深度学习重建算法对肝脏图像质量及肝转移瘤诊断的研究  

Study on Deep Learning Image Reconstruction in Image Quality and Clinical Diagnosis for Patients with Liver Metastases

作　　者：许艺馨 李辉坚 王国华[1] 王铭君 张振 XU Yixin;LI Huijian;WANG Guohua;WANG Mingjun;ZHANG Zhen(Department of Radiology,Qingdao Municipal Hospital Affiliated to Qingdao University,Qingdao Shandong 266011,China;Department of Radiology,Qingdao No.8 People’s Hospital,Qingdao Shandong 266121,China;CT Research Centre,GE China,Shanghai 200240,China)

机构地区：[1]青岛大学附属青岛市市立医院放射科,山东青岛266011 [2]青岛市第八人民医院放射科,山东青岛266121 [3]GE中国CT影像研究中心,上海200240

出　　处：《中国医疗设备》2021年第10期28-31,39,共5页China Medical Devices

摘　　要：目的对比常规滤波反投影(Filter Back Projection,FBP)和基于多模型迭代重建(Adaptive Statistical Iterative Reconstruction-Veo,ASiR-V)算法,评估深度学习图像重建(Deep Learning Image Reconstruction,DLIR)算法对提高肝转移瘤患者CT图像质量和诊断信心的临床应用可能性。方法连续收集肝转移瘤32例患者,行上腹部动态增强CT扫描,门脉期原始数据使用0%ASiR-V(FBP)、30%ASiR-V及3个重建强度的DLIR(L、M、H)进行重建。分别测量5组重建图像的肝病灶与肝组织CT值、同层面右侧椎旁肌肉噪声标准差(Standard Deviation,SD),并计算信噪比(Signal to Noise Ratio,SNR)和对比噪声比(ContrasttoNoiseRatio,CNR)。同时由2名放射科医师在图像质量、病灶显示及诊断信心方面对5组图像进行主观评分。结果图像质量方面,DLIR组客观图像质量指标均优于FBP及30%ASiR-V(P<0.001)。DLIR-H噪声最低(6.35±1.27),较FBP和30%ASiR-V分别减少70%、53%;进而在肝脏处的SNR(16.32±5.1)较FBP和30%ASiR-V分别增加了64%、53%;病灶处的SNR(10.09±4.16)较FBP和30%ASiR-V分别增加了66%、54%;病灶CNR(6.21±2.61)较FBP和30%ASiR-V分别增加了65%、53%(P<0.001)。病灶诊断信息方面,对于病灶显示及诊断信心的评分显示,DLIR得分均高于FBP、30%ASiR-V。因DLIR-H致细小结构模糊以及过度平滑,所以DLIR-M得分最高。结论DLIR与ASiR-V和FBP相比能显著降低图像噪声、提高图像质量,并提高病灶细节和诊断信心。Objective To compare the conventional filter back projection(FBP)and adaptive statistical iterative reconstructionveo(ASIR-V),and to evaluate the clinical possibility of deep learning image reconstruction(DLIR)algorithm for improving CT image quality and diagnostic confidence in patients with liver metastases.Methods A total of 32 patients with liver metastasis were continuously collected and underwent dynamic enhanced CT scanning of the upper abdomen.Raw data of portal vein phase were reconstructed using 0%ASir-V(FBP),30%ASir-V and DLIR(L,M,H)of three reconstruction intensities.The CT values of the lesion and liver tissue as well as the image noise standard deviations(SD)of right erector spinae were measured in the five groups of reconstructed images.The signal to noise ratio(SNR)and contrast to noise ratio(CNR)were calculated.At the same time,two radiologists scored the five groups of images subjectively in terms of image quality,lesion display and diagnostic confidence.Results In terms of image quality,the,objective image quality indexes of DLIR group were better than FBP and 30%ASiR-V(P<0.001).DLIR-H had the lowest noise(6.35±1.27),which was 70%and 53%lower than FBP and 30%ASIR-V,respectively.Compared with FBP and 30%ASir-V,SNR in liver(16.32±5.1)increased by 64%and 53%,respectively.The SNR of lesion(10.09±4.16)was increased by 66%and 54%compared with FBP and 30%ASiR-V.CNR of lesion(6.21±2.61)was increased by 65%and 53%compared with FBP and 30%ASiR-V(P<0.001).In terms of diagnosis information of lesions,DLIR scores were higher than FBP and 30%ASIR-V for lesion display and diagnostic confidence.DLIR-M had the highest score because DLIR-H caused fine structure blur and excessive smoothness.Conclusion Compared with ASiR-V and FBP,DLIR can significantly reduce image noise,improve image quality,and improve the details and diagnostic confidence of lesions.

关 键 词：深度学习图像重建算法 迭代重建 滤波反投影 肝转移瘤 图像质量 

分 类 号：R814.42[医药卫生—影像医学与核医学] TP391[医药卫生—放射医学]