CT影像组学预测肾透明细胞癌核分级:扫描期相及ROI勾画策略  

CT radiomics in prediction of clear cell renal cell carcinoma nuclear grading:exploration of different CT phases and ROI delineation strategies

作　　者：徐向东[1] 罗诗维 韦瑞丽 张婉丽 姚旺 丁文双[2] 庞欣蕊 王晔 杨蕊梦[1] 赖胜圣[4] XU Xiang-dong;LUO Shi-wei;WEI Rui-li(Department of Radiology,Guangzhou First people’s Hospital,Guangzhou 510180,China)

机构地区：[1]广州市第一人民医院放射科,广州510180 [2]广州市第一人民医院病理科,广州510180 [3]华南理工大学医学院,广州510006 [4]广东食品药品职业学院医疗器械学院,广州510520

出　　处：《放射学实践》2022年第12期1542-1547,共6页Radiologic Practice

基　　金：国家自然科学基金面上项目(81971574);广州市医学重点学科建设资金;广州市高水平临床重点专科建设资金;广州市分子影像与临床转化医学重点实验室基金(202201020376)。

摘　　要：目的:探讨不同CT扫描期相及感兴趣区(ROI)勾画策略对影像组学方法预测肾透明细胞癌(ccRCC)核分级效能的影响。方法:回顾性搜集具有完整4期CT扫描图像(平扫期、皮髓质期、实质期和排泄期)且经病理证实为ccRCC的137例患者的病例资料。其中,96例为低级别(Fuhrman 1级和2级)ccRCC,41例为高级别(Fuhrman 3级和4级)ccRCC。在每期图像中选取肿瘤最大层面,使用ITK-SNAP软件分别勾画出病灶最大层面的2D-ROI并获得全瘤3D-ROI,并使用Pyradiomics软件分别提取病灶的影像组学特征。然后,采用22种特征选择方法和8种分类算法对组学特征进行筛选并构建了176个分类模型,使用五折交叉检验法验证各模型的预测效能,并采用诊断符合率、敏感度、特异度和受试者工作特征曲线下面积(AUC)评估模型的预测效能。结果:基于3D-ROI的影像组学模型鉴别高、低核级ccRCC的前5个最大AUC的平均值及相应诊断符合率的平均值高于基于2D-ROI的影像组学模型。基于平扫期的影像组学模型的前5个最大AUC的平均值优于其它3个期相。在平扫、皮髓质期、实质期和排泄期CT图像上基于3D-ROI的组学模型的最大AUC分别为0.822、0.732、0.742和0.780,基于2D-ROI的组学模型的最大AUC分别为0.738、0.692、0.710和0.674。结论:采用影像组学方法预测ccRCC核分级推荐选用平扫图像和全瘤3D-ROI。Objective:The purpose of this study was to evaluate the value of different CT scan phases and different delineation strategies in radiomics methods for the prediction of clear cell renal cell carcinoma(ccRCC)nuclear grading.Methods:137 patients with pathologically proven ccRCCs,including 96 low-grade(Fuhrman grade 1 and 2)and 41 high-grade(Fuhrman grade 3 and 4)ccRCCs,were retrospectively collected in this study.The CT images of four phases[(unenhanced phase(UP),corticomedullary phase(CMP),nephrographic phase(NP)and excretory phase(EP)]were employed in radiomics analysis.The selected axial images of the lesion in each phase were segmented with ITK-SNAP software,obtaining two-dimensional region of interest(2D-ROI)of the tumor with the largest diameter and three-dimensional region of interest(3D-ROI)of the entire tumor.Texture feature extraction was performed on 2D-ROI or 3D-ROI using Pyradiomics software,followed by 176 prediction models constructed with 22 feature selection methods and 8 classification algorithms.Five-fold cross-validation was used to evaluate the efficacy of these models.The discrimination abilities of the models were quantified by area under the receiver operating characteristic(ROC)curve(AUC),accuracy,sensitivity,and specificity.Results:The average of AUCs and corresponding accuracy of the top-5 radiomics models based on 3D-ROI were higher than those of the models based on 2D-ROI,and the ave-rage of the AUCs of the top-5 models based on unenhanced phase was higher than those of the other phases.The maximum AUCs of the 3D-ROI-based radiomics models in unenhanced phase,corticome-dullary phase,nephrographic phase and excretory phase were 0.822,0.732,0.742 and 0.780,respectively;while the maximum AUCs of 2D-ROI-based radiomics models in the four phases were 0.738,0.692,0.710 and 0.674,respectively.Conclusion:3D-ROI and unenhanced phase-based radiomics models are more recommended to be employed to predict ccRCC nuclear grading.

关 键 词：肾肿瘤 透明细胞肾细胞癌 影像组学 体层摄影术 X线计算机 扫描期相 

分 类 号：R814.42[医药卫生—影像医学与核医学] R737.11[医药卫生—放射医学]