摘要
目的:评价基于CT双期增强图像的不同深度迁移学习(DTL)模型对甲状腺良恶性结节的分类效能。方法:采用相同程序架构和相同数据集对3种DTL模型(VGG19、ResNet50和DenseNet201)的分类诊断效能进行测试和评估。以不同模型在训练集和测试集中的最高预测符合率和在验证集中的符合率、召回率、F1评分和受试者工作特性曲线(ROC)下面积作为评估模型效能的指标。结果:DenseNet201模型获得了最好的训练和测试结果,在训练集和测试集中的最高预测符合率分别为1.00和0.98;VGG19模型用时最长,其在训练集和测试集中的预测符合率分别为0.99和0.98,较DenseNet201略差;ResNet50模型用时最短,但测试结果最差,在训练集和测试集中的最高符合率分别为0.93和0.92。VGG19、ResNet50和DenseNet201模型在验证集中的平均符合率为0.96、0.92和0.98,召回率分别为0.96、0.91和0.98,F1评分分别为0.96、0.91和0.98。DenseNet201模型的ROC曲线下面积为0.98,高于VGG19模型(0.95)和ResNet50模型(0.91)。结论:基于DenseNet201的DTL模型对甲状腺CT良恶性结节具有较高的分类效能,有助于提高影像诊断准确性。
Objective:To evaluate the use of different deep transfer learning(DTL)models to differentiate malignant from benign thyroid nodules on CT images.Methods:The diagnostic efficiencies of three DTL models(VGG19,ResNet50,DenseNet201)were tested and evaluated by using the same program architecture and data set.The highest accuracy of different models in training set and testing set,as well as the accuracy,recall rates,F1 scores and area under the subjects operating characteristic curves in validation set were taken as the performance indicators.Results:DenseNet201 model obtained the best training and testing results,the highest prediction accuracy of training set and testing set were 1.00 and 0.98.VGG19 model took the longest time,and were slightly worse in terms of training and testing prediction accuracy(0.99,0.98)than DenseNet201 model.ResNet50 model took the shortest time for data analysis,but got the worst training and testing accuracy:0.93 and 0.92,respectively.In validation set,the efficacy index values of the three models(VGG19,ResNet50,DenseNet201)were as follows:average accuracy of 0.96,0.92 and 0.98,recall rate of 0.96,0.91 and 0.98,F1 score of 0.96,0.91 and 0.98,and area under the ROC curve of 0.95,0.91 and 0.98.Conclusion:DTL model based on DenseNet201 has excellent efficacy for differentiating malignant from benign thyroid nodules on CT images,and can provide valuable information to improve the diagnostic accuracy.
作者
孟名柱
潘昌杰
张铭
何光远
沈栋
陈罕奇
MENG Ming-zhu;PAN Chang-jie;ZHANG Ming(Department of Radiology,the Second Hospital of Changzhou,Changzhou 213164,China)
出处
《放射学实践》
CSCD
北大核心
2021年第8期976-980,共5页
Radiologic Practice
关键词
迁移学习
卷积神经网络
甲状腺结节
体层摄影术
X线计算机
Deep transfer learning
Convolutional neural networks
Thyroid nodules
Tomography,X-ray computed
作者简介
孟名柱(1984-),男,安徽凤阳人,硕士研究生,主治医师,主要从事人工智能辅助影像诊断;通讯作者:陈罕奇,E-mail:chenhanqiab@sina.com。
