探究深度学习算法在降低胃周血管CTA辐射剂量的可行性

To Explore the Feasibility of Deep Learning Algorithm in Reducing the Radiation Dose of Perigastric Vascular CTA

目的探究深度学习(DL)算法在降低胃周血管CT血管造影(CTA)辐射剂量中的临床应用价值。方法前瞻性搜集2022年4月至2022年5月临床上行胃周血管CTA的患者60例,应用联影uCT760进行扫描,按检查时间分为A、B两组,每组各30例。A组为常规辐射剂量组,管电压120 kVp,管电流为剂量调制等级3,迭代重组算法为Karl 5;B组为低辐射剂量组,管电压120 kVp,管电流为剂量调制等级2,Karl 5及DL(1~4)四个等级重组,获得B1~B5五个亚组。记录A、B组平均mAs、CT容积剂量指数(CTDIvol)、剂量长度乘积(DLP),计算有效辐射剂量(ED)。在各组轴位图像上测量胃左动脉起始部的腹主动脉、腹腔干、脾动脉、肝动脉及同层面腹壁皮下脂肪组织的CT值和SD值,计算上述血管的信噪比(SNR)、对比噪声比(CNR);重组容积再现(VR)及最大密度投影(MIP)图像,两名观察者采用五分法评估各组轴位、VR和MIP的图像质量、胃周血管及分支的显示情况。结果A、B组患者性别、年龄、体质量指数(BMI)差异无统计学意义(P>0.05),B组平均mAs、CTDIvol、DLP、ED相较于A组分别降低了27.00%、29.76%、33.64%、33.76%,差异均有统计学意义(P<0.05);A、B组各血管的CT值差异无统计学意义(P>0.05),A、B1组间比较,SD、SNR、CNR差异均有统计学意义(P<0.05),B2~B5随着DL算法等级的提高,图像SD值逐渐下降,SNR、CNR逐渐升高(P<0.05),B3、B4组各血管的SD值、SNR及CNR相较于A组差异无统计学意义(P>0.05),B组组内两两比较,B1与B2、B3组各血管的SD值、SNR、CNR差异无统计学意义(P>0.05);两观察者对图像主观评分一致性好(Kappa值为0.818~0.860,P<0.05),各组图像均满足诊断要求。A、B1组两观察者主观评分差异有统计学意义(P<0.05),B2~B5组图像质量的主观评分随着DL等级的提高呈现先高后低的趋势,B4组得分最高(4.70±0.47,4.47±0.57)。结论DL算法能够有效降低图像噪声,在低辐射剂量的情况下,提高胃周血管CTA图像质量,其中DL 3为推荐的最佳重组等级。

Objective To explore the application value of the deep learning reconstruction algorithm in low-dose gastric arterial CT angiography.Methods A total of 60 patients for gastric arterial CT angiography were prospectively enrolled from April to May 2022.All patients were scanned using the 128-slice multidetector CT scanning of United Imaging Healthcare,and divided into groups A and B with 30 cases in each group according to the examination time.Group A was the normal radiation dose group,and 120 kVp,dose modulation 3 and Karl 5 algorithm reconstruction were used.Group B was the low radiation dose group,and 120 kVp,dose modulation 2,Karl 5 and DL algorithm reconstruction with 1-4 levels(B1-B5)were used.The mean mAs,volume CT dose index(CTDIvol)and dose length product(DLP)of the two groups were recorded,and the effective dose(ED)was calculated.The CT and SD values of abdominal aorta,celiac trunk,splenic artery,and hepatic artery in each group were measured.Subcutaneous adipose tissue of the same slice of the abdominal wall was measured as the background.The signal-to-noise ratio(SNR)and contrast-to-noise ratio(CNR)were calculated.The three-dimensional VR and MIP images were also obtained for qualitative image quality evaluation.Two observers used the five-point method to evaluate the image quality of VR and MIP,the display of gastric blood vessels and branches.Results There were no statistically significant differences in gender,age and BMI between the two groups(P>0.05).Compared with group A,the average mAs,CTDIvol,DLP and ED of group B were decreased by 27.00%,29.76%,33.64%and 33.76%,respectively.No significant difference was observed among CT values at abdominal aorta,celiac trunk,splenic artery and hepatic artery in groups A and B(P>0.05).A significant difference was observed between groups A and B1 in SD value,SNR and CNR(P<0.05).With the increase of DL level in group B,SD values decreased gradually,SNR values and CNR values increased gradually(P<0.05).Images in B3 and B4 groups had the same SD values,SNR,CNR of vessels,and background noise as group A(all P>0.05).There was no significant difference in SD value,SNR and CNR between B1,B2 and B3 groups(P>0.05).The two observers had good consistency in the evaluation of image quality and the display of gastric blood vessels(Kappa value:0.818-0.860,P<0.05).Each image can satisfy the requirements of diagnosis.There was a statistically significant difference in subjective scores between group A and B1(P<0.05).The subjective scores for displaying gastric blood vessels and their branches in B2-B5groups increased with the increase of DL level initially then gradually decreased(P<0.05).But B4 provided higher subjective scores for displaying perigastric vascular and their branches with the highest subjective score(4.70±0.47,4.47±0.57).Conclusion Deep learning algorithm can effectively reduce image noise and improve CTA image quality under the condition of low radiation dose,among which DL3 is the recommended best reconstruction.