the realism of the generated images, suggesting their high quality. The authors propose using diffusion models to create annotated synthetic datasets for diverse AI training in healthcare diagnostics.