A Multi-Model Ensemble YOLO Framework for Automated Detection of Dental Pathologies in Low- Quality Panoramic Radiographs

Accurate identification of dental pathologies in panoramic radiographs is vital for effective diagnosis and treatment planning; however, manual interpretation remains time- consuming, subjective, and prone to human error. This paper presents a deep learning-based object detection framework to automate the classification of dental diseases from low-quality panoramic X-rays. To address these issues, we propose a novel multi-model ensemble pipeline based on YOLOv11n architectures, where individual models are trained for each dental pathology and combined through a pseudo-labelling process to generate a refined, high-quality dataset. A consolidated YOLOv11m model trained on this pseudo-labelled dataset demonstrated a 6.4% relative performance improvement compared to transfer learning. Further evaluation using a new, high-quality 62-class dataset indicated that models trained from scratch outperform those fine-tuned via transfer learning. The findings highlight the critical influence of data quality on model performance and provide a robust comparative analysis of training strategies. Overall, this study introduces an effective multi-model ensemble methodology and establishes a foundation for developing accurate, automated diagnostic systems for low-quality panoramic dental radiographs in digital dentistry.