深度学习CT超高分辨图像组学特征稳定性研究  

Study on Radiomics Features Stability of CT Super-Resolution Images Based on Deep Learning

作　　者：孙瑶 王祥 彭思粤 潘博洋 李清楚 陈如潭 龚南杰 萧毅 刘士远 SUN Yao;WANG Xiang;PENG Siyue;PAN Boyang;LI Qingchu;CHEN Rutan;GONG Nanjie;XIAO Yi;LIU Shiyuan(Department of Imaging Diagnosis,Second Affiliated Hospital,Naval Medical University;Research and Development Department,RadioDynamic Healthcare;Inter national Innovation Center of Tsinghua University)

机构地区：[1]海军军医大学第二附属医院放射诊断科 [2]苏州影动医疗科技有限公司研发部 [3]清华国际创新中心

出　　处：《中国医学计算机成像杂志》2022年第2期196-202,共7页Chinese Computed Medical Imaging

基　　金：国家自然科学基金(82001812)。

摘　　要：目的:构建基于深度学习的胸部CT图像超高分辨率模型,并评价生成的超高分辨CT (SRCT)图像质量,探讨其对放射组学特征稳定性的影响。方法:对2020年10月至2021年5月期间行肺结节复查的53例患者进行了512×512矩阵高分辨率CT (HRCT)、1 024×1 024矩阵超高分辨率CT (UHRCT)靶扫描,基于24例患者的7 680幅HRCT图像、2 160幅UHRCT图像构建了基于深度学习的胸部CT图像超分辨率模型,并生成SRCT图像。3名放射科医生使用利克特5分法评分表对29名患者的SRCT、HRCT和UHRCT图像的噪声、条纹伪影、结节边缘、小血管清晰度、正常肺实质的均匀性和整体图像质量进行主观视觉评估,并使用组学软件提取29名患者的SRCT、HRCT和UHRCT图像肺结节的放射组学特征。采用单因素重复测量方差分析比较SRCT、HRCT和UHRCT图像的噪声、条纹伪影、结节边缘、小血管清晰度、正常肺实质的均质性和整体图像质量。使用组内相关系数(ICC)计算3名放射科医生在SRCT、HRCT和UHRCT图像上肺结节的组学特征一致性。结果:对于噪声和条纹伪影的评分由高到低依次为HRCT、SRCT、UHRCT (P<0.05);对于小血管清晰度和整体图像质量的评分由高到低依次为SRCT、UHRCT、HRCT (P<0.05);对于结节边缘和正常肺实质的均质性的评分SRCT与UHRCT无差异(P>0.05),但明显优于HRCT (P<0.05)。SRCT与HRCT、UHRCT图像的组学特征总体较为一致,SRCT和HRCT的组学特征ICC平均值最高(3名医生结果分别为0.90、0.90、0.93)。结论:该研究构建的深度学习模型显著提高了HRCT图像质量,生成的SRCT图像的组学特征稳定性较好。Purpose: The super-resolution model of chest CT images based on deep learning was constructed, and the quality of the generated super-resolution CT(SRCT) was evaluated, and its influence on the stability of radiomics features was explored. Methods: Fifty-three patientswho underwent follow-up for pulmonary nodules between October 2020 and May 2021 underwent high-resolution CT(HRCT) with a 512×512 matrix and ultra-high resolution CT(UHRCT) target scanning with a 1024×1024 matrix. Based on 7 680 HRCT images and 2 160 UHRCT images of24 patients, a chest CT image super-resolution model based on deep learning was constructed and SRCT images were generated. Three independent radiologists visually evaluated the HRCT, UHRCT and SRCT images for noise, streak artifact, nodular edge, clarity of small vessels, homogeneity of the normal lung parenchyma, and overall image quality using 5-point Likert scoring system. Radiomics features of pulmonary nodules in the HRCT, UHRCT and SRCT images were extracted by using radiomics software. Single-factor repeated measure ANOVA was used to compare noise, streak artifact, nodular edge, clarity of small vessels, homogeneity of the normal lung parenchyma,and overall image quality between HRCT, UHRCT and SRCT images. Intraclass correlation coefficient(ICC) was used to calculate consistency of radiomics features of pulmonary nodules on HRCT, UHRCT and SRCT images of three radiologists. Results: The scores of noise and streak artifact were from high to low in turn for HRCT, SRCT,and UHRCT(P<0.05). The scores of clarity of small vessels and overall image quality were from high to low in turn for SRCT, UHRCT, and HRCT(P<0.05). SRCT has no different from UHRCT(P>0.05), but was much better performed than HRCT(P<0.05) in nodular edge and homogeneity of the normal lung parenchyma. The radiomics features of SRCT, HRCT and UHRCT were generally consistent. The ICC of SRCT and HRCT images had the highest mean ICC in three comparison groups for all the radiomics features(0.90, 0.90, and 0.93 for three

关 键 词：深度学习 肺结节 计算机体层成像 放射组学 

分 类 号：R445.3[医药卫生—影像医学与核医学]