基于定量影像学生物标志物构建肾透明细胞癌术前精准分级预测模型的可行性

目的：探讨利用CT影像组学技术筛选定量影像学生物标志物进而构建肾透明细胞癌术前WHO/ISUP分级精准预测模型的可行性。方法：回顾性收集2010年1月至2019年10月在丽水市中心医院经手术病理证实且病理分级明确的肾透明细胞癌（ccRCC）患者72例，其中高分化组（I+II级）52例，低分化组（III+IV级）20例。收集入组患者的术前皮髓期CT图像进行影像组学分析筛选影像学生物标志物进而训练预测模型。首先利用ITK-SNAP软件在皮髓期图像手动逐层勾画靶病灶（最大病灶）边缘融合成全病灶三维感兴趣区（VOI），随后利用A.K软件提取纹理参数，并基于R语言筛选特征性纹理参数作为影像学生物标志物，进一步基于上述纹理参数计算各个患者的Rad-score，最终构建ccRCC精准分级预测模型。结果：本研究共提取出396 个纹理特征，并利用Lasso降维联合10倍交叉验证法筛选出了皮髓期CT图像的5个特征性纹理参数作为预测生物标志物，分别为均方根、峰度、相关性、熵、惯性，并计算得到了各个患者相对应的Rad-score。基于Radscore构建ccRCC术前分级预测模型，其曲线下面积为0.891（95%CI =0.797～0.952），敏感度和特异度分别为84.6%和85.3%。结论：基于皮髓期CT图像的影像组学技术构建的术前ccRCC精准分级预测模型具有较高的准确度、特异度和敏感度，利用该方法进行分级预判是可行的。

Feasibility study of preoperative precise graded prediction model for renal clear cell carcinoma based on quantitative imaging biomarkers

Objective: To explore the feasibility of constructing accurate prediction model of WHO/ISUP classification for clear cell renal cell carcinoma (ccRCC) by screened quantitative imaging biomarkers via CT imaging based radiomics. Methods: Retrospective collection of 72 patients with ccRCC who were confirmed by surgery and pathology in Lishui Central Hospital from January 2009 to October 2018. All patients were underwent abdominal CT scanning and enhanced three-phase scanning before surgery, who were divided as highdifferentiation group (I+II grade, 52 cases) and low-differentiation group (III+IV grade, 20 cases).Preoperative arterial CT images of the enrolled patients were collected for screening imaging biomarkers by radiomics analysis and then trained predictive models.Firstly, the ITK-SNAP software was used to manually delineate the target lesion (maximum lesion) edge into the whole lesion three-dimensional region of interest (VOI); secondly, the texture parameters were extracted as imaging biomarkers by A.K software, and the characteristic texture parameters are screened based on R language. The Rad-score of each patient was calculated based on the above texture parameters, and the accurate hierarchical prediction model of renal clear cell carcinoma was finally constructed. Results: A total of 396 texture features were extracted from this study, and five characteristic texture parameters of the CT image of the arterial phase were screened by Lasso reduction dimension combined with10-fold cross-validation method as the predictive biomarkers, which were root mean square (RMS), kurtosis,correlation, entropy, inertia, and calculated the corresponding Rad-score for each patient. The preoperative grading prediction model of renal clear cell carcinoma was constructed based on rad-score, and it was found that the area under the curve (ROC) was 0.891 (95%CI=0.797-0.952), and the sensitivity and specificity were high up to 84.6% and 85.3%, respectively. Conclusion: The preoperative precise grading prediction model constructed by angiography based on arterial CT images has high accuracy, specificity and sensitivity, which is feasible for grading prediction of renal clear cell carcinoma