Abstract
Background/Objectives: This study internally evaluated deep learning models for classifying mandibular condylar position on sagittal cone-beam computed tomography (CBCT) images using a predefined joint-space-based reference standard. Methods: In this single-center retrospective study, 1040 temporomandibular joint images from 560 patients were categorized as anterior, centric, or posterior using Pullinger and Hollender’s method. Patients were allocated to training, validation, and test sets in an 80:10:10 ratio (448, 56, and 56 patients, respectively), yielding 823, 107, and 110 joint images. Three object-detection models (YOLOv8x, YOLOv8n, and RF-DETR Nano) were evaluated on the same test set for condylar region localization and position classification. Performance was evaluated using accuracy, precision, recall, F1-score, and mean average precision (mAP). Results: YOLOv8x correctly classified 97 of 110 test images, achieving an accuracy of 0.88 (patient-cluster bootstrap 95% confidence interval, 0.81–0.95) and an mAP@0.5 of 0.89. YOLOv8n and RF-DETR Nano correctly classified 84 and 82 images, with accuracies of 0.76 and 0.75 and mAP@0.5 values of 0.80 and 0.81, respectively. The mAP@0.5:0.95 values for YOLOv8n and RF-DETR Nano were 0.60 and 0.59, respectively. YOLOv8x’s mAP@0.5:0.95 was 0.66. For YOLOv8x, recall/F1-score were highest for anterior class (0.94/0.91) and lowest for centric class (0.79/0.83), with precision ranging from 0.88 to 0.89 across classes. Intra- and inter-observer classification agreement in 156 joints was κ = 0.90 and 0.97, respectively. Conclusions: The YOLOv8x-based model achieved the highest overall accuracy among the three models in this comparison, indicating that it can classify mandibular condylar position on sagittal CBCT images according to predefined categories.