Objective: To explore the value of CT-based radiomics in differential diagnosis of retroperitoneal neuroblastoma (NB) and ganglioneuroblastoma (GNB) in children.
Methods: A total of 172 children with NB and 48 children with GNB were assigned into the training set and testing set at the ratio of 7∶3 using a random stratified sampling method. Radiomics features were extracted and selected from non-enhanced and post-enhanced CT images. Based on the subset of optimal features, a multivariate regression model was used to establish the radiomics models for each phase and the combined radiomics models. The ROC curves of the models were drawn, and the evaluation indexes such as AUC, accuracy, sensitivity and specificity of these models were calculated and compared.
Results: A total of 1218 radiomics features were extracted from the CT images acquired in non-enhanced (NP), arterial (AP) and venous phases (VP), from which 4 features from the NP model, 3 features from the AP model, 2 features from the VP model and 5 features from the combined model were selected. The AUC of the NP model in the training set and testing set was 0.840 (95% CI: 0.778-0.902) and 0.804 (95% CI: 0.699-0.899), respectively, as compared with 0.819 (95%CI: 0.759-0.877) and 0.815 (95%CI: 0.697-0.915) for the AP model, 0.730 (95%CI: 0.649-0.803) and 0.751 (95%CI: 0.619-0.869) for the VP model, and 0.861 (95%CI: 0.809-0.910) and 0.827 (95%CI: 0.726-0.915) for the combined model.
Conclusion: Radiomics signature based on non-enhanced and post-enhanced CT images can be helpful for distinguishing retroperitoneal NB and GNB in children. Compared with the first-order histogram features, textural features can better reflect the difference of the lesions. NP, AP and VP models have similar classification efficacy in differentiating retroperitoneal NB and GNB. The efficacy of the combined model is similar to that of the NP and AP models, but superior to that of the VP model.
目的: 基于平扫和增强CT的影像组学分析在鉴别儿童腹膜后神经母细胞瘤（NB）和节细胞性神经母细胞瘤（GNB）中的价值。
方法: 纳入172例NB和48例GNB患儿，按7∶3的比例分层随机抽样划分为训练集和测试集。分别从平扫期、动脉期和静脉期CT图像中提取并筛选影像组学特征，基于最优特征子集采用多变量回归模型建立各期以及三期复合的影像组学模型，绘制模型ROC曲线，计算并比较各期模型的AUC、准确度、灵敏度及特异性等评价指标。
结果: 从平扫期、动脉期和静脉期CT图像中分别提取了1218个影像组学特征，最终筛选出平扫期模型4个特征、动脉期模型3个特征、静脉期模型2个特征以及三期复合模型5个特征。平扫期模型在训练集中的AUC为0.840（95%CI: 0.778~0.902），测试集中AUC为0.804（95%CI: 0.699~ 0.899）。动脉期模型在训练集中的AUC为0.819（95%CI: 0.759~0.877），测试集中AUC为0.815（95%CI: 0.697~0.915）。静脉期模型在训练集中的AUC为0.730（95%CI: 0.649~0.803），测试集中AUC为0.751（95%CI: 0.619~0.869）。三期复合模型在训练集中的AUC为0.861（95%CI: 0.809~0.910），测试集中AUC为0.827（95%CI: 0.726~0.915）。
结论: 基于平扫和增强CT的影像组学特征有助于区分儿童腹膜后NB和GNB，纹理特征相对于一阶直方图特征能更好的反映病灶的差异。平扫期、动脉期和静脉期影像组学模型均可较好鉴别儿童腹膜后NB和GNB。三期复合模型与平扫期、动脉期模型效能相似，但优于静脉期模型。
Keywords: children; computed tomography; ganglioneuroblastoma; neuroblastoma; radiomics.