基于增强CT的影像组学用于术前儿童肾母细胞瘤病理分型鉴别可行性分析

To Study the Feasibility of Radiomics Based on Contrast-Enhanced CT for Preoperative Pathological Classification of Wilms Tumor in Children

摘要: 目的：本实验为探讨基于增强CT的影像组学分析对术前儿童肾母细胞瘤病理分型预测的可行性，从而避免活检及术前化疗对于患儿病理分型、临床分期的干扰，为临床治疗提供更加准确的指导。方法：通过回顾性分析113患儿腹部增强CT影像学资料，勾画ROI区间，并提取944种影像组学特征，采用LASSO回归进行影像组学特征筛选，用筛选后的组学特征分别建立SVM、随机森林、Logistic预测模型。绘制受试者工作特征曲线(ROC)评价其预测效能。结果：最终三种预测模型ROC曲线下面积(AUC)分别为随机森林0.934、SVM 0.869、Logistic 0.739，三者两两间行Delong检验，p值均小于0.05，提示两两之间有显著差异。结论：基于增强CT影像学资料，影像组学用于术前鉴别肾母细胞瘤病理分型可行，其中随机森林算法效果最佳，该方法可为患儿的个性化诊疗提供决策支持。

Abstract: Objective: To explore the feasibility of contrast-enhanced CT-based radiomics analysis in preopera-tive prediction of the pathological classification of nephroblastoma in children, to avoid the inter-ference of biopsy and preoperative chemotherapy on the pathological classification and clinical staging of children and to provide more accurate guidance for clinical treatment. Methods: The ab-dominal contrast-enhanced CT imaging data of 113 children were retrospectively analyzed, ROI in-tervals were delineated, and 944 radiomics features were extracted. LASSO regression was used to select radiomics features, and SVM, Random Forest, and Logistic prediction models were built. The receiver operating characteristic (ROC) curve is drawn to evaluate the predictive efficacy. Results: The areas under the ROC curve (AUC) of the three prediction models were 0.934 for random forest, 0.869 for SVM, and 0.739 for Logistic. Delong test was conducted in pairs of the three. Long runs showed significant differences between any two of the three predictive models (all p < 0.05). Con-clusions: Based on contrast-enhanced CT imaging data, radiomics can be used to identify pathologi-cal types of Wilms tumor prior to surgery, with the Random Forest algorithm having the best per-formance. This approach can provide decision support for personalized diagnosis and treatment of children.