基于YOLOv5s和超声图像的儿童肠套叠特征检测模型  

Child Intussusception Feature Detection Model Based on YOLOv5s and Ultrasound Images

作　　者：陈星[1] 俞凯[1,2] 袁贞明 黄坚[3,4,5] 李哲明 CHEN Xing;YU Kai;YUAN Zhenming;HUANG Jian;LI Zheming(School of Information and Technology,Hangzhou Normal University,Hangzhou 311121,China;Hangzhou Hele Tech Company,Hangzhou 311121,China;Data and Information Department,The Children's Hospital of Zhejiang University School of Medicine,Hangzhou 310051,China;Sino-Finland Joint AI Laboratory for Child Health of Zhejiang Province,Hangzhou 310051,China;AI Lab,National Clinical Research Center for Child Health,Hangzhou 310051,China)

机构地区：[1]杭州师范大学信息科学与技术学院,浙江杭州311121 [2]杭州和乐科技有限公司,浙江杭州311121 [3]浙江大学医学院附属儿童医院数据信息部,浙江杭州310051 [4]浙江-芬兰儿童健康人工智能联合实验室,浙江杭州310051 [5]国家儿童健康与疾病临床医学研究中心AI实验室,浙江杭州310051

出　　处：《杭州师范大学学报（自然科学版）》2024年第1期10-19,共10页Journal of Hangzhou Normal University(Natural Science Edition)

基　　金：国家重点研发计划项目(2019YFE0126200);国家自然科学基金面上项目(62076218);浙江省医药卫生科技计划项目(2019ZH004)。

摘　　要：为帮助医生快速寻找到儿童腹部超声中肠套叠的病变特征并实现肠套叠超声诊后数据的快速质检,文章将目标检测算法应用于儿童腹部超声图像检测肠套叠“同心圆”征.首先探索了基于YOLOv5s的儿童肠套叠检测模型,发现该模型检测肠套叠“同心圆”征的精确度、召回率、F 1分数、mAP@0.5、FPS以及参数量等方面均优于Faster RCNN.进一步,为解决肉眼难以观察的“同心圆”征的检测问题,使用双向特征金字塔网络,并将注意力机制加入YOLOv5s网络,形成基于YOLOv5s_BiFPN_SE框架的儿童肠套叠“同心圆”征检测模型.该模型检测的精确率、召回率、F 1分数、mAP@0.5分别达到了91.33%、90.73%、91.03%、88.77%,性能更优于YOLOv5s.For the purpose of helping doctors to quickly identify the lesions of intussusception in children s abdominal ultrasound and achieving the rapid quality inspection of ultrasound diagnosis data,this paper applied the target detection algorithm to detect the"concentric circle"feature of intussusception in children s abdominal ultrasound images.Firstly,a YOLOv5s based detection model for pediatric intussusception was explored,which had the improved precision,recall,F 1 score,mAP@0.5,FPS,and parameter quantity compared to Faster RCNN.Furthermore,a bidirectional feature pyramid network was used to solve the detection problem of the"concentric circle"which was difficult to observed by naked eyes.The attention mechanism was added into the YOLOv5s network to form a detection model based on YOLOv5s_BiFPN_SE framework.The accuracy,recall,F 1 score,mAP@0.5 could reach 91.33%,90.73%,91.03%,and 88.77%respectively,which represented better performance than YOLOv5s.

关 键 词：目标检测 肠套叠 超声图像 “同心圆”征 双向特征金字塔网络 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]