深度学习图像重建算法在门静脉CT血管成像中的应用研究  

Application research of deep learning image reconstruction algorithm in CT portal venography

作　　者：张悦 蒋伟 郭月飞[1] 郭焯欣 张可[1] 马昆 孟占鳌 Zhang Yue;Jiang Wei;Guo Yuefei;Guo Zhuoxin;Zhang Ke;Ma Kun;Meng Zhan'ao(Department of Radiology,the Third Affiliated Hospital of Sun Yat-sen University,Guangzhou 510630,China;General Electric Medical Treatment China CT Image Research Center,Guangzhou 510623,China)

机构地区：[1]中山大学附属第三医院放射科,广州510630 [2]通用电气医疗中国CT影像研究中心,广州510623

出　　处：《中华放射学杂志》2024年第6期653-660,共8页Chinese Journal of Radiology

基　　金：中山大学附属第三医院国家自然科学基金培育专项(2021GZRPYMS06);中山大学附属第三医院“五个五”工程(2023WW605)。

摘　　要：目的探讨深度学习图像重建(DLIR)在100 kV门静脉CT血管成像(CTPV)降低辐射剂量且提高图像质量的可行性,并与传统120 kV CTPV进行图像质量比较。方法前瞻性纳入中山大学附属第三医院2021年6月至2022年12月行上腹部CTPV检查的患者,分为120 kV标准剂量组(S组)和100 kV低剂量组(L组)。S组采用自适应统计迭代-Veo算法(ASIR-V)及中等强度的DLIR(S-DM)进行重建,ASIR-V的权重分别采用40%、60%和80%(S-AV40、S-AV60和S-AV80)。L组采用60%和80%2种权重的ASIR-V(L-AV60和L-AV80)和中、高强度的DLIR(L-DM和L-DH)进行重建。2名医师分别进行客观指标测量,包括门静脉主干(MPV)、门静脉右支(RPV)、门静脉左支(LPV)、肝实质和右侧竖脊肌的CT值和SD值,并且计算信噪比(SNR)和对比噪声比(CNR)。3名医师分别对8组图像进行5分盲法评价。记录两组容积CT剂量指数(CTDIvol)和剂量长度乘积(DLP),计算有效剂量(ED)。采用单因素方差分析对客观指标、Kruskal-Wallis H检验对主观评分进行统计分析。结果共纳入100例患者,标准剂量组(S组)和低剂量组(L组)各50例。L组的有效剂量比S组减少了51%。MPV、RPV、LPV的SD、SNR和CNR在8组图像之间差异有统计学意义(P值均<0.001)。L-DH的SD与S-AV80组间差异无统计学意义(P>0.05),但显著低于其余6组(P值均<0.05);除MPV的CNR在L-DH与S-AV80组之间差异无统计学意义(P>0.05)外,3支门静脉的SNR和CNR在L-DH组均高于其他组,差异具有统计学意义(P值均<0.001)。L-DH组的主观图像质量评分最高(4.61±0.16),除和S-DM组(4.31±0.19)外的其余6组差异均有统计学意义(P值均<0.001)。医师间的客观指标测量和主观图像质量评分均具有良好的一致性(ICC=0.584~0.960)。结论与常规120 kV CTPV相比,100 kV结合DLIR算法可显著降低CTPV 51%的辐射剂量,并在降低噪声的同时保持更高的SNR、CNR和主观评分。高强度DLIR算法获得的100 kV CTPV图像质量最佳。Objective To explore the feasibility of deep learning image reconstruction(DLIR)in reducing radiation dose and improving image quality in 100 kV portal vein CT angiography(CTPV),and compare the image quality with traditional 120 kV CTPV.Methods Consecutive 100 patients who underwent upper abdominal CTPV examinations at the Third Affiliated Hospital of Sun Yat-sen University from June 2021 to December 2022 were prospectively enrolled in this study.They were divided into two groups:the standard dose group(S group)using 120 kV and the low dose group(L group)using 100 kV.In the S group,adaptive statistical iterative reconstruction-Veo(ASIR-V)was employed with weights of 40%,60%,and 80%(S-AV40,S-AV60,and S-AV80,respectively),along with a medium-level DLIR(S-DM).The L group used two weights of 60%and 80%of ASIR-V(L-AV60 and L-AV80)and medium to high-intensity DLIR(L-DM and L-DH)for reconstruction.Objective evaluation measurements,including CT values and standard deviation(SD)of the main portal vein(MPV),right portal vein(RPV),left portal vein(LPV),hepatic parenchyma,and right vertical muscle,were independently measured by two radiologists.The signal-to-noise ratio(SNR)and contrast-to-noise ratio(CNR)were also calculated.The images of eight groups were blindly evaluated using a 5 points scale by three radiologists.The Volume CT dose index(CTDIvol)and dose length product(DLP)were recorded for both protocol groups,and the effective dose(ED)was calculated.Statistical analysis was performed using Single factor variance analysis for objective indicators and Kruskal-Wallis H test for subjective scores.Results The effective dose in the L group was reduced by 51%compared to the S group.There were significant statistical differences in SD,SNR and CNR among eight groups for MPV,RPV,and LPV(P<0.001 for all).The SD of L-DH showed no statistical difference compared to the S-AV80 group(P>0.05),but it was significantly lower than the other six groups(P<0.05 for all).Except for CNR in the MPV,which showed no statistically significant d

关 键 词：体层摄影术 X线计算机 自适应统计迭代重建-Veo 深度学习图像重建 门静脉CT血管成像 辐射剂量 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程] R816.5[医药卫生—放射医学]