Explainable Deep Learning Framework for Automated Dental Implant Detection

Abstract

Dental implants' classification on panoramic radiography (OPG) provides essential information regarding treatment planning, post-operative care, and clinical decision-making. However, manual assessment by experts is slow, has significant variability based on the examiner, and there are no objective methods to reliably determine which implant type is present in an OPG. In this study, we present a Fusion Deep Learning (FDL) architecture created by concatenating the last layers of three well-known convolutional neural networks (CNNs): EfficientNetB3 with ResNet50V2 and VGG19. This model was trained and validated using a dental implant dataset with 5,273 anonymized OPG images classified into four categories of implant—endosteal, subperiosteal, transosteal, and zygomatic. To provide interpretable behavior, the classifier uses three explainable AI (XAI) techniques—SHAP, LIME, and Grad-CAM. The fusion model achieved an accuracy of 97.5%, precision of 97.2%, recall of 97.5%, F1-score of 97.3%, and AUC/ROC of 99.7% on the hold-out test data, surpassing each individual backbone network by 2.4 to 4.7 percentage points. The XAI outputs indicated that the model's predictions were supported by the presence of radiographically evident clinical features associated with the implant such as contact at the bone-implant interface, thread configuration and integration with the cortical bone, rather than being related to incidental image artefacts. Combining high accuracy and interpretable predictions will provide a transparent pathway for deploying the fusion model in a clinical setting and facilitate CAD application for implant classification.