Comparative Evaluation of Vision Transformer Architectures for Dental Implant Type Classification in Panoramic Radiographs

Abstract

Accurate identification of dental implant types on panoramic radiographs can support treatment planning, longitudinal follow-up, and prosthetic decision-making when clinical records are incomplete or unavailable. This study comparatively evaluated transformer-based deep learning architectures for the automated classification of endosteal, subperiosteal, transosteal, and zygomatic implants on panoramic dental radiographs. A dataset comprising 5,107 images was divided into training (4,084 images), validation (510 images), and independent test (513 images) subsets. ViT-Base, Swin-Base, PiT-B, DeiT-Base, and ConViT-Base were trained under an identical experimental protocol and compared in terms of accuracy, precision, recall, F1-score, parameter count, and computational complexity. PiT-B achieved the strongest overall performance, with an accuracy of 0.9864, precision of 0.9714, recall of 0.9821, and an F1-score of 0.9764. Confusion-matrix and receiver operating characteristic analyses showed complete test-set classification of the zygomatic and subperiosteal categories, whereas the limited errors were concentrated between endosteal and transosteal implants and between transosteal and subperiosteal implants. PiT-B also exhibited the lowest computational burden among the evaluated architectures, with 72.74 million parameters and 211.067 GFLOPs. These findings indicate that transformer-based models can learn highly discriminative representations of implant morphology and anatomical context from panoramic radiographs, with PiT-B providing the most favorable balance between predictive performance and computational efficiency.