不同层数的Mask-RCNN模型自动检测成釉细胞瘤效能的比较  

Comparison of Efficacy of Mask-RCNN Models with Different Layers in Automatic Detection of Ameloblastoma

作　　者：赖丹琳 许亮[1] 倪涧钊 朱小峰[1] 黄晓红[1] LAI Danlin;XU Liang;NI Jianzhao;ZHU Xiaofeng;HUANG Xiaohong(Center of Stomatology,the First Affiliated Hospital,Fujian Medical University,Center of Stomatology,National Regional Medical Center,Bin hai Campus of the First Affiliated Hospital,Fujian Medical University,Fuzhou 350212,China)

机构地区：[1]福建医科大学附属第一医院口腔医学中心,福建医科大学附属第一医院滨海院区国家区域医疗中心口腔医学中心,福建福州350212

出　　处：《口腔医学研究》2023年第12期1092-1096,共5页Journal of Oral Science Research

基　　金：福建省教育厅中青年教师教育科研项目(编号:JAT200161)。

摘　　要：目的:基于人工智能深度学习,比较Mask-RCNN不同层数的ResNet模型在成釉细胞瘤CT图像检测中的效能。方法:回顾性收集2018年4月~2020年8月福建医科大学附属第一医院成釉细胞瘤患者的CT影像数据,按照标准将79名患者纳入研究。经过预处理后,共得到3566张图像,按照8∶1∶1的比例将其随机分为训练集、验证集以及测试集。采用ResNet-18、ResNet-50和ResNet-101模型进行训练,实现肿瘤的自动检测并通过Dice系数、平均精确度AP及检测时间等评价指标进行分析。结果:与ResNet-18及ResNet-50相比,ResNet-101模型自动检测的效果最好,其Dice系数为0.87,平均精确度AP(IOU 0.50∶0.95)为0.74。但该模型所需的检测时间最长,需要0.33 s。结论:不同层数的Mask-RCNN模型均可较好地实现对成釉细胞瘤的自动检测诊断,其中ResNet-101检测效果最好,但相应地需要更长的时间。Objective:To compare the performance of different layers of ResNet models in CT image detection of Ameloblastoma(AME),which uses artificial intelligence deep learning with Mask-RCNN.Methods:A retrospective analysis was carried out on CT image data of AME patients at the First Affiliated Hospital of Fujian Medical University from April 2018 to August 2020.In the study,79 patients were included based on specific criteria.After preprocessing,we obtained 3566 images,which were divided into training,validation,and test sets in an 8∶1∶1 ratio randomly.After ResNet-18,ResNet-50,and ResNet-101 models were employed for training,automatic tumor detection was achieved,and metrics such as Dice coefficient,average accuracy,and detection time were evaluated.Results:Compared to ResNet-18 and ResNet-50,the ResNet-101 model exhibited the best performance in automatic detection,with a Dice coefficient of 0.87 and an average accuracy AP(IOU 0.50∶0.95)of 0.74.However,this model also had the longest detection time,approximately 0.33 seconds.Conclusion:The Mask-RCNN model with different layers can realize the automatic detection of AME.Among the models studied,ResNet-101 demonstrated the most favorable detection performance,though a longer detection time.

关 键 词：成釉细胞瘤 深度学习 Mask-RCNN ResNet 

分 类 号：R739.8[医药卫生—肿瘤]