使用深度学习模型对头颅CT平扫图像进行二分类的初步探讨  

Development of deep learning model for classification diagnosis of non-contrast head CT examination:preliminary study

作　　者：张晓东[1] 刘想 谢辉辉 张宏[1] 任昕 赵永为[1] 王可[1] 吴静云[1] 刘婧[1] 林志勇[1] 张虽虽 刘建新[1] 王霄英[1] ZHANG Xiaodong;LIU Xiang;XIE Huihui;ZHANG Hong;REN Xin;ZHAO Yongwei;WANG Ke;WU Jing yun;LIU Jing;LIN Zhiyong;ZHANG Suisui;LIU Jianxin;WANG Xiaoying(Department of Radiology,Peking University First Hospital,Beijing 100034,China;Beijing Smart Tree Medical Technology Co.,Ltd.,Beijing 100011,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：《实用放射学杂志》2020年第10期1670-1675,共6页Journal of Practical Radiology

基　　金：北京大学第一医院青年临床研究专项基金项目(2018CR25)。

摘　　要：目的探讨使用深度残差网络(ResNet)模型鉴别头颅CT平扫(NCHCT)图像二分类诊断的可行性。方法选取NCHCT连续病例,包括CT图像及其影像诊断报告,以图像为单位,分为“无发现”组(该层面无任何异常发现),共5786幅图像;“有发现”组(该层面有1种以上的疾病或其他影像所见),共3564幅图像。将全部图像分为训练集(70%)、调优集(20%)和测试集(10%),各组图像不相同。以ResNet 50(50 layers)为二分类模型的基础架构训练二分类模型,结合梯度类激活图(Grad-CAM)技术生成激活热图。使用测试集的图像检验NCHCT二分类模型的效能。结果在测试集中(“有发现”者361幅图像,“无发现”者573幅图像),NCHCT二分类模型鉴别“有发现”与“无发现”的精确度分别为0.838和0.973,召回率分别为0.961和0.883,F1-分数分别为0.896和0.926,平均AUC为0.99。结论以ResNet为基础架构的模型在不需要对病灶进行分割的前提下,可对头CT平扫图像逐层预测二分类。Objective To investigate the feasibility of using the deep residual network(ResNet)model to discriminate the diagnosis of noncontrast head CT(NCHCT)images.Methods This study was approved by local IRB,and a retrospective cohort was constructed for the study.NCHCT images and the reports were collected.The“no significant finding”images were defined as no significant finding was detected on the NCHCT image.On the contrary,“any significant finding”images were defined as any significant finding was detected.5786 images were allocated as“no significant finding”data,and 3564 images were allocated as“any significant finding”data.NCHCT binary classification model was trained by using ResNet 50(50 layers)as the framework of the model and combined with gradient class activation maps(Grad-CAM)as a method for generating model activation heating maps.The cohort was randomly divided as training(70%),validation(20%),and testing dataset(10%)for the model training.Results In the testing dataset(“any significant finding”=361 images,“no significant finding”=573 images),the accuracy of the NCHCT binary classification model were 0.838 and 0.973,the recall rates were 0.961 and 0.883,the F1-scores were 0.896 and 0.926,and the mean AUC was 0.99.Conclusion The ResNet-based model can perform slice-by-slice classification of head CT images without the segmentation of lesions.

关 键 词：深度学习 人工智能 计算机体层成像 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] R814.42[自动化与计算机技术—控制科学与工程]