基于无先验解耦网络模型的下颌第三磨牙检测应用研究

Detection of mandibular third molars based on none-priori decoupling network model

目的:探讨基于无先验解耦网络模型在曲面断层片下颌第三磨牙(智齿)自动识别、影像学分类中的应用,帮助医患双方实现准确高效的实时诊断。方法:回顾2020年6月—12月于苏州大学附属第三医院就诊的患者曲面断层片,标注(Winter分类法)后,一致性资料973例。以YOLO模型为基础,引入不依赖标注框先验信息的方法简化模型,同时采用检测头解耦的优化思想,建立无先验解耦网络模型。有效数据集分为训练集和测试集进行训练推理,采用PASCAL VOC和COCO数据集评判准则进行统计分析。结果:人工智能模型对智齿目标的检测,漏检率为0%;对临床常见的智齿类型,VOC mAP达90%以上,COCO mAP达50%以上。结论:基于YOLO无先验解耦网络模型在曲面断层片下颌智齿的识别分类上表现出良好解读诊断效果。

Objective To explore the application of a none-priori decoupling network model in automatic identification and imaging classification of mandibular third molars(wisdom teeth)based on curved tomograms,so as to help doctors and patients achieve accurate and efficient real-time diagnosis.Methods The panoramic radiographs of the patients treated in the Third Affiliated Hospital of Soochow University in the past six months were reviewed and labeled(according to Winter classification),and 973 cases of consistent data were obtained.Based on the YOLO model,a new method was introduced to simplify the model,which does not rely on the priori information of the labeled frame,while the optimization idea of decoupling the detection head was adopted to build a priori-free decoupling network model.The valid data sets were divided into the training set and the test set for training reasoning,and PASCAL VOC and COCO data set evaluation criteria were adopted for statistical analysis.Results The missing rate of AI model detecting the wisdom tooth targets was 0%.For the common types of wisdom teeth in clinic,VOC mAP was more than 90%and COCO mAP was more than 50%.Conclusion The YOLO-based none-priori decoupling network model shows a good interpretation and diagnosis effect in the recognition and classification of mandibular wisdom teeth on curved tomographic slices.