基于增强CT影像组学联合机器学习鉴别均质性肾透明细胞癌与肾乏脂肪血管平滑肌脂肪瘤

Application of enhanced CT radiomics in differentiation of renal angiomyolipoma without visible fat from homogeneous clear cell renal cell carcinoma

目的使用基于增强CT影像组学特征联合机器学习鉴别均质性肾透明细胞癌与肾乏脂肪血管平滑肌脂肪瘤。方法回顾性分析术后病理证实的均质性肾透明细胞癌26例,肾乏脂肪血管平滑肌脂肪瘤22例。CT图像手工勾画肿瘤感兴趣区,提取组学特征,数据经过归一化及空间降维,筛选特征分别建立支持向量机(SVM)、逻辑回归(LR)、随机森林(RF)模型,进行5倍交叉验证,选取交叉验证集AUC最高的模型为最佳模型。分析临床特征确定预测因子并建立临床预测模型,利用临床所选预测因子和最佳组学模型预测值建立综合模型并绘制列线图。以Hosmer-Lemeshow拟合优度检验评价列线图的拟合度。绘制决策曲线评价列线图的净获益。结果最佳组学模型为LR模型,经bootstrap法内部验证模型AUC值为0.836(95%CI:0.701~0.927)。综合模型AUC值为0.869(95%CI:0.740~0.949),列线图校正曲线具有良好的一致性,模型的决策曲线也获得了良好的净获益。结论结合临床特征与影像组学特征具有较好的鉴别肾均质性透明细胞癌与肾乏脂肪血管平滑肌脂肪瘤能力。

Objective Using the enhanced CT radiomics in differentiation of renal angiomyolipoma without visible fat from homogeneous clear cell renal cell carcinoma.Methods Retrospective analysis was performed on 28 cases of homogeneous clear cell renal cell carcinoma and 22 cases of renal angiomyolipoma without visible fat,which were confirmed by postoperative pathology.The arterial and venous phase images manual sketch tumor area were concerned,About 1706 radiomics features,data after normalization and by space dimension reduction,were extracted from the image.Features were screened to establish LR,SVM,RF model respectively,five-fold cross validation and selecting cross validation were used to set the highest AUC model as the best model.The clinical features were analyzed to determine the predictors and the clinical model.The comprehensive model was established by multi-factor binary Logistic regression based on the predicted values of the selected optimal radiomics model and clinical predictors.The nomogram based on combined with clinical factor and radiomics model.The test of Hosmer-Lemeshow was used to evaluate the fitness of the line chart.Decision curve analysis was applied for clinical use.Results The best radiomics model was LR model,and the AUC of the model verified by Bootstrap method was 0.836(95%CI:0.701~0.927).The AUC value of the comprehensive model was 0.869(95%CI:0.701~0.927),and the calibration curve of the nomogram showed good consistency.Decision curve analysis verified the clinical usefulness of the predictive nomogram.Conclusion The combination of clinical factors and radiomics features has a strong ability to distinguish renal angiomyolipoma without visible fat from homogeneous clear cell renal cell carcinoma.