基于差异化网络分析的CT平扫期影像组学模型可有效预测及解释肾透明细胞癌WHO/ISUP核分级

An unenhanced CT radiomics model using differential network analysis could predict and explain WHO/ISUP grade of clear cell renal cell carcinoma effectively

目的探讨基于差异化网络分析特征选择的CT影像组学模型预测肾透明细胞癌世界卫生组织/国际泌尿病理学会(WHO/ISUP)核分级的可行性,以及探讨模型特征的可解释性。方法回顾性分析175例肾透明细胞癌患者的CT图像以及临床病理资料,其中训练集105例,测试集70例。使用ITK-SNAP和PyRadiomics软件平台对肿瘤的图像进行分割和特征提取。在训练集中,通过差异化网络分析进行特征选择,并构建不同期相的WHO/ISUP核分级预测模型。选择性能最佳的期相模型,并与其他机器学习模型和临床模型比较检验其性能。通过K-M生存分析、Cox回归分析以及风险评分,对最佳期相模型中的特征进行重要性评估,以验证其可解释性。结果平扫期模型的预测效能最好,验证集中的曲线下面积均为0.76,优于其他机器学习模型以及临床模型(P<0.05)。K-M生存分析、Cox回归分析以及风险评分分析结果均显示平扫期模型中的影像组学特征可预测无进展生存期。结论基于差异化网络分析特征选择构建的平扫期模型不仅可以有效预测肾透明细胞癌的WHO/ISUP核分级,同时该模型的特征也能很好的预测无进展生存期,有较强的可解释性。

Objective To predict the World Health Organization/International Society of Urological Pathology(WHO/ISUP)grade of clear cell renal cell carcinoma(ccRCC)using a CT radiomics model based on differential network analysis feature selection and to explore the interpretability of features.Methods A total of 175 ccRCC patients with CT and clinicopathologic data were enrolled in the study,which included a training set of 105 patients and a test set of 70 patients.Segmentation images and feature extraction were performed using ITK-SNAP and PyRadiomics software.The WHO/ISUP grade prediction model was constructed using differential network analysis feature selection with different phases in the training set.The best phase model was then compared with other machine learning models and clinical models.The importance of features in the best phase model was evaluated using K-M survival analysis,Cox regression analysis and risk scores to verify their interpretability.Results The non-contrast phase model was identified as the best phase model,with an AUC of 0.76 in the validation set,performing better than other machine learning models and clinical models(P<0.05).Kaplan-Meier survival analysis,Cox regression analysis and risk score analyses revealed that the non-contrast phase model could predict progression-free survival well.Conclusion The differential network analysis feature selection was used to select the non-contrast phase model,which could accurately predict the WHO/ISUP grade of ccRCC.The characteristics of this model could predict progression-free survival with strong interpretability.