人工智能模型迭代重建对低剂量腹部增强CT图像质量的影响

Effect of artificial intelligence model-based iterative reconstruction on low-dose abdominal enhanced CT image quality

目的:评估人工智能模型迭代重建(AIIR)对低剂量腹部增强CT图像质量的影响,并寻求最佳AIIR等级。方法:对52例患者行低剂量腹部增强CT检查,采用滤波反投影(FBP)重建、混合迭代重建(HIR)及AIIR(5~1级)。评估FBP、HIR及AIIR图像质量的客观参数。两位放射科医师采用盲法按5分法对FBP、HIR及AIIR图像的噪声程度评分(1分为噪声非常小;5分为噪声大不能接受),并按5分法对AIIR图像的蜡状伪影程度评分(1分为无蜡状伪影;5分为完全不能接受的蜡状伪影)。结果:动脉期FBP、HIR及AIIR 5~1级图像上肝脏噪声分别为(22.6±3.1)、(14.6±2.1)、(9.3±0.9)、(7.9±0.8)、(6.4±0.8)、(4.8±0.9)和(3.3±1.1)Hu,静脉期相应图像上肝脏噪声分别为(23.8±3.7)、(15.5±2.4)、(9.7±1.1)、(8.3±0.9)、(6.7±0.8)、(5.0±0.8)和(3.4±1.0)Hu。FBP、HIR及AIIR 5~1级图像上,腰大肌噪声分别为(25.4±4.1)、(16.6±2.7)、(10.3±1.7)、(8.8±1.7)、(7.3±1.7)、(5.6±1.8)和(4.0±2.0)Hu,髂腰肌相应噪声分别为(23.1±4.0)、(15.5±3.2)、(10.2±2.2)、(9.1±2.3)、(7.8±2.5)、(6.5±2.6)和(5.3±2.8)Hu。AIIR图像上肝脏、腰大肌及髂腰肌噪声均明显低于FBP图像和HIR图像(P<0.01)。腹主动脉在FBP、HIR、AIIR 5~1级图像上噪声分别为(27.0±3.9)、(18.0±2.8)、(10.9±1.7)、(9.4±1.6)、(7.7±1.6)、(6.0±1.7)和(4.4±1.9)Hu,相应的信噪比(SNR)分别是(19.4±4.3)、(29.1±6.4)、(48.1±9.1)、(56.2±10.6)、(68.8±13.9)、(90.1±21.5)和(135.0±49.6)。门静脉在FBP、HIR、AIIR 5~1级图像上噪声依次分别是(26.7±4.4)、(17.6±3.5)、(11.2±3.0)、(9.8±3.1)、(8.2±3.3)、(6.6±3.6)和(5.0±3.9)Hu,SNR依次分别是(7.6±2.1)、(11.6±3.4)、(18.4±3.5)、(21.5±4.4)、(26.2±6.0)、(34.6±9.5)和(52.5±21.2)。AIIR图像与FBP图像及HIR图像相比,腹主动脉和门静脉噪声明显减低(P<0.01),SNR明显增加(P<0.01)。医师1动脉期FBP、HIR、AIIR 5~1级图像噪声评分分别为(3.58±0.50)、(2.35±0.48)、(1.42±0.50)、(1.08±0.27)、(1.02±0.14)、(1.00±0.00)和(1.00±0.00),静脉期相应数值分别为(3.63±0.49)、(2.35±0.56)、(1.31±0.47)、(1.08±0.27)、(1.00±0.00)、(1.00±0.00)和(1.00±0.00);医师2动脉期FBP、HIR、AIIR 5~1级图像噪声评分分别为(3.60±0.50)、(2.48±0.51)、(1.52±0.51)、(1.10±0.30)、(1.06±0.24)、(1.06±0.24)和(1.00±0.00),静脉期相应数值分别为(3.65±0.48)、(2.54±0.50)、(1.50±0.51)、(1.17±0.38)、(1.02±0.14)、(1.00±0.00)和(1.00±0.00);5种AIIR图像噪声程度评分均显著低于FBP和HIR图像(P<0.01)。医师1动脉期AIIR 5~1级蜡状伪影程度评分分别为(1.04±0.19)、(1.12±0.32)、(1.92±0.48)、(2.73±0.49)和(4.15±0.54),静脉期分别为(1.02±0.14)、(1.19±0.40)、(2.02±0.42)、(3.10±0.41)和(4.02±0.37);医师2动脉期AIIR 5~1级图像蜡状伪影程度评分分别为(1.21±0.41)、(1.23±0.43)、(2.10±0.41)、(2.88±0.55)和(4.12±0.43),静脉期分别为(1.12±0.32)、(1.33±0.47)、(2.29±0.46)、(3.17±0.38)和(4.00±0.40)。蜡状伪影程度AIIR 5级和4级动脉期及静脉期图像评分显著低于AIIR 3~1级图像(P<0.01),且AIIR 5级与4级图像之间评分无明显差异(P>0.05)。结论:AIIR能够显著提升低剂量腹部增强CT的图像质量,其中5级和4级AIIR图像质量最优。

