基于深度卷积神经网络的甲状腺超声图像良恶性结节识别方法研究  

Identification of benign and malignant nodules in thyroid ultrasound images based on deep convolutional neural network

作　　者：姚文君[1,2] 殷超然 朱宏庆 江健敏 庞小溪 孙怡宁 Yao Wenjun;Yin Chaoran;Zhu Hongqing;Jiang Jianmin;Pang Xiaoxi;Sun Yining(Dept of Radiology,The Second Affiliated Hospital of Anhui Medical University,Hefei 230601;Intelligent Mechanization Institute,Hefei Institute of Physical Matter,Chinese Academy of Sciences,Hefei 230031;Dept of Ultrasonography,Anhui No.2 Provincial People′s Hospital,Hefei 230011;School of Electronic and Information Engineering,Anhui University,Hefei 230601;Dept of Nuclear Medicine,The Second Affiliated Hospital of Anhui Medical University,Hefei 230601)

机构地区：[1]安徽医科大学第二附属医院放射科,合肥230601 [2]中国科学院合肥物质研究院智能机械研究所,合肥230031 [3]安徽省第二人民医院超声科,合肥230011 [4]安徽大学电子信息工程学院,合肥230601 [5]安徽医科大学第二附属医院核医学科,合肥230601

出　　处：《安徽医科大学学报》2023年第5期854-858,共5页Acta Universitatis Medicinalis Anhui

基　　金：安徽省自然科学基金(编号:2008085QH406);安徽医科大学第二附属医院临床研究培育计划(编号:2020LCYB05);安徽省转化医学研究院科研基金(编号:2021zhyx-C45)。

摘　　要：目的探讨深度卷积神经网络在甲状腺结节超声图像的自动检测和良恶性分类中应用价值。方法回顾性选取1012幅甲状腺结节的超声图像并对其进行标记,构建YOLOv5网络模型,精准定位甲状腺结节所在位置并自动裁减结节所在区域,同时构建GoogLeNet网络模型对裁减后结节的图像进行良恶性分类。结果在所采集的数据集中,目标检测网络对甲状腺结节位置检测的平均精确度均值为96.2%;分类网络对良恶性结节分类的敏感度为0.885,特异度为0.822,准确度为0.866,AUC值为0.92,显著高于AlexNet模型(AUC=0.81)、VGG模型(AUC=0.86)和MobileNet模型(AUC=0.76)。结论深度卷积神经网络模型对超声图像中的甲状腺良恶性结节具有较高的定位和识别能力,有助于提高影像自动诊断的准确性。Objective To explore the clinical application value of deep convoluti onal neural network for automatic detection and classification of benign and malignant thyroid nodules ultrasound images.Methods A total of 1012 ultrasound images of thyroid nodules were retrospectively selected and labeled.The YOLOv5 network model was constructed to accurately locate the location of thyroid nodules and automatically trim the area of the nodules.At the same time,a GoogLeNet network model was constructed to classify benign and malignant nodules after reduction.Results In the collected data set,the mean accuracy of the target detection network for thyroid nodule location detection was 96.2%.Meanwhile,the sensitivity,specificity,accuracy and AUC of the classification network for benign and malignant nodules were 0.885,0.822,0.866 and 0.92 respectively,which were significantly higher than those of the AlexNet model(0.81),VGG model(0.86)and MobileNet model(0.76).Conclusion The deep convolutional neural network model has high localization and recognition ability for benign and malignant thyroid nodules in ultrasound images,which is helpful to improve the accuracy of automatic image diagnosis.

关 键 词：甲状腺结节 超声图像 深度卷积神经网络 YOLOv5网络 

分 类 号：R816.6[医药卫生—放射医学] R320.1140[医药卫生—临床医学]