多模态MRI影像组学术前预测肾透明细胞癌Fuhrman核分级  

Multimodal MRI radiomics for preoperative predicting Fuhrman nuclear grade of clear cell renal cell carcinoma

作　　者：邢兆宇[1] 沈立文 潘靓[2] 孙军[2] 陈杰[2] 沈楠[2] 俞胜男[2] 邢伟[2] 张龙江 Xing Zhaoyu;Shen Liwen;Pan Liang;Sun Jun;Chen Jie;Shen Nan;Yu Shengnan;Xing Wei;Zhang Longjiang(Department of Urology,Third Affiliated Hospital of Soochow University,Changzhou 213003,China;Department of Radiology,Third Affiliated Hospital of Soochow University,Changzhou 213003,China;Department of Radiology,General Hospital of Eastern Theater Command,Nanjing 210002,China)

机构地区：[1]苏州大学附属第三医院泌尿外科,常州213003 [2]苏州大学附属第三医院影像科,常州213003 [3]解放军东部战区总医院影像科,南京210002

出　　处：《中华放射学杂志》2022年第7期785-791,共7页Chinese Journal of Radiology

基　　金：国家自然科学基金(82171901,82001763);江苏省卫生健康委员会重点项目(K2019023);常州市国际合作科研项目(CZ20210031);常州市卫生健康委员会重大项目(ZD202010);常州市应用基础研究项目(CJ20210147,CJ20200115);常州市卫生健康委员会青苗人才计划(CZQM2020014)。

摘　　要：目的探讨多模态MRI影像组学对肾透明细胞癌(ccRCC)术前Fuhrman核分级的预测价值。方法回顾性分析2011年4月至2021年4月就诊于苏州大学附属第三医院的129例经手术病理证实为ccRCC的患者资料,采用随机数表法按7∶3的比例随机分为训练集(90例)和验证集(39例)。根据术后病理结果,将FuhrmanⅠ、Ⅱ级纳入低级别组(96例,训练集65例、验证集31例),FuhrmanⅢ、Ⅳ级纳入高级别组(33例,训练集25例、验证集8例)。由2名放射科医师于T1WI、T2WI、Dixon纯水相、Dixon纯脂相、磁敏感加权成像(SWI)、血氧水平依赖(BOLD)图像上手动勾画感兴趣区,每个ROI分别提取396个纹理特征。在训练集采用组内相关系数、Mann-Whitney U检验、最小冗余-最大相关法、最小绝对收缩与选择算子法行特征降维,获取最佳纹理特征,采用logistic回归构建多模态影像组学模型,采用受试者操作特征(ROC)曲线评估模型在训练集和验证集中鉴别高、低级别ccRCC的效能。结果共筛选出4个SWI、1个T2WI以及1个BOLD纹理特征用于建模。训练集和验证集中多模态影像组学模型鉴别高、低级别ccRCC的ROC曲线下面积(95%CI)分别为0.859(0.770~0.923)和0.883(0.740~0.964),特异度分别为95.4%和87.1%,灵敏度分别为68.0%和87.5%,准确度分别为87.8%和87.2%。结论基于T2WI、SWI和BOLD图像建立的多模态MRI影像组学模型术前预测ccRCC Fuhrman核分级具有较高的效能。Objective To investigate the value of multimodal MRI radiomics in the preoperative prediction of Fuhrman nuclear grade of clear cell renal cell carcinoma(ccRCC).Methods A total of 129 patients with ccRCC confirmed by pathology from April 2011 to April 2021 in Third Affiliated Hospital of Soochow University were collected,and the imaging and clinicopathological data were retrospectively analyzed.All patients were divided into training set(n=90)and validation set(n=39)at the ratio of 7∶3 using random indicator method.According to the postoperative pathological results,Fuhrman gradesⅠandⅡwere included in the low grade group(96 cases,65 cases in the training set and 31 cases in the validation set),and Fuhrman gradesⅢandⅣwere included in the high grade group(33 cases,25 cases in the training set and 8 cases in the validation set).Two radiologists manually delineated regions of interest(ROI)on T1WI,T2WI,Dixon-water,Dixon-fat,susceptibility weighted imaging(SWI),blood oxygen level dependent(BOLD)images,and 396 texture features were extracted from each ROI.In the training set,intra-class correlation coefficient,Mann-Whitney U test,minimum redundancy maximum relevance and least absolute shrinkage and selection operator method were used to reduce the dimension of features to obtain the best texture features.The logistic regression was used to develop the multimodal radiomics model,and the receiver operating characteristic(ROC)curve was used to evaluate the effectiveness of the model in identifying high and low-grade ccRCC in training set and validation set.Results Four SWI,one T2WI and one BOLD texture features were selected for modeling.The areas under the ROC curve(95%CI)of the multimodal radiomics model for identifying high and low grade ccRCC in the training and validation sets were 0.859(0.770-0.923)and 0.883(0.740-0.964),with the specificity at 95.4%and 87.1%,the sensitivity at 68.0%and 87.5%,the accuracy at 87.8%and 87.2%,respectively.Conclusion The multimodal MRI radiomics model based on T2WI,SWI and BOLD

关 键 词：癌 肾细胞 磁共振成像 影像组学 Fuhrman核分级 

分 类 号：R445.2[医药卫生—影像医学与核医学] R737.11[医药卫生—诊断学]