探讨深度学习算法降低胸部CT辐射剂量的应用价值:体模研究

To Explore the Value of Deep Learning Algorithm in Reducing Chest CT Radiation Dose:Phantom Study

目的通过比较低管电流扫描联合深度学习(DL)算法重组图像与常规扫描的胸部CT图像,探讨DL算法在低剂量胸部CT对肺部结节测量和图像质量的影响。方法于仿真胸部体模中随机放置12个大小、密度均不同的人工球形结节。应用联影uCT760对体模进行扫描。依据管电流和不同算法重组进行分组:A组(常规扫描组):100 mAs+滤波反投影法(FBP),B1组:55 mAs+FBP,B2组:55 mAs+DL,C组:25 mAs+DL,各组均采用骨算法和标准算法重组。记录CT容积剂量指数(CTDIvol)、剂量长度乘积(DLP)。在骨算法下使用计算机辅助诊断软件(CAD)检测模拟肺结节,记录结节长径,并计算长径和体积测量的绝对错误率(APE)=100×(V_(测量)-V_(实际))/V_(实际)。在标准算法下测量主动脉弓、心脏最大层面及椎旁软组织的CT值及噪声(SD)值,计算信噪比(SNR)、对比噪声比(CNR)。由两位观察者采用独立双盲法对图像质量及肺结节显示进行5分制评价。采用配对样本t检验比较A、B、C组间及B组组内图像SD值、SNR、CNR及肺结节长径、体积的APE差异;两位放射科医师对图像主观质量评分一致性采用Kappa检验,A、B、C组间及B组组内图像主观质量评分比较采用Wilcoxon符号秩检验。结果A组和B2组结节APE直径和APE体积差异无统计学意义(P>0.05)。C组结节APE直径和APE体积大于A组。B1组和B2组,结节APE直径和APE体积差异无统计学意义(P>0.05)。采用DL算法重组图像的主动脉弓、心脏最大层面SNR高于FBP重组图像,椎旁肌SD值低于FBP重组图像。两位观察者对各组图像质量主观评分一致性较好(Kappa值0.756~0.873)。A组和B2、C组图像质量主观评分无统计学差异(P>0.05),A组图像质量主观评分高于B1组,差异有统计学意义(P<0.05)。B组、C组有效辐射剂量(ED)较A组分别降低44.9%、75.1%。结论应用低管电流扫描联合DL算法重组进行胸部CT成像,在显著降低辐射剂量的同时,能获得与常规扫描质量相当的图像,且不会影响肺结节的检出和测量,具有较好的临床应用价值。

Objective The effects of DL algorithm on the measurement and image quality of pulmonary nodules in low-dose chest CT were investigated by comparing the reconstructed images of low tube current scanning combined with deep learning(DL)algorithm with conventional chest CT images.Methods Twelve artificial spherical nodules with different sizes and densities were randomly placed in the simulated chest body model.The Phantom was scanned by uCT760.Group A:100 mAs+FBP(conventional scanning group),Group B1:55 mAs+FBP,Group B2:55 mAs+DL,and Group C:25 mAs+DL.Bone algorithm and standard algorithm were used for reconstruction in all groups.Volume CT dose index(CTDIvol)and dose length product(DLP)were recorded.The simulated pulmonary nodules were detected by computer aided diagnosis software(CAD)under the bone algorithm,the length and diameter of the nodules were recorded,and the absolute error rates of the length and diameter measurements were calculated[APE=100×(V measurem-V actual)/V actual].The CT and SD values of aortic arch,maximum layer of heart and paravertebral soft tissue were measured under the standard algorithm,and SNR and CNR were calculated.Image quality and pulmonary nodule display were evaluated on a 5-point scale by two independent double-blind methods.Paired sample T test was used to compare the SD values,SNR,CNR and APE differences in the length and volume of pulmonary nodules between groups A,B,C and within group B.The two radiologists used Kappa test for the consistency of subjective quality scores of images,and Wilcoxon symbolic rank test was used for the comparison of subjective quality scores of images between groups A,B and C and within group B.Results There were no significant differences in APE diameter and APE volume between group A and group B2(P>0.05).The APE diameter and APE volume of nodules in group C were larger than those in group A.There were no significant differences in APE diameter and APE volume between group B1 and group B2(P>0.05).The SNR of maximum layer of aortic arch and heart reconstructed by DL algorithm was higher than that of FBP reconstructed image,and SD of paraspinal muscle was lower than that of FBP reconstructed image.The subjective scores of the two observers for each group were consistent(Kappavalue=0.756-0.873).There was no statistical difference in subjective score of image quality between group A and group B2 and C(P>0.05),and the subjective score of image quality in group A was higher than that in group B1(P<0.05).The effective radiation dose of group B and C was 44.9%and 75.1%lower than that of group A,respectively.Conclusion Low tube current scanning combined with DL algorithm reconstruction for chest CT imaging can significantly reduce the radiation dose and obtain images of the same quality as conventional scanning,without affecting the detection and measurement of pulmonary nodules,which has good clinical application value.