深度学习图像重建算法对图像质量的影响:体模研究  

Impact of deep learning reconstruction algorithm on image quality:a phantom study

作　　者：付芮 张成亮 沈云 陈大治[3] 町田治彦 蔡磊 FURui;ZHANGCheng-liang;SHEN Yun(Department of Radiology,Yinchuan Hospital of Traditional,Yinchuan 750001,China)

机构地区：[1]银川市中医医院放射科,宁夏银川750001 [2]GEHC中国CT影像研究中心(CTRC) [3]宁夏回族自治区人民医院放射科,宁夏银川750001 [4]東京女子医科大学附属足立医療センター放射線科,123-8558

出　　处：《放射学实践》2023年第12期1617-1622,共6页Radiologic Practice

摘　　要：目的:基于体模研究探讨深度学习图像重建(DLIR)算法的图像质量是否优于传统的自适应统计迭代重建算法(ASIR-V)。方法:使用GE Revolution Apex CT机对分辨率体模(JIS体模)进行常规剂量(CTDI为11.50 mGy;A组)和低剂量(CTDI为4.22 mGy;B组)扫描。对A组的原始扫描数据采用常规30%权重ASIR-V进行图像重建,B组采用滤波反投影(FBP)、30%-ASIR-V、60%-ASIR-V、90%-ASIR-V及不同级别的深度学习重建(DLIR-L、DLIR-M、DLIR-H)算法进行图像重建。在每组重建图像上测量JIS体模试管内各ROI(圆形,大小135 mm2)的CT值及其SD(图像噪声),计算图像的信噪比(SNR);同时由3位放射科医师对扫描图像的噪声和分辨率(空间分辨率/密度分辨率)两个方面对不同重建算法的图像采用5分法(5分,图像清晰;4分,图像基本清晰;3分,满足临床诊断要求;2分,图像部分勉强辨认;1分,完全不能辨认)进行主观评分。采用单因素方差分析比较不同重建算法图像上测得的CT值、SD及SNR,采用Fleiss Kappa检验比较医师之间主观评分的一致性。结果:B组的CTDI为4.22 mGy,较A组(11.50 mGy)下降了63.3%。在相同权重迭代算法(30%-ASIR-V)下,低剂量与常规剂量相比图像的SD升高了68.33%,低剂量图像噪声和图像分辨率的主观评分低于常规剂量。在低剂量组中,FBP、30%-ASIR-V、60%-ASIR-V、90%-ASIR-V及DLIR-L、DLIR-M、DLIR-H重建图像的SD分别为31.33±2.51、25.21±2.47、19.43±1.76、13.76±1.39和21.33±2.09、21.33±2.09、13.84±1.42,与FBP算法相比,不同权重迭代算法和不同级别的DLIR算法均可减低图像噪声(P<0.05),且随着ASIR-V权重或DLIR级别的升高,SD值越低。图像分辨率主观评分随着ASiR-V权重的升高而降低,但随着DLIR级别的升高而升高。与FBP、30%-ASIR-V、60%-ASIR-V和DLIR-L、DLIR-M算法相比,DLIR-H算法可显著降低图像噪声(P均<0.05);DLIR-H算法的图像噪声和分辨率的主观评分高于其它算法(P均<0.05)。结论:应�Objective:To compare the impact of deep learning iterative reconstruction(DLIR)algorithm with adaptive statistical iterative reconstruction algorithm veo(ASIR-V)on image quality in a phantom study.Methods:The GE Revolution Apex CT was used to perform routine dose(CTDI of 11.5mGy;Group A)and low dose(CTDI of 4.22mGy;Group B)scans on the JIS phantom.The original scan data from Group A was reconstructed using the conventional 30%-weighted ASIR-V;while in Group B,filtered back projection(FBP),30%-ASIR-V,60%-ASIR-V,90%-ASIR-V,and low,medium and high levels of DLIR(DLIR-L,DLIR-M,DLIR-H)was used for image reconstruction.The CT va-lues and their standard deviation(SD)values(image noise)were measured within the ROIs(area of 135mm 2)in the JIS phantom tubes on the images with each reconstruction algorithm in both groups,and the signal-to-noise ratio(SNR)was calculated.Three radiologists subjectively scored the images based on noise and resolution(spatial/contrast resolution)using a 5-point scale(5 points=very good;4 points=generally good;3 points=meeting clinical diagnostic requirements;2 points=partially recognizable;1 point=completely unrecognizable).One-way ANOVA was used to compare the CT values,SD values,and SNR of the images reconstructed using different algorithms,with a significance level of P<0.05 indicating a statistically significant difference.Fleiss Kappa test was used to assess the consistency of the subjective scoring.Results:The CTDI for Group B was 4.22mGy,which decreased by 63.3%compared to Group A(11.5mGy).Under the same weighted iterative algorithm(30%-ASIR-V),the SD of the images was elevated by 68.33%in the low dose compared with the conventional dose,and the subjective scores of image noise and image resolution were lower in the low dose than in the conventional dose.In the low-dose group,the SD of FBP,30%-ASIR-V,60%-ASIR-V,90%-ASIR-V,DLIR-L,DLIR-M and DLIR-H reconstructed images were 31.33±2.51,25.21±2.47,19.43±1.76,13.76±1.39,21.33±2.09,21.33±2.09 and 13.84±1.42,respectively.Compared with th

关 键 词：体层摄影术 X线计算机 图像质量 深度学习 图像重建 低剂量 

分 类 号：R814.42[医药卫生—影像医学与核医学]