Performance of artificial intelligence and human observers in the identification of dental implant systems: The effect of deep learning algorithms

Objective: This study aims to comparatively evaluate the effectiveness of artificial intelligence (AI) systems used to identify dental implant systems (DIS) and investigate their effect on human observers' performance in identifying DIS. Materials and Methods: This study retrospectively screened panoramic radiographs of patients who underwent implant treatment at Aydın Adnan Menderes University, Faculty of Dentistry, between January 2014 and April 2024. A total of 380 panoramic radiographs, encompassing 1143 implant images representing five DIS (Nucleoss T4, Nucleoss T6, Dentium Superline, Nobel Replace Tapered, and NTA Regular) that met inclusion criteria, were included in the study. The included radiographs were annotated using the Roboflow Annotate tool (Roboflow, Inc., Des Moines, Iowa, USA) and subsequently underwent preprocessing and data augmentation. Following augmentation, the final dataset consisted of 810 images split into 80% for training, 10% for validation, and 10% for testing. For DIS identification, the AI-based object detection algorithms YOLOv8l and YOLOv10l were employed. Confusion matrices, sensitivity, precision, F1 score, and mean Average Precision (mAP) values were used to assess the algorithms' performance. For the human observer group, 12 research assistants and faculty members with a minimum of one-year experience in dental implant treatment were randomly selected. Multiple-choice forms, prepared with and without AI assistance, were completed by the observers. Confusion matrices, sensitivity, precision, and F1 scores were calculated to evaluate the observers' performance in identifying DIS. True Positive (TP) rates were used to compare overall performance among the observers. Statistical analyses were performed using the Shapiro-Wilk test and Wilcoxon Signed Rank Test. Results: Sensitivity, precision, F1 score, and mAP values for YOLOv8 were 0.94, 0.94, 0.94, and 0.96, respectively, while YOLOv10 showed values of 0.95, 0.94, 0.95, and 0.97. Although overall accuracy values were similar, the YOLOv10 algorithm demonstrated a narrower range of variation. Except for the sensitivity values of the Nucleoss T4 group, AI provided statistically significant improvements across all DIS groups in all metrics analyzed (p<0.05). The overall performance of human observer groups in identifying DIS was observed to be lower than that of the AI systems. Conclusions: Deep learning algorithms demonstrated high accuracy in identifying DIS. The use of AI significantly enhanced human observers' performance in identifying DIS, indicating that AI systems can assist clinicians in identifying DIS. Keywords: Dental Implant, Artificial Intelligence, Deep Learning, Human Observer