Yang Lin, Chen Guan-Yu
College of Electrical Engineering and New Energy, China Three Gorges University, Yichang, China.
Department of Experimental Orofacial Medicine, Philipps University Marburg, Marburg, Germany.
Int Dent J. 2025 Jul 22;75(5):100898. doi: 10.1016/j.identj.2025.100898.
Deep learning methods have been proven to be effective in detecting dental caries in visible light images. However, existing research involves inadequate categories and mainly focuses on local lesion areas. This study aims to use advanced deep learning models to achieve caries detection based on tooth instances (where all teeth in images are detected) and fine-grained classification according to the International Caries Detection and Assessment System (ICDAS). To address the potential instability under complex scenarios, we propose 2 correction methods that incorporate background knowledge.
A total of 1200 selected high-quality intraoral images were expanded to 8,754 images using data augmentation techniques, and each tooth inside was annotated. Three advanced models, YOLO-v8, YOLO-v9, and YOLO-NAS, were trained and tested on the dataset. In the stage of postprocessing, predicted categories were corrected with a weighted average of scores, and confidence scores were adaptively adjusted based on the spatial relationships of teeth.
The proposed methods improved the mean Average Precision (mAP) scores by 4.7% (p < .01/Mann-Whitney-U-test), 2.8% (p < .01), and 4.4% (p < .01) across the 3 models, with the highest score of 72.9% on YOLO-v8. Precision and recall increased by 3.8% and 5.6%, respectively, while FPS decreased from 83.1 to 78.1. Especially improved the scores for moderate caries and demonstrated greater robustness.
The primary objectives were achieved, and the 2 proposed correction methods bring an effective improvement to the existing algorithm framework. It's expected to promote the application of artificial intelligence and inspire further research.
This research is of clinical value due to its functional innovation: the finer classification should assist dentists formulate personalized treatment strategies. Focusing on the detailed evaluation of each tooth should help deliver better and personalised clinical care.
深度学习方法已被证明在可见光图像中检测龋齿有效。然而，现有研究涉及的类别不足，且主要集中在局部病变区域。本研究旨在使用先进的深度学习模型基于牙齿实例（即检测图像中的所有牙齿）实现龋齿检测，并根据国际龋齿检测与评估系统（ICDAS）进行细粒度分类。为解决复杂场景下的潜在不稳定性，我们提出了两种纳入背景知识的校正方法。
使用数据增强技术将总共1200张精选的高质量口腔内图像扩展到8754张图像，并对其中的每颗牙齿进行标注。在该数据集上对三种先进模型YOLO-v8、YOLO-v9和YOLO-NAS进行训练和测试。在后处理阶段，用分数的加权平均值校正预测类别，并根据牙齿的空间关系自适应调整置信度分数。
所提出的方法使三种模型的平均精度均值（mAP）分数分别提高了4.7%（p <.01/曼-惠特尼-U检验）、2.8%（p <.01）和4.4%（p <.01），在YOLO-v8上得分最高，为72.9%。精确率和召回率分别提高了3.8%和5.6%，而每秒帧数（FPS）从83.1降至78.1。尤其提高了中度龋齿的分数，并表现出更强的鲁棒性。
实现了主要目标，所提出的两种校正方法对现有算法框架带来了有效改进。有望促进人工智能的应用并激发进一步研究。
本研究因其功能创新具有临床价值：更精细的分类应有助于牙医制定个性化治疗策略。关注每颗牙齿的详细评估应有助于提供更好的个性化临床护理。
