An XAI-Driven Hybrid YOLOv8-LRP Architecture for the Detection of Impacted Mandibular Third Molars in Panoramic Radiography

The accurate detection of impacted mandibular third molars (i-mTM) in panoramic radiographs (PRs) is of vital importance for diagnosis and treatment planning. In this study, a hybrid YOLOv8-LRP architecture was proposed to enhance the interpretability of predictions using deep learning (DL)based YOLOv8 architectures for i-mTM detection, by integrating Layer-wise Relevance Propagation (LRP), an Explainable Artificial Intelligence (XAI) technique. A new dataset was created by labeling 608 PRs obtained from the public m-TM dataset. The performance of the proposed architecture was evaluated on this new dataset using a comprehensive ablation test. The optimized YOLOv8x architecture achieved superior performance metrics such as 99.59% precision, 100% recall, 99.80% F1 Score, and 99.50% mAP50. Heatmaps obtained from XAI analysis performed with LRP demonstrated that the proposed method provides explainability appropriate for the problem. In a performance comparison with state-of-the-art methods in the literature, the proposed hybrid method demonstrated the best performance across all metrics. These results clearly demonstrate that it offers a powerful and transparent tool for diagnostic decision support systems focused on detecting i-mTM in dentistry.