|
|
本文已被:浏览 1115次 下载 756次 |
![本文二维码信息](/download_upload_file.aspx?file_name=/uploadfile/two_dimensional_code_img/ajsmmu/2022/8/20211217.png) 码上扫一扫! |
基于深度学习的4种肺部超声征象分类 |
段晓倩1,陈建刚1,2,王茵3,秦伟1,曹羽成1,马烨波1,王卓然1,魏高峰4,沈梦君3* |
|
(华东师范大学通信与电子工程学院, 上海市多维度信息处理重点实验室, 上海 200241;华东师范大学通信与电子工程学院, 上海市多维度信息处理重点实验室, 上海 200241;上海中医药大学中医智能康复教育部工程研究中心, 上海 201203;同济大学附属上海市肺科医院超声科, 上海 200092;海军军医大学(第二军医大学)海军医学系医工交叉协同创新中心, 上海 200433;1. 华东师范大学通信与电子工程学院, 上海市多维度信息处理重点实验室, 上海 200241) |
|
摘要: |
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1 500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1 388张、B线图像1 375张、胸腔积液图像1 384张、肺实变图像1 398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。 |
关键词: 肺 超声检查 深度学习 深度残差网络 四分类 |
DOI:10.16781/j.CN31-2187/R.20211217 |
投稿时间:2021-12-05修订日期:2022-05-15 |
基金项目:上海市科学技术委员会科研计划项目(19DZ2203300),上海市2021年度"科技创新行动计划"医学创新研究专项(21Y11902500),上海市肺科医院2020年度院级临床研究项目(FKLY20015). |
|
Classification of 4 pulmonary ultrasound signs based on deep learning |
DUAN Xiao-qian1,CHEN Jian-gang1,2,WANG Yin3,QIN Wei1,CAO Yu-cheng1,MA Ye-bo1,WANG Zhuo-ran1,WEI Gao-feng4,SHEN Meng-jun3* |
(Shanghai Key Laboratory of Multidimensional Information Processing, College of Communication and Electronic Engineering, East China Normal University, Shanghai 200241, China;Shanghai Key Laboratory of Multidimensional Information Processing, College of Communication and Electronic Engineering, East China Normal University, Shanghai 200241, China;Engineering Research Center of Intelligent Rehabilitation of Traditional Chinese Medicine, Ministry of Education, Shanghai University of Traditional Chinese Medicine, Shanghai 201203, China;Department of Ultrasound, Shanghai Pulmonary Hospital, Tongji University, Shanghai 200092, China;Medical and Industry Collaborative Innovation Center, Faculty of Naval Medicine, Naval Medical University (Second Military Medical University), Shanghai 200433, China;1. Shanghai Key Laboratory of Multidimensional Information Processing, College of Communication and Electronic Engineering, East China Normal University, Shanghai 200241, China) |
Abstract: |
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 (1 500 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,1 388 images of A lines,1 375 images of B lines,1 384 images of pleural effusion,and 1 398 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. |
Key words: lung ultrasonography deep learning deep residual network four kinds of classification |
|
|
|
|
|