Tooth Image Synthesis Using Customized GAN Architectures: A Deep Learning-Based Solution for Data Scarcity in Dentistry

The scarcity of high-quality, annotated dental radiographic data poses a significant challenge to the advancement of Artificial Intelligence (AI) applications in dentistry. This study proposes a customized Generative Adversarial Network (GAN) framework for the generation of anatomically realistic synthetic dental images, focusing on individual tooth-level synthesis from panoramic radiographs. Utilizing a dataset of 612 annotated images categorized according to the FDI tooth notation system, the model was trained separately on cropped single-tooth images and full panoramic views. Architectural innovations such as residual blocks, self-attention modules, and an auxiliary classifier were incorporated to enhance image fidelity and stability. Quantitative evaluation using Fréchet Inception Distance (FID) demonstrated superior image quality in single-tooth generation compared to panoramic synthesis, with the best-performing classes achieving FID scores below 100. These results validate the model's potential to augment dental datasets for downstream AI applications, while promoting ethical data practices by reducing the dependency on clinical image acquisition.