<![CDATA[The early detection of dental diseases is essential for preventing severe oral health complications. However, automated lesion detection utilizing intraoral images remains highly challenging due to severe tooth overlap, occlusion, and visually similar anatomical structures. Under these complex conditions, conventional single-stage object detectors frequently produce redundant and inaccurate bounding boxes, which significantly degrades localization precision. To explicitly resolve this problem, this study proposes a robust multi-scale ensemble learning strategy that integrates bounding box predictions from YOLOv5 and YOLOv8 through a Weighted Boxes Fusion (WBF) mechanism. Unlike traditional post-processing techniques such as Non-Maximum Suppression (NMS) and Soft-NMS, the proposed method fuses overlapping bounding boxes by leveraging confidence-weighted spatial aggregation, thereby preserving critical detection information. Extensive experiments were conducted on a publicly validated intraoral image dataset comprising four distinct clinical classes: caries, cavity, cracks, and normal teeth. Quantitative evaluations demonstrate that the proposed WBF ensemble approach substantially outperforms single- model baselines. The integrated model achieves a mean Average Precision (mAP@0.5) of 66.14%, a Precision of 66.47%, and an Intersection over Union (IoU) of 90.83%, representing a massive improvement over the baseline mAP values of approximately 36 to 37%. Furthermore, rigorous statistical testing validates that these performance gains are highly significant (p < 0.05). Ultimately, these findings indicate that the proposed ensemble framework provides a reliable, high-precision solution for intraoral dental lesion localization, offering substantial viability for real-world clinical diagnostic applications.]]>
