摘要
目的:通过与单纯人工阅片进行比较,探讨基于DenseNet网络深度学习的人工智能肺结节自动检测系统鉴别肺结节良恶性的价值和优势。方法:搜集2015年1月-2017年12月本院510例肺结节CT检查病例,由医师组(按照从事胸部CT诊断的年限分为高级医师组和初级医师组)和人工智能组(基于DenseNet网络深度学习的人工智能系统)分别对所有肺结节进行良恶性的诊断,以病理结果为金标准,分别统计各组在不同大小肺结节(直径≤10 mm、10 mm<直径≤20 mm以及直径>20 mm)良恶性诊断上的敏感度、特异度及符合率,并通过卡方检验进行统计分析。结果:在510例肺结节的诊断中,人工智能组诊断敏感度(93.14%)与高级医师组(91.14%)间差异无统计学意义(P>0.05),与初级医师组(61.43%)间的差异具有统计学意义(P=0.000);而诊断特异度(95.63%)及符合率(93.92%)均高于医师组(初级56.25%、59.80%;高级58.75%、80.98%),差异均有统计学意义(P=0.000)。在≤10 mm的肺结节中,人工智能组的诊断敏感度、特异度及符合率均高于高级医师组(90.38%,92.96%,91.43%;78.85%,64.79%,73.14%;所有P=0.000);在10 mm<直径≤20 mm和直径>20 mm肺结节组中,人工智能组的诊断敏感度(92.25%,97.12%)与高级医师组(95.77%,97.12%)间的差异均无统计学意义(P=0.211和1.000),但诊断特异度(98.33%,96.55%)及符合率(94.06%,96.99%)均高于高级医师组(51.67%,58.62%;82.67%,88.72%),差异均有统计学意义(P<0.05)。在不同大小的三组结节中,人工智能组诊断敏感度、特异度及符合率均高于初级医师组(敏感度:90.38%vs.17.31%,92.25%vs.70.42%,97.12%vs.93.27%;特异度:92.96%vs.85.92%,98.33%vs.43.33%,96.55%vs.10.34%;符合率:91.43%vs.45.14%,94.06%vs.62.38%,96.99%vs.75.19%),除两组在≤10 mm肺结节中的诊断特异度(P=0.361>0.05)和>20 mm肺结节中的诊断敏感度(P=0.211>0.05)的差异无统计学意义外,其它指标的组间比较差异均有统计学意义(P均<0.05)。结论:相较于人工诊断,应用人工智能(DenseNet网络深度学习)技术对肺结节的良、恶性进行诊断具有良好、可靠的诊断准确性。
Objective:To explore the value and advantages of artificial intelligence(AI)lung nodule automatic detection system based on DenseNet network deep learning on CT images to identify benign and malignant lung nodules.Methods:A total of 510 cases of pulmonary nodules were collected from January 2015 to December 2017.The diagnosis of benign and malignant nodules was made by the physician groups(senior group and primary group,according to years of experience in chest CT diagnosis)and the AI group(based on DenseNet network deep learning)respectively.The pathological results were used as the gold standard.The sensitivity,specificity and accuracy of each group in the diagnosis of benign and malignant lung nodules were statistically analyzed.Results:In the diagnosis of 510 pulmonary nodules,the diagnostic sensitivity of AI group was not significantly different from that of the senior physician group(93.14%vs.91.14%;P>0.05),but significantly different from that of the primary physician group(61.43%;P=0.000).The specificity and accuracy of AI group(95.63%,93.92%)were higher than those in physician groups(primary:56.25%,59.80%;advanced:58.75%,80.98%)with significant statistical differences(all P=0.000).On this basis,according to the diameter(D)of the pulmonary nodules D≤10mm,10<D≤20mm and D>20mm,510 cases were further subdivided into three groups(A,B and C).In group A,the diagnostic sensitivity,specificity and accuracy of the AI group(90.38%,92.96%,91.43%)were higher than those of the senior physician group(78.85%,64.79%,73.14%)with significant statistical difference(P=0.000);in group B and group C,the diagnostic sensitivity of the AI group(92.25%,97.12%)was not significantly different from that of the senior physician group(95.77%,97.12%;P=0.211,1.000),but the diagnostic specificity and accuracy(98.33%,96.55%;94.06%,96.99%)were higher than those of the senior physician group(51.67%,58.62%;82.67%,88.72%)with statistical difference(P<0.05).In group A,B and C,the diagnostic sensitivity,specificity and accuracy of AI group were all higher than those of the primary physician group(sensitivity:90.38%vs.17.31%,92.25%vs.70.42%,97.12%vs.93.27%;specificity:92.96%vs.85.92%,98.33%vs.43.33%,96.55%vs.10.34%;accuracy:91.43%vs.45.14%,94.06%vs.62.38%,96.99%vs.75.19%);There was no statistical difference only between the AI group and the primary physician group in the diagnostic specificity in group A(P=0.361)and the diagnostic sensitivity in group C(P=0.211).Conclusion:Artificial intelligence(DenseNet network deep learning)demonstrates more reliable diagnostic accuracy than manual diagnosis in the diagnosis of benign and malignant pulmonary nodules.
作者
戴正行
胡春洪
王希明
陈琦
夏菁
姚柳
刘稳
DAI Zheng-xing;HU Chun-hong;WANG Xi-ming(Department of Radiology,Xishan People's Hospital,Jiangsu 214000,China)
出处
《放射学实践》
北大核心
2020年第4期484-488,共5页
Radiologic Practice
基金
苏州市民生科技示范工程(SS201808)
江苏数字创新诊疗装备应用示范研究(2017YFC0114300)。
