摘要
甲状腺超声图像由于对比度低、边缘不清晰、高噪声和周围组织复杂难辨等问题,给医生诊断甲状腺疾病造成困难。针对此问题,采用Cascade Rcnn目标检测算法,分别以ResNet50、Resnet101以及融合压缩激励注意力模块SE-ResNet50、SE-ReNet101为主干网络,对从某三甲医院获取的1513例(其中良性结节832例,恶性结节681例)甲状腺超声图像,在专业超声科医生的指导下进行预处理,制作本次实验使用的标准coco格式数据集。采用迁移学习的方式将从Imagenet大型数据库上预训练得到的权重迁移到本次实验模型结构中,经过4个主干网络的实验结果对比,以SE-ResNet101为主干网络的Cascade Rcnn算法,在结节定位和判别方面,实现了精确率92.4%,召回率86.2%,特异性95.1%,F1值89.2%,mAP值82.4%的检测效果,对辅助医生进行甲状腺超声图像的诊断具有一定的临床指导意义。
Thyroid ultrasound images have low contrast,unclear edges,high noise,and the surrounding tissues are complex and difficult to distinguish,making it extremely difficult for doctors to diagnose thyroid diseases.To overcome this problem,Cascade Rcnn target detection algorithm was used in this work,with ResNet50,Resnet101 and fusion compression incentive attention modules SE-ResNet50,SE-ReNet101 as the backbone network.There were 1513 cases thyroid ultrasound images(including 832 cases benign nodules and 681 cases malignant nodules)obtained from a third-class hospital.Under the guidance of professional sonographers,the data were preprocessed into the standard coco format data set.The weights obtained from the pre-training of the large Imagenet database by transfer learning were migrated to this experimental model structure.Comparing with the experimental results of the four backbone networks,Cascade Rcnn algorithm with SE-ResNet101 as the backbone network achieved an accuracy of 92.4%,recall rate of 86.2%,specificity of 95.1%1,F1 value of 89.22%,and mAP value of 82.4%.The detection result of nodule localization and classification of benign and malignant was of clinical guiding significance for assisting doctors in the diagnosis of thyroid ultrasound images.
作者
章浩伟
李占齐
刘颖
李淼
Zhang Haowei;Li Zhanqi;Liu Ying;Li Miao(School of Health Science and Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Zhengzhou Yihe Hospital of Henan Province,Zhengzhou 450000,China)
出处
《中国生物医学工程学报》
CAS
CSCD
北大核心
2022年第1期64-72,共9页
Chinese Journal of Biomedical Engineering
基金
上海市科委医学引导项目(12401907700)
微创励志创业基金(183852272)。
