基于深度神经网络的颈动脉斑块超声图像诊断技术研究

Study of diagnosis for carotid plaque by ultrasound images based on deep neural network

目的构建颈动脉斑块超声图像数据集,探讨深度学习技术对颈动脉斑块自动分类诊断的应用价值。方法选取354例颈动脉斑块患者和254例正常成人的超声图像,每例均采集2幅颈部动脉图像,构建共包含1216幅颈动脉超声图像的数据集;基于已构建的颈动脉超声图像数据集对传统的HOG+SVM方法和14种不同结构的深度神经网络模型进行训练;通过分类精确率、召回率、精确率和召回率的调和平均值(F1)确定现有的颈动脉斑块超声图像分类性能最好的深度神经网络模型。结果通过综合比较15种不同的颈动脉斑块超声图像分类方法,得出性能最好的模型为深度残差网络模型ResNet50,其精确率、召回率和F1值分别为97.36%、97.32%和97.34%。结论基于ResNet的颈动脉超声图像自动诊断方法能够准确地区分颈动脉斑块与正常颈动脉超声图像,为后续临床应用提供了技术参考。

ObjectiveTo construct the carotid plaque ultrasound image dataset,and to explore the application value of deep learning technology in the automatic classification and diagnosis of carotid plaque.MethodsThe ultrasound images of354 patients with carotid plaque and 254 normal adults were selected.Two carotid artery images were collected in each case,and a carotid ultrasound image dataset containing 1216 images was constructed.Then,traditional method HOG+SVM and 14 different deep neural network models were trained based on proposed dataset.Finally,the best network model was determined based on 3evaluation indexes:precision,recall and F1 score.Results for carotid plaque diagnosis,the depth residual network model ResNet50 had the best performance,and its precision,recall and F1 values were 97.36%,97.32%and 97.34%,respectively.Conclusion image based on ResNet can accurately distinguish carotid plaque and normal carotid ultrasound image,which provides a technical reference for clinical application.