深度学习图像重建算法在头部非增强扫描中的价值:尸头研究  

The value of deep learning image reconstruction algorithm in non-contrast head CT scans:a cadaver study

作　　者：张凯[1] 刘洪川 张帅 曾令明 李玉明[1] 师珂 李真林[1] ZHANG Kai;LIU Hongchuan;ZHANG Shuai;ZENG Lingming;LI Yuming;SHI Ke;LI Zhenlin(Department of Radiology,West China Hospital of Sichuan University,Chengdu 610041,China;CT Research Center,GE Healthcare China,Shanghai 200000,China)

机构地区：[1]四川大学华西医院放射科,四川成都610041 [2]GE医疗中国CT影像研究中心,上海200000

出　　处：《实用放射学杂志》2022年第1期139-143,共5页Journal of Practical Radiology

基　　金：四川省科技计划项目(2019YFS0522);四川大学华西医院学科卓越发展1·3·5工程项目(ZYGD18019)。

摘　　要：目的 深度学习图像重建算法(DLIR)在头部非增强CT中降噪及降低辐射剂量的可能性研究.方法 采用256排CT对成人尸头进行常规剂量(A组)和低剂量(B组)扫描,2组原始数据分别采用滤波反投影法(FBP)和DLIR低(DLIR-L)、中(DLIR-M)、高(DLIR-H)3档重建.测量组织CT值、噪声(SD)值,计算信噪比(SNR)、对比噪声比(CNR);同时对图像噪声和锐利度进行主观评分.比较2组辐射剂量及不同重建算法对图像质量的影响.结果 与FBP相比,DLIR显著降低各组织的SD值(下降14.52%~43.66%),并增加SNR(增加22.07%~82.39%)(P<0.05),且随DLIR重建档位升高,SD值降低,SNR和CNR升高.2组DLIR图像的噪声评分均明显高于FBP.B组DLIR图像的锐利度评分均明显高于FBP(P<0.05).与A组FBP相比,除B组DLIR-M灰质的SD值和SNR值外,其余组织在DLIR-M和DLIR-H的SD值均明显降低而SNR值均明显升高(P<0.05);B组DLIR-M和DLIR-H的噪声得分明显高于A组FBP重建而且锐利度无显著降低.B组容积CT剂量指数(CTDIvol)较A组降低了31.74%.结论 DLIR在常规剂量和低剂量模式下都可以降低头部各组织噪声,增加图像SNR.DLIR-M、DLIR-H档可以显著提高低剂量模式获得的图像质量.Objective To explore the feasibility of deep learning image reconstruction(DLIR)algorithm to reduce image noise and radiation dose in non-contrast head CT seanning.Methods Adult cadaver heads were scanned with a 256-row CT scanner using two modes:routine dose(mode A)and low dose(mode B).For each mode,the raw data were reconstructed with DLIR(low,medium,and high)and filtered back projection(FBP).CT value and standard deviation(SD)value of tissues were measured to calculate the signal-to-noise ratio(SNR)and contrast-to-noise ratio(CNR).Also,the image noise and sharpness were evaluated by radiologists.The radiation dose of two modes and the image quality of different algorithms were compared.Results DLIR significantly reduced the SD values(14.52%-43.66%)and increased SNR(22.07%-82.39%)(P<0.05)compared with FBP.SD values decreased and SNR,CNR increased with the increment of DLIR level.The noise scores of DLIR in each mode were significantly higher than FBP.The sharpness scores of DLIR were significantly higher than FBP(PV0.05)in mode B.Compared with FBP of mode A,SD values of all tissues in DLIR-M and DLIR H in mode B were significantly lower,and the SNR significantly higher(P<0.05).except for the SD value and SNR of grey matter in DLIR-M.The noise scores in DLIR-M and DLIR H in mode B were significanily higher than FBP in mode A while no statistical difference in sharpness scores.Volume CT dose index(CTDIvol)of mode B was reduced by 31.74% compared with mode A.Conclusion DLIR can reduce the noise and increase the SNR in both routine and low dose modes.DLIR-MH can significantly improve the image quality under low dose mode.

关 键 词：计算机体层成像 深度学习重建算法 头部 尸体 

分 类 号：R814.42[医药卫生—影像医学与核医学] R816.1[医药卫生—放射医学]