摘要
目的:比较管电压100kVp下使用FBP(滤波反投影法)及不同等级迭代重建算法胸部CT图像在AI肺结节中检测效能并寻找较优重建算法。方法:使用西门子双源(Definition FLASH)CT搜集189例门诊肺结节筛查患者行双能量扫描,管电压100 kVp,分别采用FBP及SAFIRE1、3、5级重建算法重建图像得出1 mm肺部图像,测量各组图像平均信噪比(SNR)与对比噪声比(CNR);同时将薄层图像导入AI辅助诊断软件(北京推想公司)行结节自动检测,将检测结果与金标准进行比较得出不同算法结节检测敏感度、准确度、假阳性率,分别对全部结节及实性、亚实性、≥4 mm、<4 mm结节不同算法敏感度、准确度、假阳性率行统计学分析,P<0.05为差异具有统计学意义。结果:100kVp下FBP、SAFIRE1、SAFIRE3、SAFIRE5重建图像CNR和SNR随着迭代等级增加逐渐提升,差异具有统计学意义(P<0.05),AI自动检测全部肺结节各种重建算法敏感度分别为73.5%、70.9%、73.7%、73.3%,差异无统计学意义(P>0.05),假阳性率分别为1.36/CT、1.49/CT、2.31/CT、4.87/CT,FBP算法结节自动检测假阳性率最低,差异具有统计学意义(P<0.05),准确度分别为72.2%、69.4%、61.7%、43.1%,差异具有统计学意义(P<0.05)。结论:管电压100kVp下不同重建算法AI检测全部肺结节、实性结节、亚实性结节≥4mm和<4mm结节敏感度均无显著差别,但FBP假阳性率最低,因此检测效能最优。
Objective:To investigate the influence of iterative reconstruction algorithm on deep learning(DL)-based artificial intelligence diagnostic system in detecting pulmonary nodules on 100kVp CT scans.Methods:CT images of 189 patients for lung nodule screening were acquired using a dual-source scanner(SOMATOM Definition Flash,Siemens Healthineers)at 100kVp.Images were reconstructed with filtered back projection(FBP)algorithm and sinogram affirmed iterative reconstruction(SAFIRE)algorithm with varying weights of 1,3,and 5.The signal to noise ratio(SNR)and contrast to noise ratio(CNR)of reconstructed images were measured.The DL-based AI diagnostic system(InferRead CT Lung Research,Infervision,Beijing)was employed to detect nodules from reconstructed images.The detection results were compared with the gold standard,and the sensitivity and accuracy and false positive rate of nodule detection with different algorithms were obtained.False positive rate was statistically analyzed,P<0.05 was considered statistically significant.Results:At 100kVp,the CNR and SNR of FBP,SAFIRE1,SAFIRE3,SAFIRE5 reconstructed images gradually increase with the increase of iteration level,while the difference is statistically significant(P<0.05),The sensitivity of AI automatic detection with various reconstruction algorithms was 73.5%,70.9%,73.7%,73.3%,respectively,with the difference insignificant(P>0.05).The false positive rates were 1.36/CT,1.49/CT,2.31/CT,4.87/CT,respectively,with the difference statistically significant(P<0.05)and the lowest false positive rate under FBP algorithm.The accuracy is 72.2%,69.4%,61.7%,43.1%,respectively,with the difference statistically significant(P<0.05).Conclusion:There is no significant difference in the sensitivity of AI detection of all pulmonary nodules,solid nodules,subsolid nodules,≥4mm and<4mm nodules under different reconstruction algorithms l at a tube voltage of 100kVp.However,the lowest false positive rate was achieved with FBP,indicating is the best detection efficiency.
作者
曹源
李丹阳
张扬
高道福
顾俊
张清
CAO Yuan;LI Dan-yang;ZHANG Yang(Department of Radiology,Affiliated Zhongshan Hospital Dalian University,Dalian 116001,China)
出处
《放射学实践》
北大核心
2020年第10期1324-1328,共5页
Radiologic Practice
基金
大连市科技计划项目(2015E12SF120)。
关键词
人工智能
图像处理
计算机辅助
低剂量
肺结节
体层摄影术
X线计算机
Artificial intelligence
Image processing
computer-assisted
Low-dose
Pulmonary nodules
Tomography
X-ray computed