基于深度残差网络研发辅助诊断软件用于X线胸片分类诊断  

Development of a ResNet-based CADx software for classification diagnosis in chest X-ray images

作　　者：张晓东[1] 孙兆男 任昕 周宇[1] 周雯 李建辉[1] 谢辉辉 刘婧[1] 张虽虽 李津书[1] 王霄英[1] ZHANG Xiao-dong;SUN Zhao-nan;REN Xin(Department of Radiology, Peking University First Hospital,Beijing 100034,China)

机构地区：[1]北京大学第一医院医学影像科,北京100034 [2]北京赛迈特锐医学科技有限公司,北京100011

出　　处：《放射学实践》2019年第9期952-957,共6页Radiologic Practice

摘　　要：目的:研究以深度残差网络(ResNet)为基础架构建立深度学习模型,对X线胸片(CXR)做出“有发现”与“无发现”鉴别诊断的可行性。方法:回顾性收集2017年1月1日至2018年7月1日的连续CXR图像及诊断报告,经过数据清洗后分为“无发现”组(无任何异常发现,诊断印象为“两肺心膈未见异常”,共9765例)与“有发现”组(诊断印象中提及了一种以上影像所见,共9956例)。使用ResNet152(152 layers)作为二分类模型的基础架构,结合Grad-CAM技术生成模型激活热图,训练二分类模型。数据随机分为训练集(70%)、调优集(20%)和测试集(10%)。以测试集的预测结果检测CXR二分类模型的效能。结果:在测试集中(“有发现”者1018例,“无发现”者995例),CXR二分类模型鉴别“有发现”与“无发现”的精确度分别为0.885和0.894,召回率分别为0.898和0.880,F1-分数分别为0.891和0.887,ROC曲线下面积均为0.96。结论:使用CXR二分类模型可对X线胸片做出“无发现”与“有发现”的预测。Objective: To evaluate the feasibility of residual deep neural network (ResNet) based deep learning model in the differentiation between "no significant finding" and "any significant finding" cases in chest X-ray (CXR) images. Methods: From January 1 2017 to July 1 2018,a consecutive cohort were retrospectively collected in this study.All the images were divided into the "no significant finding" group (defined as no significant finding was detected and reported) and the "any significant finding" group (defined as any significant findings that were detected or reported).Finally,9765 cases were collected in "no significant finding" group and 9956 cases were collected in "any significant finding" group.The ResNet 152 (152 layers) was used as the framework of the binary classification model and the Grad-CAM (gradient class activation maps) was used to generate model activation heating maps.The cohort was randomly divided into training (70%),validation (20%),and testing dataset (10%).The results obtained in testing dataset was considered as the performance of the model. Results: In the testing dataset,1018 cases were classified as "any significant finding",while 995 cases were classified as "no significant finding".The precision,recall and F1-scores were 0.885 vs 0.894, 0.898 vs 0.880 and 0.891 vs 0.887,respectively in terms of two group (any significant finding vs no significant finding).The areas under ROC (AUC) were both 0.96. Conclusion: The CXR binary classification model could be useful in the prediction of "no significant finding" and "any significant finding" for CXR images.

关 键 词：深度学习 人工智能 深度残差网络 胸部X线片 用例 结构化报告 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] R814.41[自动化与计算机技术—控制科学与工程]