摘要
目的 通过比较噪声功率谱、基于任务的传递函数以及病变检测能力,评估深度学习图像重建(DLIR)对提升图像质量及降低辐射剂量的潜力。方法 使用GE Revolution APEX CT扫描ACR464体模并设定8个不同噪声指数(NI=10、14、16、18、20、22、24、28),原始数据采用滤波反投影(FBP)、多模型迭代重建算法(ASiR-V)40%、ASiR-V 60%、ASiR-V 80%以及不同级别深度学习图像重建(DLIR-L、DLIR-M、DLIR-H)算法进行图像重建。通过使用imQuest软件计算不同重建算法图像噪声功率谱(NPS)、基于任务的传递函数(TTF)以及检测能力指数(d')评估图像质量。结果 在所有重建算法中,DLIR-H的NPS peak最低。随着噪声指数的增加,NPS fav均向低频率移动;DLIR-H的fav(0.24~0.27 mm^(-1))仅低于ASiR-V40%(0.26~0.28 mm^(-1))。TTF_(50%)值不受DLIR级别的影响;TTF_(50%)值较ASiR-V 60%和80%分别高出(37.44±10.85)%、(46.24±15.28)%。深度学习图像重建的大小特征检测能力均高于ASiR-V 40%;比较DLIR-H与NI=10时ASiR-V 40%病灶检测能力相当的辐射剂量,小特征辐射剂量减少约76.48%,大特征减少约72.59%。结论 深度学习图像重建不仅可以在不改变噪声纹理的情况下降低噪声、提高空间分辨率和病灶可检测性,而且具有较ASiR-V更加强大的降低辐射剂量的能力。
Objective To evaluate the potential of deep learning image reconstruction(DLIR)in improving image quality and reducing radiation dose by comparing the noise power spectrum,task-based transfer function and lesion detection capability.Methods The ACR464 phantom was scanned using GE Revolution APEX CT and eight different noise indices(NI=10,14,16,18,20,22,24,28)were set.The original data were subjected to image reconstruction using filtered back-projection(FBP),multi-model iterative reconstruction algorithms(ASiR-V)at 40%,ASiR-V at 60%,ASiR-V at 80%,and different levels of deep learning image reconstruction(DLIR-L,DLIR-M,DLIR-H)algorithms.The image quality was evaluated by using imQuest software to calculate the noise power spectrum(NPS),task-based transfer function(TTF),and detection capability index(d')of different reconstruction algorithms.Results Among all the reconstruction algorithms,the NPS peak of DLIR-H was the lowest.With the increase of noise index,both NPS and fav move towards low frequencies.The fav of DLIR-H(0.24-0.27 mm^(-1))was only 40%lower than that of ASiR-V(0.26-0.28 mm^(-1)).The TTF50%value was not affected by the DLIR level.The TTF50%value was(37.44±10.85)%and(46.24±15.28)%higher than that of ASiR-V60%and 80%,respectively.The detection ability of both large and small features in deep learning image reconstruction was 40%higher than that of ASiR-V.When comparing the radiation doses with comparable lesions detection capabilities of 40%ASiR-V at NI=10 and DLIR-H,the radiation dose for small features decreased by approximately 76.48%,and that for large features decreased by approximately 72.59%.Conclusion Deep learning image reconstruction can not only reduce noise,improve spatial resolution and lesion detectibility without changing noise texture,but also has more powerful ability to reduce radiation dose than ASiR-V.
作者
樊丽华
李明
贾永军
韩冬
于勇
郑运松
魏伟
FAN Lihua;LI Ming;JIA Yongjun;HAN Dong;YU Yong;ZHENG Yunsong;WEI Wei(Department of Medical Imaging,Affiliated Hospital of Shaanxi University of Traditional Chinese Medicine,Xianyang 712000,China;School of Medical Technology,Shaanxi University of Traditional Chinese Medicine,Xianyang 712046,China)
出处
《分子影像学杂志》
2025年第9期1064-1070,共7页
Journal of Molecular Imaging
基金
陕西省重点研发计划项目(2024SF-YBXM-524)。
关键词
深度学习图像重建
辐射剂量
图像质量
体模
deep learning image reconstruction
radiation dose
image quality
phantom
