基于多期CT的定量影像组学特征鉴别透明细胞型肾细胞癌与乏脂肪性血管平滑肌脂肪瘤的价值

The value of quantitative multiple-phase CT radiomic features analysis in differentiation of clear cell renal cell carcinoma from fat-poor angiomyolipoma

目的探讨基于多期CT定量影像组学特征筛选鉴别透明细胞型肾细胞癌(ccRCC)与乏脂肪性血管平滑肌脂肪瘤(fpAML)的CT优势扫描期相及构建最佳分类模型。方法回顾性收集2014年1月至2018年9月195例经病理证实的ccRCC(n=131)及fpAML(n=64)患者的术前CT平扫及动态多期增强扫描(皮髓质期、肾实质期、排泄期)资料,采用ITK-SNAP软件在病灶最大径层面手工勾画ROI后对其进行特征提取,采用两独立样本Mann-Whitney U检验分别筛选出CT各期扫描图像中差异有统计学意义的候选特征集。基于29种特征选择算法,采用后向淘汰原则进行特征选择(前10个),结合合成少数类过采样技术对8种分类器分别进行训练,构建出232个分类模型。对各分类模型的性能进行比较,记录ROC下面积(AUC)、准确率、敏感度和特异度,筛选出鉴别ccRCC与fpAML的CT扫描优势期相及最佳分类模型,并获得关键影像组学特征。结果131例ccRCC病灶的最大径为(3.9±1.4)cm,64例fpAML病灶的最大径为(3.5±1.7)cm,两组病灶的最大径间差异无统计学意义(P>0.05)。初始影像组学特征共102个,经统计学筛选各期图像候选特征集(P<0.05)相应的总特征数量为平扫期(n=26)、皮髓质期(n=71)、肾实质期(n=68)、肾排泄期(n=62)。232个分类模型中,获得最大AUC值的分类模型数在CT各期扫描中所占比例为:平扫期(n=106,45.7%)、皮髓质期(n=94,40.5%)、肾实质期(n=23,9.9%)、排泄期(n=9,3.9%),CT平扫期及皮髓质期的影像组学特征对分类模型的贡献优于肾实质期和排泄期,相应的最佳分类模型分别为SVM-fisher_score与Logistic Regression-RFS,所对应的AUC、准确率、敏感度和特异度分别为0.897、83%、84%、80%及0.891、83%、81%、89%。结论肾脏CT平扫期及皮髓质期的定量影像组学特征对鉴别ccRCC与fpAML的分类模型效能作用明显优于肾实质期及排泄期,通过对不同分类器和特征选择算法的组合筛选出最佳分类模型,对鉴别ccRCC和fpAML具有一定的可行性。

Objective To explore the CT dominant phase and optimal classification model in differenting clear cell renal cell carcinoma (ccRCC) from fat-poor angiomyolipoma (fpAML) through quantitative multiple-phase CT radiomic features analysis. Methods Clinical and imaging data of 195 cases pathologically confirmed ccRCC (n=131) and fpAML (n=64) were retrospectively studied. All the patients underwent non-contrast enhanced CT scans and dynamic multi-phase (corticomedullary phase, medullary phase and excretion phase) contrast-enhanced CT scans. Regions of interest (ROIs) were manually delineated based on the selected image slices with the maximal diameter of the lesion using ITK-SNAP software, followed by the acquisition of candidate CT radiomic feature sets from each phase with statistically significant differences by using Mann-Whitney U test. Then, using the synthetic minority oversampling technique (SMOTE), 232 classification models which are composed of 29 different feature selection algorithms (top 10 features were chosen by the backward elimination method) and 8 different classifiers were constructed. Employing the 5-fold cross-validation method, the performance of each classification models for each phase was evaluated using accuracy (ACC), sensitivity (SEN), specificity (SPE) and area under receiver operating characteristic curve (AUC), to acquire dominant CT phases and the optimal classification models for distingushing ccRCC and fpAML, along with the key imaging radiomic features. Results In this study, the mean maximal diameter of ccRCC and fpAML lesions were (3.9±1.4) cm, and (3.5±1.7) cm, respectively, and there was no statistically significant difference in the size of the tumor between two groups (P>0.05). From 102 initial imaging feature sets, the total number of candidate imaging feature sets (P<0.05) were: non-enhanced phase (n=26), corticomedullary phase (n=71), medullary phase (n=68), excretion phase (n=62). Among the 232 classification models through different combination of classifiers and feature selectors, the amount of classification models which achieved the maximum of AUC value (AUCmax) from different CT phases were: non-enhanced phase (n=106, 45.7%), corticomedullary phase (n=94, 40.5%), medullary phase (n=23, 9.9%), excretion phase (n=9, 3.9%). Imaging features from non-enhanced phase and corticomedullary phase yielded higher performance compared with medullary phase and excretion phase, with the corresponding optimal prediction models were SVM-fisher_score (AUC: 0.897, ACC: 83%, SEN: 84%, SPE:80%) and Logistic Regression-RFS (AUC: 0.891, ACC: 83%, SEN: 81%, SPE:89%), respectively. Conclusions The quantitative imaging features from non-enhanced and corticomedullary phase have better performance among proposed classification models than that from medullary phase and excretion phase. Furthermore, it is feasible to acquire proper combination of feature selection and classifiers to achieve high performance in identifying ccRCC and fpAML.