Pano-GAN: A Deep Generative Model for Panoramic Dental Radiographs

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Dental Radiology

Background:

• Scarcity of high-quality dental imaging data hinders research and education.
• Generative Adversarial Networks (GANs) show potential for synthetic data generation.

Purpose of the Study:

• To develop and evaluate a GAN for generating synthetic dental panoramic radiographs.
• To address data limitations in dental research and education through synthetic image creation.

Main Methods:

• A deep convolutional GAN (DCGAN) with Wasserstein loss and gradient penalty (WGAN-GP) was utilized.
• Training was performed on 2322 dental panoramic radiographs, focusing on dentoalveolar regions.
• Data preprocessing involved cleaning and standardization; models varied critic iterations, features, and denoising.

Main Results:

• Generated radiographs depicted dentoalveolar structures moderately but contained artifacts.
• Expert evaluation revealed a trade-off: non-denoised data yielded better fine structures (mandibular canal, trabecular bone).
• Denoised data models provided superior overall image clarity, sharpness, and realism.

Conclusions:

• GANs can generate synthetic dental radiographs, offering a potential solution for data scarcity.
• Image quality is influenced by denoising and model parameters, requiring careful optimization.
• Further research into GAN architectures is warranted for advanced dental imaging applications.