A Comparison ofshcU-Net Based GAN and U-net Based GAN in Adult Dental Segmentation

Abstract

Teeth are among the most diverse organs in vertebrates, exhibiting significant morphological and functional variation. Tooth segmentation is a specialized area within dental imaging and digital dentistry that focuses on accurately delineating individual teeth in various imaging modalities, such as Cone Beam Computed Tomography (CBCT) scans. Teeth segmentation and classification are critical processes in dental imaging, accurately delineating individual teeth and classifying each image pixel into objects of interest. This task is essential for various dental applications, such as diagnosis and treatment planning. Tooth segmentation plays a critical role in forensic identification and analysis processes; therefore, accurate and efficient segmentation methods are needed. Forensic odontology is pivotal in personal identification during mass calamities, sexual assault cases, and child abuse investigations, often relying on dental remains when other forms of evidence are unavailable. This study compares the performance of U-Net-based GAN with semi-hybrid convolution (shcU-Net) and traditional U-Net-based GAN models for teeth segmentation in digital forensics. In this study, the proposed shcU-Net-based GAN model uses only depth-separable convolution in the decoder layer, while it offers a hybrid structure in which depth-separable convolution and hybrid convolution are used together in the decoder layer. Although shcU-Net-based GAN used half the parameters compared to U-Net-based GAN architecture, it achieved almost the same performance. While U-Net-based GAN achieved 0.99 validation accuracy and validation precision performance values, shcU-Net-based GAN achieved 0.98 validation accuracy and 0.92 validation precision values, respectively. © 2025 Elsevier B.V., All rights reserved.