基于三维卷积神经网络肺结节深度学习算法模型临床效能初步评估  

A preliminary clinical evaluation of a 3D convolutional neural network based deep learning system

作　　者：王祥 李清楚[1] 邵影 邹勤 孙安 陈彦博 陈如谭 高耀宗 刘士远[1] 萧毅[1] WANG Xiang;LI Qing-chu;SHAO Ying(Department of Radiology,Changzheng Hospital of Naval Military Medical University,Shanghai 200003,China)

机构地区：[1]海军军医大学(原第二军医大学)附属长征医院影像科,上海200003 [2]上海联影智能医疗科技有限公司,上海201210

出　　处：《放射学实践》2019年第9期942-946,共5页Radiologic Practice

基　　金：上海市科学技术委员会基金项目(NO:17411952400);国家重点研发计划政府间合作项目(NO:2016YFE0103000);上海市卫计委智慧医疗项目(NO:2018ZHYL0101);科技部国家重点研发计划(NO:2018YFC0116404)

摘　　要：目的:为评价人工智能模型的应用价值,本研究在专家共识的基础上建立了肺结节标准测试集,对前期建立的一种基于三维卷积神经网络肺结节深度学习算法模型进行验证,评价该模型的临床效能和限度。方法:基于胸部CT肺结节数据标注与质量控制专家共识建立标准测试数据集,对前期建立的基于三维卷积神经网络的肺结节深度学习算法模型及传统CAD系统(Siemens syngo.via VB 3.0 和Philips ISP V8)进行检验,在肺结节检出灵敏度、精准度以及平均每例假阳性个数等多个指标方面进行优效验证。结果:针对测试数据集中的肺结节,Syngo.via工作站检出灵敏度为36%,精准度为69%,平均每例假阳性1.2个;Philips ISP工作站肺结节检出灵敏度为34%,精准度为73%,平均每例假阳性0.9个;三维卷积神经网络的肺结节深度学习算法模型检出灵敏度为90%,精准度为71%,平均每例假阳性2.8个。结论:该三维卷积神经网络算法模型相较于传统CAD系统,肺结节检出灵敏度显著提升。由于训练数据集的偏倚等问题,灵敏度仍有进一步提升的空间。通过针对性地补充训练数据集,如增加磨玻璃结节的比重,可进一步提升肺结节检出灵敏度。改进之后的模型有望成为影像医生肺癌筛查工作的得力助手。Objective: To assess the application value of artificial intelligence model,a standardized testing set constructed according to the expert consensus were used to validate a prior developed 3D convolutional neural network (3D CNN) algorithm and evaluate its clinical efficacy and limitation. Methods: A standardized testing dataset was constructed based on the expert consensus on the rule and quality control of pulmonary nodule annotation based on thoracic CT and used to test a prior developed 3D CNN algorithm and two traditional CAD systems (Siemens syngo.via VB 3.0 and Philips ISP V8.Several metrics).Several evaluation indicators,such as detection sensitivity,precision and average number of false positives per subject (average FP),were calculated. Results: Testing with the standardized chest CT dataset,the sensitivity,precision and average FP of Syngo.via VB 3.0 were 36%,69%,and 1.2,respectively.The three indicators of Philips ISP V8 were 34%,73%,and 0.9,respectively.The studied 3D CNN algorithm obtained a sensitivity of 90%,a precision of 71%,and an average FP of 2.8. Conclusion: Compared to the traditional CAD systems,the 3D CNN algorithm shows a significant improvement on detection sensitivity of lung nodules.However,the detection sensitivity still needs further improvement due to the unbalanced training dataset.Studying with a more diversified training dataset,e.g.,more included ground-glass nodules,may further improve the detection sensitivity.The AI system can help the radiologists in lung cancer screening in a more efficient way after improvement.

关 键 词：肺结节 人工智能 卷积神经网络 计算机辅助筛查 体层摄影术 X线计算机 磨玻璃结节 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] R734.2[自动化与计算机技术—控制科学与工程]