Implant Thread Shape Classification by Placement Site from Dental Panoramic Images Using Deep Neural Networks

Abstract

Purpose: In this study, we aimed to classify an implant system by comparing the types of implant thread shapes shown on radiographs using various Convolutional Neural Networks (CNNs), particularly Xception, InceptionV3, ResNet50V2, and ResNet101V2. The accuracy of the CNN based on the implant site was compared. Materials and Methods: A total of 1000 radiographic images, consisting of eight types of implants, were preprocessed by resizing and CLAHE filtering, and then augmented. CNNs were trained and validated for implant thread shape prediction. Grad-CAM was used to visualize class activation maps (CAM) on the implant threads shown within the radiographic image. Results: Averaged over 10 validation folds, each model achieved an AUC of over 0.96: AUC of 0.961 (95% CI 0.952–0.970) with Xception, 0.973 (95% CI 0.966-0.980) with InceptionV3, 0.980 (95% CI 0.974-0.988) with ResNet50V2, and 0.983 (95% CI 0.975-0.992) with ResNet101V2. Accuracy was higher in the posterior region than in the anterior area in all four models. Most CAMs highlighted the implant surface where the threads were present; however, some showed responses in other areas. Conclusion: The CNN models accurately classified implants in all areas of the oral cavity according to the thread shape, using radiographic images.