A Multinational Study of Orientation-Aware Deep Learning for Automated Detection of Retained Roots and Periodontal Bone Loss

Introduction and aims Accurate radiographic assessment of retained roots (RR) and periodontally compromised teeth (PCT) are critical for dental diagnosis. This study aimed to develop and evaluate deep learning models for the automated detection and classification of these pathologies in panoramic radiographs. Specifically, we investigated the efficacy of Oriented Bounding Box (OBB) versus Axis-Aligned Bounding Box (AABB) annotation strategies in capturing the anatomical angulation of dental structures across a diverse, multi-national dataset. Methods A dataset comprising 4768 panoramic radiographs from multi-national sources was annotated into seven clinically distinct classes using both OBB and AABB protocols. Two state-of-the-art architectures, You Only Look Once v11 (YOLOv11) and Real-Time Detection Transformer (RT-DETR), were trained and evaluated under identical hyperparameters. Model performance was assessed using mean Average Precision (mAP) and loss metrics. Qualitative analysis was performed to evaluate localization fidelity in complex scenarios, such as overlapping roots or severe crowding. Results OBB annotation models consistently outperformed AABB counterparts, demonstrating superior precision in isolating obliquely oriented and overlapping structures. Among all architectures, the YOLOv11-OBB-Small model provided the best trade-off between detection performance and computational efficiency, achieving scores of 85.3% on the validation set and 83.3% on the test set. While transformer-based RT-DETR models exhibited competitive accuracy, they incurred higher computational costs during training. Qualitative assessment confirmed that OBB reduced background noise inclusion and improved boundary delineation compared to AABB. Conclusions OBB annotations improved the anatomical representation of RR and PCT compared to AABB, leading to consistent gains in detection performance. The YOLOv11-OBB-Small model achieved the best trade-off in accuracy, computational efficiency, and real-time inference capability among the evaluated models. Clinical relevance Higher diagnostic accuracy and low latency performance of the YOLOv11-OBB-Small model can support real-time clinical use. Orientation-sensitive algorithms may enhance consistency, reliability, and efficiency in automated dental diagnostics in routine practice.