低剂量CT对肺结节AI检测效能及影像组学特征的影响:体模研究

Effect of Low-dose CT on AI Detection Efficacy and Radiomics Features of Pulmonary Nodules:A Phantom Study

目的:探讨低剂量CT及不同低剂量扫描方案对肺结节人工智能(AI)检测效能和CT影像组学特征的影响。方法:于仿真胸部体模中随机放置4种大小(直径5mm、8mm、10mm、12mm)、2种密度(100 HU、-630HU)的人工球形结节。应用uCT760对体模进行常规剂量扫描(A组:120kV+200mA)和低剂量扫描(B组:B1组为120kV+50mA,B2组为100kV+90mA)。各组均采用骨滤波(B_SHARP_C),重建层厚1mm。重建图像传至AI辅助诊断系统进行肺结节自动检测,记录结节的大小、密度、体积,并计算长径、体积测量的绝对错误率。以3D-Slicer软件对A和B组图像中肺结节手动勾画感兴趣区(ROI),采用Pyradiomics平台提取每个ROI的影像组学特征107个,包括14个形状特征、18个一阶统计量和75个纹理特征。采用单因素方差分析比较3种扫描条件下结节长径和体积的测量准确度[以绝对错误率(APE)表示]大小。使用一致性相关系数(CCC)评估常规剂量和低剂量及不同低剂量扫描方案所获得的图像上肺结节分割结果的组学特征一致性。结果:相较于A组,B1和B2的有效剂量(ED)分别降低75.0%、73.3%。A组、B1组和B2组肺结节检出率均为100%,肺结节定性一致,长径、体积的APE差异均无统计学意义(P>0.05)。随着CCC阈值增加,可重复组学特征占比均呈明显下降趋势。低剂量组所提取的一阶统计量和纹理特征与常规剂量组一致性均较差,而形状特征一致性较好;相同低剂量下,不同低剂量实现方式所提取的一阶统计量和纹理特征致性也较差,实性结节更为显著。结论:低剂量CT下AI系统的肺结节检测结果是可信的;但低剂量CT及不同低剂量扫描方案所提取的组学特征与常规剂量一致性均较差,尤其是一阶统计量和纹理特征。

Purpose:To investigate the effects of low-dose CT and different low-dose scanning protocols on artificial intelligence(AI)detection efficacy and CT radiomics features of pulmonary nodules.Methods:Artificial spherical nodules of 4 sizes(5 mm,8 mm,10 mm,12 mm in diameter)and 2 densities(100 HU,-630 HU)were randomly placed in the simulated chest phantom.Conventional scan(group A:120 kV+200 mA)and low-dose scan(group B:120 kV+50 mA in group B1,100 kV+90 mA in group B2)were applied to the phantom with uCT760.Bone filtering(B_SHARP_C)was used in each group,and the reconstruction slice thickness was 1 mm.The reconstructed images were transmitted to the Al-aided diagnosis system for automatic detection of pulmonary nodules,and the size,density,and volume of nodules were recorded,and the absolute percentage error(APE)of length-diameter and volume measurements was calculated.The regions of interest(ROIs)of lung nodules in group A and B images were manually outlined with 3D-Slicer software,and the pyradiomics platform was used to extract 107 radiomics features for each ROI,including 14 shape features,18 first-order statistics and 75 texture features.Oneway ANOVA was used to compare the APE of nodal length-diameter and volume measurement under 3 scanning conditions.The consistency of the radiomics features of lung nodule segmentation results on images obtained with conventional and low dose and different low dose scanning protocols was assessed using the consistency correlation coefficient(CCC).Results:Compared with group A,the effective dose(ED)of B1 and B2 was reduced by 75.0%and 73.3%,respectively.The pulmonary nodules detection rate was 100%in groups A,B1 and B2,with consistent pulmonary nodule characterization and no statistical significant differences in APE for length diameter and volume(P>0.05).With the increase of CCC threshold,the reproducible histological feature ratios all showed a significant decreasing trend.The first-order statistics and texture features extracted from the low-dose group were less consistent with those of the conventional dose group,while the shape features were more consistent;the first-order statistics and texture features extracted from different low-dose protocol were also less consistent,and the solid nodules were more significant.Conclusion:The results of lung nodule detection by the AI system at low-dose CT were plausible;however,the radiomics features extracted by both low-dose CT and different low-dose scanning protocols were in poor agreement with conventional dose,especially the first-order statistics and texture features.