Pediatric Dental Panoramic Image Segmentation Using Conditional Generative Adversarial Networks

Dental caries in children requires early detection and management to prevent permanent damage and support optimal development. This study develops a Conditional Generative Adversarial Network (CGAN) model based on Residual U-Net for the semantic segmentation of pediatric dental Cone Beam Computed Tomography (CBCT) panoramic images. The dataset consists of 193 pairs of two-dimensional images with a resolution of 2000x942 pixels, obtained from patients aged 2–13 years, and annotated by expert dentists. Images were pre-processed via cropping, normalization using CLAHE, and random data splitting (70% training, 20% validation, 10% testing). Data augmentation was applied to the images and masks using the Albumentations library to increase diversity and prevent overfitting. The CGAN model was trained for 10,000 epochs, optimized using RMSprop with a learning rate of 5e-4 for the generator and 1e-3 for the discriminator. The combined loss function incorporated Binary Crossentropy and Mean Absolute Error with a 1:100 weight ratio. Evaluation on the test data employed Intersection over Union (IoU) and Dice score metrics. The IoU score reached 97.07%, while the Dice score was 98.51%, indicating exceptionally high segmentation performance. These results demonstrate that the CGAN architecture can produce precise dental masks on medical images. The proposed model has the potential to support early diagnosis by pediatric dentists, accelerate the image evaluation process, and serve as a foundation for developing artificial intelligence-based diagnostic systems in dentistry.