Transfer Learning-Based Classification of Dental Implant Types from Panoramic Radiographic Images

Accurate recognition of dental implant type from panoramic radiographs is relevant for prosthetic planning, follow-up assessment, implant maintenance, and retrospective clinical documentation. Manual identification may be difficult when implant borders are partially visible, when metal artifacts obscure relevant structures, or when different implant types show overlapping radiographic appearances. This study investigated a transfer learning-based strategy for four-class dental implant classification using panoramic radiographic images. The dataset included 5,107 images from endosteal, subperiosteal, transosteal, and zygomatic implant categories and was divided into 3,573 training, 764 validation, and 770 test images using a 70/15/15 protocol. Four pretrained convolutional models, ConvNeXt-Base, ResNet-101, EfficientNet-B5, and Xception, were trained under the same experimental setting and evaluated using accuracy, macro precision, macro recall, macro F1-score, AUROC, computational complexity, and inference speed. EfficientNet-B5 achieved the highest test performance, with 98.70% accuracy, 97.94% macro precision, 97.88% macro recall, 97.88% macro F1-score, and 99.98% macro AUROC. Confusion matrix analysis showed that errors mainly occurred among endosteal, subperiosteal, and transosteal implants, while all zygomatic test images were correctly classified. The findings support the potential of transfer learning for implant type recognition from panoramic radiographs, but external validation and leakage-safe patient-level evaluation remain necessary before clinical deployment.