Objective:To evaluate the effect of artificial intelligence model-based iterative reconstruction(AIIR)on low-dose abdominal enhanced CT image quality and to determine the optimal AIIR level.Methods:Fifty-two patients underwent low-dose abdominal enhanced CT with filter back projection(FBP)reconstruction,hybrid iterative reconstruction(HIR),and AIIR(level 5~1).Objective parameters of image quality for FBP,HIR,and AIIR images were evaluated.Noise was graded by two blinded radiologists for FBP,HIR,and AIIR images on a 5-point scale(1=very low noise;5=unacceptable increased noise).Wax artifacts were also graded by two blinded radiologists for AIIR images on a 5-point scale(1=no wax artifact;5=completely unacceptable wax artifact).Results:On the FBP,HIR,and AIIR(level 5~1)images,the liver noise was(22.6±3.1),(14.6±2.1),(9.3±0.9),(7.9±0.8),(6.4±0.8),(4.8±0.9)and(3.3±1.1)Hu in the arterial phase,and that was(23.8±3.7),(15.5±2.4),(9.7±1.1),(8.3±0.9),(6.7±0.8),(5.0±0.8)and(3.4±1.0)Hu in the venous phase,respectively.On the FBP,HIR,and AIIR(level 5~1)images,the noise of the psoas major muscle was(25.4±4.1),(16.6±2.7),(10.3±1.7),(8.8±1.7),(7.3±1.7),(5.6±1.8)and(4.0±2.0)Hu,while the noise of the iliopsoas was(23.1±4.0),(15.5±3.2),(10.2±2.2),(9.1±2.3),(7.8±2.5),(6.5±2.6)and(5.3±2.8)Hu,respectively.The noise of the liver,psoas major muscle,and iliopsoas on the AIIR images was all significantly lower than that on the FBP and HIR images(P<0.01).The noise of the abdominal aorta was(27.0±3.9),(18.0±2.8),(10.9±1.7),(9.4±1.6),(7.7±1.6),(6.0±1.7)and(4.4±1.9)Hu,while the signal-to-noise ratios(SNRs)of the abdominal aorta were(19.4±4.3),(29.1±6.4),(48.1±9.1),(56.2±10.6),(68.8±13.9),(90.1±21.5)and(135.0±49.6)on the FBP,HIR,and AIIR(level 5~1)images,respectively.The noise of the portal vein was(26.7±4.4),(17.6±3.5),(11.2±3.0),(9.8±3.1),(8.2±3.3),(6.6±3.6)and(5.0±3.9)Hu,while the SNRs of the portal vein were(7.6±2.1),(11.6±3.4),(18.4±3.5),(21.5±4.4),(26.2±6.0),(34.6±9.5)and(52.5±21.2)on the FBP,HIR,and AIIR(level 5~1)images,respectively.Compared with the FBP and HIR images,the AIIR images exhibited lower noise and higher SNRs for the abdominal aorta and portal vein(P<0.01).The noise scores of the FBP,HIR,and AIIR(level 5~1)images were(3.58±0.50),(2.35±0.48),(1.42±0.50),(1.08±0.27),(1.02±0.14),(1.00±0.00)and(1.00±0.00)in the arterial phase and(3.63±0.49),(2.35±0.56),(1.31±0.47),(1.08±0.27),(1.00±0.00),(1.00±0.00)and(1.00±0.00)in the venous phase for radiologist 1;furthermore,they were(3.60±0.50),(2.48±0.51),(1.52±0.51),(1.10±0.30),(1.06±0.24),(1.06±0.24)and(1.00±0.00)in the arterial phase and(3.65±0.48),(2.54±0.50),(1.50±0.51),(1.17±0.38),(1.02±0.14),(1.00±0.00)and(1.00±0.00)in the venous phase for radiologist 2,respectively.The noise scores of the AIIR images were obviously lower than those of the FBP and HIR images(P<0.01).The wax artifact scores of the AIIR(level 5~1)images were(1.04±0.19),(1.12±0.32),(1.92±0.48),(2.73±0.49)and(4.15±0.54)in the arterial phase and(1.02±0.14),(1.19±0.40),(2.02±0.42),(3.10±0.41)and(4.02±0.37)in the venous phase for radiologist 1;moreover,they were(1.21±0.41),(1.23±0.43),(2.10±0.41),(2.88±0.55)and(4.12±0.43)in the arterial phase and(1.12±0.32),(1.33±0.47),(2.29±0.46),(3.17±0.38)and(4.00±0.40)in the venous phase for radiologist 2,respectively.The wax artifact scores of the AIIR images with levels 5 and 4 were obviously lower than those of the AIIR images with levels 3~1(P<0.01),and there was no obvious difference in wax artifact scores between AIIR images with level 5 and level 4.Conclusion:AIIR can greatly increase the image quality of low-dose abdominal enhanced CT,and AIIR images with levels 5 and 4 are optimal.