深度学习计算机辅助诊断系统检测胸部DR肋骨骨折的验证性研究  

A confirmatory study of deep learning based computer-aided diagnosis system for detecting chest DR rib fractures

作　　者：张倩[1] 王大为 隋时[1] 潘珅[1] 林坤[1] 孙娜[1] 刘浩蒂[1] 侯阳[1] ZHANG Qian;WANG Da-wei;SUI Shi;PAN Shen;LIN Kun;SUN Na;LIU Hao-di;HOU Yang(Department of Radiology,Shengjing Hospital of China Medical University,Shenyang 110004,China;Beijing Infervision Technology Co.,Ltd.,Beijing 100025,China)

机构地区：[1]中国医科大学附属盛京医院放射科,辽宁沈阳110004 [2]北京推想科技有限公司,北京100025

出　　处：《中国临床医学影像杂志》2022年第6期430-434,共5页Journal of China Clinic Medical Imaging

基　　金：中国医科大学附属盛京医院345人才工程。

摘　　要：目的:验证应用深度学习算法的计算机辅助诊断(DL-CAD)系统检测胸部DR片肋骨骨折的应用价值。方法:回顾性收集110例因外伤拍摄胸部正位DR片的患者。DL-CAD系统及4名放射科主治医生(5~9年诊断经验)分别阅片,采用灵敏度、特异度、准确率指标比较两者对骨折病例的筛检效能;精准度、召回率及F1值比较两种阅片模式在骨折定位诊断中的效能;采用阅片时间评价诊断效率。结果:DL-CAD系统识别患者是否存在肋骨骨折的表现与放射科医生的平均水平相比无统计学差异,ROC曲线下面积(AUC)分别为0.824 vs. 0.833(P>0.05),二者的敏感度,特异度及准确率分别为87.5%vs. 88.3%、73.9%vs. 69.6%和81.8%vs. 80.5%,差异无统计学意义(均P>0.05)。在骨折的定位诊断方面,DL-CAD系统的诊断效能逊于放射科医生,二者的精准度、召回率及F1值分别为60.6%vs. 62.1%、55.7%vs. 73.1%和0.58 vs. 0.67,差异有显著的统计学意义(均P<0.05);侧肋的诊断准确率高于后肋。DL-CAD系统每例阅片时间较医生显著缩短((2.00±0.35) s vs.(108.00±6.36) s,P<0.01)。结论:DL-CAD系统可高效的进行肋骨骨折检测,在筛检骨折病例上的效能与放射科医生相当,但其骨折定位诊断能力有待进一步提升。Objective: To verify the application value of a deep learning based computer-aided diagnosis(DL-CAD) system in detecting rib fractures on chest DR. Methods: The chest DR images of 110 patients with trauma were analyzed retrospectively. DL-CAD system and 4 attending radiologists(5~9 years of diagnostic experience) read the images respectively. The indexes of sensitivity, specificity and accuracy were used to compare the screening efficiency of the two methods for fracture cases. The accuracy, recall rate and F1 value were used to compare the diagnostic efficiency of the two models in fracture localization diagnosis, and the image reading time was used to evaluate the diagnostic efficiency. Results: There was no significant difference between the performance of the DL-CAD system in identifying patients with rib fractures and the average level of radiologists. The area under the ROC curve(AUC) was 0.824 vs. 0.833(P>0.05), respectively, and the sensitivity, specificity and accuracy rate were 87.5% vs. 88.3%, 73.9% vs. 69.6% and 81.8% vs. 80.5%, respectively, and the difference was not statistically significant(all P>0.05). In the localization diagnosis of fracture, the diagnostic efficiency of DL-CAD system was inferior to that of radiologists. The accuracy, recall rate and F1 value of the two methods were 60.6% vs. 62.1%, 55.7% vs.73.1% and 0.58 vs. 0.67 respectively, with significant difference(all P<0.05), the diagnostic accuracy of lateral ribs was higher than that of posterior ribs. The reading time of DL-CAD system was significantly shorter than that of radiologists((2.00±0.35) s vs.(108.00±6.36) s, P<0.01). Conclusion: The DL-CAD system can efficiently detect rib fractures, its efficiency in screening fracture cases is equivalent to that of radiologists, but its ability of fracture location needs to be further improved.

关 键 词：肋骨骨折 诊断 计算机辅助 

分 类 号：R683[医药卫生—骨科学] R319[医药卫生—外科学]