基于深度学习的4种肺部超声征象分类

Classification of 4 pulmonary ultrasound signs based on deep learning

目的探究基于深度残差网络ResNet152对4种常见肺部超声征象的分类。方法前瞻性收集2020年6月至9月在同济大学附属上海市肺科医院超声科进行超声检查患者的超声图像,分别采集A线、B线、胸腔积液、肺实变的肺部超声图像各1500张。选择其中清晰、未被骨骼遮挡、征象单一的图像,最终入选A线图像1388张、B线图像1375张、胸腔积液图像1384张、肺实变图像1398张。采用深度残差网络ResNet152进行分类模型的训练和验证,并在完全独立于训练集和验证集的测试集上测试模型的泛化能力。以精确率、准确度、特异度、召回率和F1指数评价深度分类模型的分类性能,并通过混淆矩阵直观地展示分类结果。结果基于深度残差网络的分类模型分类A线、B线、胸腔积液和肺实变4种征象的精确率分别为97.51%、87.31%、85.42%、93.70%,召回率分别为90.38%、86.97%、94.25%、91.18%。4种肺征象的整体分类精确率为90.99%,准确度为90.70%,特异度为96.85%,F1指数为90.50%,表现出优秀的分类特性。结论基于深度残差网络的肺部超声分类模型表现出较高的分类特性,有潜力辅助超声医师做出诊断。

Objective To explore the classification of 4 common pulmonary ultrasound signs based on deep residual network ResNet152.Methods We prospectively collected ultrasound images of A lines,B lines,pleural effusion and lung consolidation(1500 at each site)from Jun.to Sep.2020 in Shanghai Pulmonary Hospital of Tongji University.Images that were clear,uncomplicated and not obscured by bones were selected.Finally,1388 images of A lines,1375 images of B lines,1384 images of pleural effusion,and 1398 images of lung consolidation were included.Deep residual network ResNet152 was used to train and verify the classification model,and the generalization ability of the model was tested on the test set completely independent of the training set and the validation set.Accuracy rate,precision rate,specificity rate,recall rate and F1-score were used as the evaluation criteria for classification,and the classification results were visualized by the confusion matrix.Results The accuracy rates of the classification model based on the deep residual network to classify the 4 signs of A lines,B lines,pleural effusion and lung consolidation were 97.51%,87.31%,85.42%and 93.70%,respectively,and the recall rates were 90.38%,86.97%,94.25%and 91.18%,respectively.The overall classification accuracy rate of the 4 lung signs was 90.99%,the precision rate was 90.70%,the specificity rate was 96.85%,and the F1-score was 90.50%,showing excellent classification characteristics.Conclusion The pulmonary ultrasound classification model based on deep residual network shows high classification characteristics,and has potential to assist ultrasound diagnosis.