【摘要】 目的:基于胸部模体探讨不同强度深度学习重建算法（DLIR）对低剂量CT图像上肺结节显示及测量的影响。方法:采用包括纵隔、支气管血管束、软组织及骨骼的成年男性胸部仿真模型，内置直径（体积）为5 mm（66 mm³）、8 mm（268 mm³）和10 mm（523 mm³）的实性结节（SN）及磨玻璃结节（GGN）,对其进行低剂量CT扫描(100 kVp、40 mA,CTDI_VOI=0.84 mGy),采用标准卷积核的自适应统计迭代重建算法（ASIR-V）及中档（DLIR-M）和高档（DLIR-H）深度学习重建算法分别进行图像重建。在肺组织内放置ROI（面积100 mm²）测量肺组织CT值的标准差（SD）作为肺组织噪声(N_肺组织)。选用肺结节CT影像辅助检测系统自动计算得到10 mm SN及10 mm GGN CT值的SD(即N_结节)。计算3组图像上肺组织以及直径10 mm的SN和GGN的信噪比（SNR）及对比噪声比（CNR）,以及所有结节的CT值和体积及其偏差度。对各组图像上肺组织及所有结节的噪声、支气管血管束的锐利度、SN和GGN的显示情况进行主观评价。结果:（1）在3组图像上，DLIR-H的肺组织、SN和GGN的噪声均为最低，肺组织的SNR、SN和GGN的SNR和CNR均为最高（P均<0.05）,肺组织、SN和GGN显示情况的主观评分为最高（P均<0.001）。（2）三组图像上，三种直径SN的平均CT值偏差度的总体差异均无统计学意义（P均>0.05）,三种直径GGN的平均CT值偏差度的总体差异均有统计学意义（P均<0.001）;对于直径10 mm及5 mm的GGN,DLIR-M和DLIR-H图像上测得的平均CT值偏差度均小于ASIR-V（P均<0.001）,DLIR-M和DLIR-H图像上测得的平均CT值偏差度的差异无统计学意义（P>0.05）;对于直径8 mm的GGN,ASIR-V图像测得的平均CT值偏差度均小于DLIR-M、DLIR-H图像（P均<0.001）,而DLIR-M与DLIR-H图像上测得的此指标值的差异无统计学意义（P=0.535）。（3）三组重建图像上测得10 mm、8 mm直径的SN及GGN体积偏差度的总体差异均无统计学意义（P均>0.05）;对于直径5 mm的SN,ASIR-V与DLIR-M组间、DLIR-M与DLIR-H组间均无统计学差异（P均>0.05）,ASIR-V图像测得偏差度小于DLIR-H（P=0.020）;对于直径5 mm GGN,在DLIR-M、DLIR-H图像上测得的体积偏差度均较ASIR-V图像更小（P均<0.05）,而DLIR-M、DLIR-H图像间偏差度的总体差异无统计学意义（P=0.476）。（4）主观评价结果：三组重建图像上肺组织噪声评分的总体差异具有统计学意义（H=15.58,P<0.001）,在DLIR-H图像上该指标评分高于ASIR-V和DLIR-M图像（P<0.05）,而ASIR-V与DLIR-M图像之间该指标评分的总体差异无统计学意义（P=0.849）;三组重建图像的支气管血管束锐利度以及SN和GGN可见度主观评分的总体差异均无统计学意义（P均>0.05）。结论:低剂量下DLIR图像对结节CT值和体积的测量及显示情况与ASIR-V算法相当，而且在DLIR-M和DLIR-H图像上测得的5 mm GGN的平均CT值及体积更准确。对于肺结节的低剂量CT筛查及随访，采用深度学习重建算法是值得推荐的。更多还原

【Abstract】 Objective:To evaluate the effect of chest low-dose CT combined with different strengths of TrueFidelity^TM deep-learning image reconstruction（DLIR） for the display and measurement of pulmonary nodules in the chest phantom.Methods:The adult male chest phantom（including mediastinum, bronchial vascular bundles, soft tissue and skeleton） implanted with solid and ground-glass nodules（SN and GGN） of 5mm（66mm³）,8mm（268mm³） and 10mm（523mm³） in diameter（volume） was scanned by CT with low dose(100kVp, 40mA,CTDI_VOI=0.84mGy),and the raw data was reconstructed with Adaptive Statistical Iterative Reconstruction-Veo（ASIR-V） and the M-strength and H-strength of DLIR（DLIR-M and DLIR-H） at standard kernel for image reconstruction.A ROI（area of 100mm²） was placed in lung tissue to measure the standard deviation of lung tissue’s CT value, which was recorded as lung tissue noise(N_Lung).The CT values’ standard deviation（SD） of 10mm SN and 10mm GGN was automatically calculated by pulmonary nodules CT image au-xiliary detection system, which was recorded as nodule noise(N_Nod).The signal-to-noise ratio（SNR） and contrast-to-noise ratio（CNR） of lung tissue, 10mm SN and GGN,as well as the mean CT values and volumes of all nodules and their deviations were calculated.The noise of lung tissue and all no-dules, the sharpness of bronchial vascular bundles, and the display of solid nodules and ground glass nodules in each group of images were subjectively assessed.Results:（1）The noise of lung tissue, SN and GGN in the DLIR-H images were all the lowest, and the SNR of lung tissue, the SNR and CNR of SN and GGN were all the highest among the three groups（all P<0.05）,and the noise scores of lung tissue, SN and GGN were all the highest（all P<0.001）.（2）There was no significant overall difference in the mean CT value’s deviation of all SN measured in the three groups of reconstructed images（all P>0.05）; and the mean CT value’s deviation of all GGN in the three groups of images was statistically significant（all P<0.001）:the mean CT value’s deviation of 10mm and 5mm GGN was less in DLIR-M,DLIR-H than ASIR-V images（all P<0.001）,and the difference between DLIR-M and DLIR-H images was not statistically significant（P>0.05）.The mean CT value’s deviation of 8mm GGN is smaller in ASIR-V than DLIR-M and DLIR-H images（all P<0.001）,and there is no significant over all difference between DLIR-M and DLIR-H images（P=0.535）.（3）There was not significantly diffe-rence in the volume’s deviation of SN and GGN with a diameter of 10mm and 8mm among various reconstructed images（all P>0.05）.There was no statistical overall difference between DLIR-M,and ASIR-V,DLIR-H of 5mm SN（all P>0.05）,and the deviation of ASIR-V image was less than DLIR-H（P=0.020）.The volume’s deviation of 5mm GGN was smaller on DLIR-M and DLIR-H than ASIR-V images（all P<0.05）,but there was no statistical difference between DLIR-M and DLIR-H images（P=0.476）.（4）The overall difference of lung tissue noise subjective score among the three groups of reconstructed images was statistically significant（H=15.58,P<0.001）,and the subjective score of DLIR-H images was higher than ASIR-V and DLIR-M（P<0.05）,but there was no statistical overall difference between ASIR-V and DLIR-M（P=0.849）.There was all no significant overall difference in the sharpness of bronchial vascular bundle and the subjective score of visibility of SN and GGN among the three groups（all P>0.05）.Conclusion:The measurement and display of the nodules’ CT value and volume in DLIR-M and DLIR-H images are comparable to those of ASIR-V images at low dose, and the measurement of the mean CT value and volume for 5mm GGN is more accurate especially.It is recommended for pulmonary nodules’ low-dose screening and follow-up to combine with deep learning reconstruction algorithms.更多还原