摘要
目的评估基于深度学习算法的计算机辅助检测系统(CAD),在低剂量胸部CT中肺结节检出能力。方法95例低剂量胸部CT纳入观察CAD,1名低年资医师(5年工作经验)、1名高年资医师(10年以上)分别独立阅片,对比CAD与医师检出总灵敏度、以及不同密度、大小及位置分组中灵敏度差异,计算CAD假阳性率。结果三者共检出结节149个,总灵敏度及实性结节组灵敏度,CAD与低、高年资医师无统计学差异(69%&66%、60%,P=0.389;64%&71%、65%,P=0.620);亚实性结节组及纯磨玻璃密度结节组,CAD灵敏度分别为83%、83%,显著高于低年资医师(55%,P=0.005;55%,P=0.010),与高年资医师差异不显著(60%,P=0.021;58%,P=0.018)。直径<4mm组,CAD灵敏度31%,显著低于低年资医师(80%,P<0.001)及高年资医师(41%,P=0.001);直径≥4mm及≥6mm组,CAD灵敏度分别为89%、98%,明显高于低年资医师(59%,P<0.001;71%,P=0.009)及高年资医师(70%,P=0.003;71%,P=0.009);直径>8mm组,三者灵敏度无明显差异(P=0.222)。肺外周部及中央部结节,三者检出灵敏度无统计学差异(外周部P=0.599、中央部P=0.087)。CAD假阳性0.65个/例。结论基于深度学习的CAD系统在低剂量胸部CT肺结节检出总灵敏度与医师相当,4mm及以上结节CAD检出灵敏度明显高于放射医师,且伴随较低假阳性;对亚实性结节CAD检出优于放射医师,尤其纯磨玻璃密度结节检出灵敏度较以往有明显提升。
Objective To evaluate the performance of computer-aided detection(CAD)system based on deep learning algorithm for detection of pulmonary nodules on low-dose CT screening.Methods 95 low-dose screening CT examinations were analyzed by CAD software and 2 radiologists(one with more then 10 years of experience,resident in 5th yeons)independently.All detected nodules were classified by density,size and location.The sensitivity was compared betwen radiologists and CAP.The false positive was calculated.Results 149 nodules were detected.There was no significant difference in total sensitivity and solid nodules sensitivity among CAD and the radiologists(69%&66%,60%,P=0.389;64%&71%,65%,P=0.620).For the SSN and pGGN detection,the sensitivity of CAD(83%,83%)was significantly higher than that of resident(55%,P=0.005;55%,P=0.010).However no significant differemce was observed with senior radiologist(60%,P=0.021;58%,P=0.018).In group with diameter<4 mm,CAD sensitivity was 31%,which was significantly lower than the resident(80%,P<0.001)and the senior radiologist(41%,P=0.001).In groups with diameter≥4 mm and≥6 mm,CAD sensitivity(89%,98%)ware significantly higher than that of resident(59%,P<0.001;71%,P=0.009)and senior radiologsit(70%,P=0.003;71%,P=0.009);In the group with diameter>8 mm,there was no significant difference in sensitivity among the three groups(P=0.222).No significant difference was found in the detection sensitivity of peripheral and central pulmonary nodules among the three groups(P=0.599,P=0.087).CAD false positive was 0.65/case.Conclusion The overall sensitivity of the deep learning based CAD system in the detection of pulmonary nodules in low-dose chest CT is similar to radiologists.In the detection of nodules arger than≥4 mm,the sensitivity is higher than radiologists.The sensitivity of CAD is higher than that of radiologists in SSN detection,especially the sensitivity of the detection of pGGN is obviously improved.
作者
孟晓燕
顾慧
王锡明
朱海峰
张莹
MENG Xiaoyan;GU Hui;WANG Ximing;ZHU Haifeng;ZHANG Ying(Department of Radilolgy,Civil Aviation General Hospital,Beijing 100123,P.R.China;Department of Medical Imaging,Shandong Provincial Hospital Affiliated to Shandong University,Jinan 250021,P.R,China)
出处
《医学影像学杂志》
2019年第12期2042-2046,共5页
Journal of Medical Imaging