Lam Quang Tuan, Le Minh Huu Nhat, Lee I-Ta, Le Nguyen Quoc Khanh
School of Dentistry, College of Oral Medicine, Taipei Medical University, Taipei, Taiwan.
AIBioMed Research Group, Taipei Medical University, Taipei, Taiwan.
Evid Based Dent. 2025 Jul 23. doi: 10.1038/s41432-025-01180-1.
Recent advancements in the You Only Look Once (YOLO) algorithm show promise for dental caries diagnosis. We aimed to evaluate the diagnostic performance of different YOLO versions using photographic and radiographic images for caries detection.
We searched PubMed (MEDLINE), EMBASE, Web of Science, and Scopus for studies up to December 12, 2024. Studies using any YOLO version for caries detection were included. Binary diagnostic accuracy data were extracted to calculate pooled sensitivity, specificity, and area under the curve (AUC) using a bivariate random-effects model. Quality was assessed with QUADAS-2 and the Radiomics Quality Score (RQS). This review is registered in PROSPERO (CRD42024615440).
We included 15 studies in the systematic review and 14 in the meta-analysis. Overall, YOLO-based models achieved a pooled sensitivity of 79.3% and specificity of 84.9%, with an AUC of 0.832. YOLO using radiographic images demonstrated higher specificity (92.5% vs 72.0%) and AUC (0.847 vs 0.735) than using photographic images, while sensitivity was similar (78.6% vs 80.0%). Differences between YOLO versions (v5 and earlier vs v6 and later) and the use of external validation did not significantly affect diagnostic accuracy.
Radiograph-based YOLO models showed superior specificity to photograph-based models, reflecting the higher diagnostic detail of radiographs. However, photographic approaches are completely radiation-free and more accessible, which could benefit screening in low-resource settings. Newer YOLO versions did not significantly outperform older versions, likely due to the limited complexity of the task and dataset constraints in current studies.
YOLO algorithms provide a reliable tool for dental caries detection. Radiograph imaging combined with YOLO offers enhanced diagnostic specificity, while even older YOLO versions remain effective for caries detection in practice.
“你只看一次”（YOLO）算法的最新进展显示出在龋齿诊断方面的前景。我们旨在评估不同版本的YOLO算法使用照片和X光图像进行龋齿检测的诊断性能。
我们检索了截至2024年12月12日的PubMed（MEDLINE）、EMBASE、Web of Science和Scopus数据库中的研究。纳入使用任何版本的YOLO算法进行龋齿检测的研究。提取二元诊断准确性数据，使用双变量随机效应模型计算合并敏感性、特异性和曲线下面积（AUC）。使用QUADAS-2和放射组学质量评分（RQS）评估质量。本综述已在PROSPERO（CRD42024615440）注册。
我们在系统评价中纳入了15项研究，在荟萃分析中纳入了14项研究。总体而言，基于YOLO的模型合并敏感性为79.3%，特异性为84.9%，AUC为0.832。与使用照片图像相比，使用X光图像的YOLO算法显示出更高的特异性（92.5%对72.0%）和AUC（0.847对0.735），而敏感性相似（78.6%对80.0%）。YOLO不同版本（v5及更早版本与v6及更高版本）之间的差异以及外部验证并未显著影响诊断准确性。
基于X光片的YOLO模型比基于照片的模型具有更高的特异性，这反映了X光片更高的诊断细节。然而，摄影方法完全无辐射且更易于获取，这可能有利于在资源匮乏地区进行筛查。较新的YOLO版本并没有显著优于旧版本，这可能是由于当前研究中任务的复杂性有限和数据集的限制。
YOLO算法为龋齿检测提供了一种可靠的工具。X光成像与YOLO相结合可提高诊断特异性，而即使是较旧版本的YOLO算法在实际龋齿检测中仍然有效。
Zotero 插件