基于深度学习的曲面体层片颌骨病变辅助诊断技术研究  

Research on deep learning assisted diagnosis technology of jaw lesions using panoramic radiographs

作　　者：高歌 刘畅 曾梦雨 彭俊杰 郭际香[2] 汤炜[1] GAO Ge;LIU Chang;ZENG Mengyu;PENG Junjie;GUO Jixiang;TANG Wei(State Key Laboratory of Oral Diseases&National Center for Stomatology&National Clinical Research Center for Oral Diseases&West China Hospital of Stomatology,Sichuan University,Chengdu 610041,China;Machine Intelligence Laboratory,College of Computer Science,Sichuan University,Chengdu 610065,China)

机构地区：[1]口腔疾病防治全国重点实验室、国家口腔医学中心、国家口腔疾病临床医学研究中心、四川大学华西口腔医院,四川成都610041 [2]四川大学计算机学院机器智能实验室,四川成都610065

出　　处：《口腔疾病防治》2024年第10期789-796,共8页Journal of Prevention and Treatment for Stomatological Diseases

基　　金：四川省自然科学基金(2024NSFSC0659);四川大学华西口腔医学院探索与研究项目(RD-01-202304)。

摘　　要：目的 探讨深度学习应用于曲面体层片辅助诊断颌骨透射病变、颌骨阻射病变的效果,以减少漏诊,辅助医生早期筛查、提高诊断准确性。方法 本研究通过四川大学华西口腔医院伦理委员会批准。以443例曲面体层片为研究对象,构建YOLO v8m-p2神经网络模型,将标注后的图像分为训练集354例,验证集45例和测试集44例,用于模型训练、验证和测试。采用精确率、召回率、F-1分值、G分值、mAP50评价模型的检测性能。结果 443例曲面体层片涵盖颌骨常见的良性病变,其中颌骨透射病变数量为318,包括含牙囊肿、牙源性角化囊肿、成釉细胞瘤3类病变;颌骨阻射病变数量为145,包含特发性骨硬化、牙瘤、牙骨质瘤、牙骨质-骨结构不良4类病变,样本有良好的代表性。YOLO v8m-p2神经网络模型识别颌骨病变的性能:精确率为0.887,召回率为0.860,F-1分值为0.873,G分值为0.873,mAP50为0.863。其中,含牙囊肿、牙源性角化囊肿、成釉细胞瘤召回率分别为0.833、0.941、0.875。结论 YOLO v8m-p2神经网络模型应用于初步检测口腔曲面体层片中的颌骨透射病变及颌骨阻射病变以及多分类检测颌骨透射病变时诊断性能表现良好,可辅助医生筛查曲面体层片的颌骨疾病。Objective To study the effect of deep learning applied to the assisted diagnosis of radiolucent lesions and radiopaque lesions of the jaws in panoramic radiography and to reduce the missed diagnosis,with early screening to assist doctors to improve the diagnostic accuracy.Methods This study was approved by the Ethics Committee of the West China Stomatological Hospital of Sichuan University.The YOLO v8m-p2 neural network model was constructed with 443 panoramic images as a subject to read.The labeled images were divided into 354 training sets,45 verification sets,and 44 test sets,which were used for model training,verification,and testing.Accuracy,recall,F-1 score,G score,and mAP50 were used to evaluate the detection performance of the model.Results 443 panoramic images covered the common benign lesions of the jaw,the number of radiolucent lesions of the jaw was 318,containing dentigerous cyst,odontogenic keratocyst,and ameloblastoma.The number of radiopaque lesions was 145,containing idiopathic osteoscle-rosis,odontoma,cementoma,and cemento-osseous dysplasia;the samples are well representative.The accuracy of the YOLO v8m-p2 neural network model in identifying jaw lesions was 0.887,and the recall,F-1 score,G score,and mAP50 were 0.860,0.873,0.873,and 0.863,respectively.The recall rates of dentigerous cyst,odontogenic keratocyst,and ameloblastoma were 0.833,0.941,and 0.875,respectively.Conclusion YOLO v8m-p2 neural network model has good diagnostic performance in preliminary detection of radiolucent and radiopaque lesions of the jaws in panoramic radiography and multi-classification monitoring of radiolucent lesions of jaws,which can assist doctors to screen jaw dis-eases in panoramic radiography.

关 键 词：颌骨囊肿 颌骨肿瘤 影像诊断 曲面体层片 人工智能 深度学习 目标检测 YOLO v8m 神经网络模型 

分 类 号：R78[医药卫生—口腔医学]