肺不同分区结节及其大小、密度对人工智能检出效果的影响  

Impact of the Size and Density of Pulmonary Nodules in Different Areas on Diagnostic Accuracy of Artificial Intelligence.

作　　者：罗艺 余建群[2] 彭礼清[2] 张文钊[2] Luo Yi;Yu Jianqun;Peng Liqing(.Department of Radiology,The First People's Hospital of Long quanyi District,Chengdu,Sichuan 610100;Department of Radiology,West China Hospital,Sichuan University,Chengdu,Sichuan 610041,China.)

机构地区：[1]成都市龙泉驿区第一人民医院放射科,四川成都610100 [2]四川大学华西医院放射科,四川成都610041

出　　处：《四川医学》2021年第9期937-942,共6页Sichuan Medical Journal

摘　　要：目的探讨人工智能(AI)对肺不同分区的结节及其大小、密度对AI检出结节效果的影响。方法分析220例胸部CT图像,由2名高级影像诊断医师确定的真性结节数与AI检出的结节数进行比较,按照肺不同分区、不同最大径和不同密度记录真阳性结节数、假阳性结节数和假阴性结节数,并计算真阳性结节检出率(TPNDR)、假阳性结节检出率(FPNDR)、假阴性结节漏检率(FNNMR),并对以上不同分组间AI检出结节的TPNDR、FPNDR、FNNMR进行统计分析。结果①AI在肺不同分区的TPNDR、FPNDR和FNNMR区间差异均有统计学意义(P<0.05),AI在胸膜下区的TPNDR(97.1%)、胸膜粘连区的FPNDR(31.6%)和肺门区的FNNMR(19.8%)显著高于同组其他肺分区(P<0.05)。②AI在不同最大径分组的TPNDR、FPNDR和FNNMR组间差异均无统计学意义(P>0.05),但各结节最大径分组中,TPNDR、FNNMR在各肺分区间的差异均有统计学意义(P<0.05),5 mm≤最大径≤10 mm组和10 mm<最大径≤30 mm组的FPNDR在各肺分区间的差异均有统计学意义(P<0.05)。AI对不同大小结节的TPNDR在胸膜下区的较高,对较大结节(最大径≥5 mm)的FPNDR在胸膜粘连区较高。③AI在不同结节密度分组的TPNDR、FPNDR和FNNMR组间差异均无统计学意义(P>0.05),但各结节密度分组中,TPNDR、FPNDR和FNNMR在各肺分区间的差异均有统计学意义(P<0.05)。AI对磨玻璃结节的TPNDR在肺门区最高(96.3%,P=0.011),FPNDR和FNNMR在胸膜粘连区最高(62.5%,P=0.000;18.2%,P=0.011)。AI对部分实性结节和实性结节的TPNDR均在胸膜下区最高(94.0%,P=0.035;99.7%,P=0.000),FNNMR均在肺门区最高(19.0%,P=0.035;27.6%,P=0.000),对部分实性结节的FPNDR在中央区最高(28.8%,P=0.016),对实性结节的FPNDR在胸膜粘连区最高(27.4%,P=0.002)。结论结节位置对AI检出结节效果存在影响,结节的大小、密度对AI检出结节效果不存在影响。Objective To evaluate the influence of the size and density of pulmonary nodules in different areas on diagnostic accuracy of artificial intelligence.Methods CT images of 220 cases were analyzed,the number of true pulmonary nodules were determined by two senior imaging diagnostics physicians.After that,compared with the number of nodules detected by AI,the number of true positive nodules,false positive nodules and false negative nodules were recorded respectively based on different pulmonary areas,different diameter groups and different density groups.The true positive nodule detection rate,false positive nodule detection rate,false negative nodule missed rate were calculated.TPNDR,FPNDRand FNNMRwere statistically analyzed between different groups.Results①There were statistically significant differences in TPNDR,FPNDRand FNNMRamong different pulmonary areas(P<0.05).TPNDRin subpleural area(97.1%),FPNDRin pleural adhesion area(31.6%)and FNNMRin hilar area(19.8%)were significantly higher than those in other pulmonary areas of the same group(P<0.05).②There was no significant difference in TPNDR,FPNDRand FNNMRamong different maximum diameter groups(P>0.05),but in each maximum diameter group,the differences of TPNDR,FNNMRin different pulmonary areas were statistically significant(P<0.05).In 5 mm≤the maximum diameter≤30 mm group and 10 mm<the maximum diameter≤30 mm group,the differences of FPNDRin different pulmonary areas were statistically significant(P<0.05).TPNDRfor nodules of different diameter were relatively high in the subpleural region,while FPNDRfor larger nodules(maximum diameter≥5 mm)were relatively high in the pleural adhesion region.③There was no significant difference in TPNDR,FPNDRand FNNMRamong different density groups(P>0.05),but in each density group,the differences of TPNDR,FNNMRin different pulmonary areas were statistically significant(P<0.05).TPNDRfor ground glass nodules was the highest in the hilar area(96.3%,P=0.011),and FPNDRand FNNMRwere the highest in the pleural adhesion area(

关 键 词：人工智能 CT 肺结节 

分 类 号：R814.42[医药卫生—影像医学与核医学] R563[医药卫生—放射医学]