Comparative Analysis of GAN-based Segmentation Models on Dental Panoramic Radiography Dataset

Abstract

Especially in criminal investigations, the identification of the victim is essential. The branch of forensic medicine that uses the method of identification from the teeth of the victims is called forensic odontology. In forensic odontology, physical information about the individual can be obtained from the bone and enamel structure of the teeth. Panoramic, periapical, and cephalometric imaging techniques are the most commonly used in the odontological identification of the individual. Forensic odontology is increasingly recognized for its essential role in personal identification during mass disasters, sexual assault cases, and child abuse investigations. Deep learning algorithms have recently successfully detected dental disorders such as caries, periodontal bone loss, and apical lesions. Generative adversarial networks (GAN) models have mainly achieved high segmentation performance in medical images. In this study, GAN models were designed and comparatively analyzed using U-Net, Volumetric convolutional neural network (V-Net), spatial and channel Squeeze-Excitation-based U-Net(scSEU-Net), Transformer-based U-Net (TransU-Net), and U-Net like pure Transformer (SwinU-Net) segmentation architectures which are widely used in the literature as generators. As a result of the comparative analyses, scSEU-Net-based GAN achieved the highest performance values with 0.8826 Thresholded Dice(DSC), 0.7901 Thresholded Intersection over Union (Thresh-IoU), 0.9805 Accuracy (ACC), 0.9268 Precision (PREC), and 0.9001 Recall (REC).