基于卷积神经网络成人肋骨骨折CT自动检测和分类的应用研究  

Automatic detection of adult rib fractures on CT using convolutional neural network

作　　者：周清清 王佳硕 唐雯 张荣国 ZHOU Qing-qing;WANG Jia-shuo;TANG Wen;ZHANG Rong-guo(Department of Radiology,The Affiliated Jiangning Hospital of Nanjing Medical University,Jiangsu 211100,China)

机构地区：[1]南京医科大学附属江宁医院医学影像科,江苏南京211100 [2]中国药科大学,江苏南京210006 [3]北京推想科技有限公司,北京100000

出　　处：《影像诊断与介入放射学》2020年第1期27-31,共5页Diagnostic Imaging & Interventional Radiology

基　　金：南京市卫健委课题(YKK17226)

摘　　要：目的探讨基于卷积神经网络(CNN)胸部CT平扫自动检测和分类肋骨骨折的准确性和可行性。方法回顾性搜集A医院2011年1月~2019年1月974例成人肋骨骨折患者,另外收集2019年1月B医院25例,C医院25例成人肋骨骨折患者作为多中心测试集进行鲁棒性验证。三种骨折类型(新鲜骨折、愈合期骨折和陈旧性骨折)的相应CT图像被自动检测并输出为结构化报告。采用精准度、召回率和F1值作为衡量CNN模型诊断效能的指标。检测/诊断时间、精准度、灵敏度、fROC曲线用来比较CNN模型的结构化报告和放射科主治医生的诊断效能。结果CNN模型在所有测试集上具有良好的鲁棒性(平均精准度、平均召回率、平均F1值均≥0.8)。新鲜骨折和愈合期骨折的检测效率略高于陈旧性骨折(平均精准度:0.829,0.867>0.814;平均召回率:0.875,0.870>0.827;平均F1值:0.851,0.868>0.821)。CNN模型输出的结构化报告达到放射科主治医师诊断水平,并且CNN模型的检测时间平均缩短132.07 s。结论利用CNN模型可在较短的时间内自动检测并分类肋骨骨折,达到放射科主治医师的诊断水平,且该模型具有一定的鲁棒性和可行性。Objective To investigate the feasibility of automatic detection and classification of rib fractures on thorax CT scan based on convolutional neural network (CNN). Methods 974 adult patients from hospital A from January 2011 to January 2019, 25 adult patients from hospital B and 25 from hospital C in January 2019 were included in the multicenter testing sets for robustness validation in this study. Three types including acute, healing and old fracture with corresponding CT were detected automatically and recorded in structured reports. Precision, recall and F1-score were selected as metrics to measure the performance of CNN model. Detection/diagnosis time, precision, sensitivity and fROC were employed to compare the diagnostic efficiency of structured reports from CNN model and attending radiologists. Results The robustness of the model was good on all testing sets (all mean precision, recall, and F1-score > 0.8). The detection efficiency was higher for acute (mean precision=0.829, mean recall=0.875, mean F1-score=0.851) and healing (0.867, 0.870, 0.868) fractures than that of old fracture (0.814, 0.827, 0.821). The structured report output by CNN model matched the diagnostic performance of attending radiologists and the detection time of the model was reduced by 132.07 s on average. Conclusion Our CNN model can automatically detect and categorize rib fractures in a shorter time with diagnostic performance matching that of attending radiologists.

关 键 词：肋骨骨折 卷积神经 深度学习 人工智能 体层摄影术 X线计算机 结构化报告 

分 类 号：R816.8[医药卫生—放射医学] R683[医药卫生—临床医学]