基于影像组学的儿童神经母细胞瘤风险分级预测研究

Radiomics-based risk prediction research in pediatric neuroblastoma

关键词:  影像组学  神经母细胞瘤  风险分级模型

Key words:  radiomics  neuroblastoma  risk stratification model

摘要:

目的 本研究旨在基于对比增强CT(Contrast enhanced computed tomography,CECT)图像影像组学技术,寻找能够有效预测NB风险分级的关键指标,并构建相应的预测模型,助力NB的精准用药。方法 本研究采用回顾性方法收集2018年8月至2023年12月浙江大学医学院附属儿童医院收治的的患者,收集患者的临床信息、影像资料、实验室检测结果等,并使用R包Pyradiomics从CECT图像中提取影像组学特征,由影像科医生利用CECT图像统计影像学危险因子(image defined risk factor,IDRF)。将影像组学特征、IDRF特征和临床信息特征经过筛选后分别利用逻辑回归、支持向量机和随机森林算法建立风险分级预测模型。利用AUC值、DCA曲线和校正曲线等指标评估模型性能,并构建诺模图。结果 本研究最终纳入111例NB患者,其中高危组71人,中低危组40人。通过分析影像组学特征、IDRF特征和临床信息特征,最终筛选出7个影像组学特征、4个IDRF特征和4个临床特征分别构建影像组学模型、临床预测模型和IDRF模型。影像组学模型的AUC值介于0.82到0.85之间,临床预测模型的AUC值介于0.87到0.96之间,而IDRF模型的AUC值介于0.80到0.81之间。DCA曲线和校正曲线分析显示,影像组学模型与IDRF模型的联合应用在稳定性和临床获益方面均得到了显著提升。结论 本研究构建的影像组学模型在NB风险分级预测上展现出良好的预测能力,为NB的个体化用药提供了重要参考依据。

Abstract:

OBJECTIVE This study aims to use radiomics technology based on contrast - enhanced CT (CECT) images. It seeks to identify key indicators that can effectively predict neuroblastoma (NB) risk stratification. Then, it will build corresponding prediction models to help with precise drug use for NB. METHODS In this study, a retrospective method was adopted to collect patients admitted to Children's Hospital Zhejiang University School of Medicine from August 2018 to December 2023. The clinical information, imaging data, laboratory test results, etc. of the patients were collected, and the radiomics features were extracted from the CECT images using the R-package Pyradiomics. The image defined risk factor (IDRF) was statistically analyzed by radiologists using CECT images. After screening the radiomics features, IDRF features and clinical information features, the risk classification prediction models were established respectively by using logistic regression, support vector machine and random forest algorithm. Model performance was evaluated using AUC values, decision curve analysis (DCA) curves, and calibration curves, and nomograms were developed. RESULT A total of 111 NB patients were finally included in this study, among whom 71 were in the high-risk group and 40 were in the medium-low risk group. By analyzing the radiomics features, IDRF features and clinical information features, finally 7 radiomics features, 4 IDRF features and 4 clinical features were screened out to construct the radiomics model, clinical prediction model and IDRF model respectively. The AUC values of the radiomics model ranged from 0.82 to 0.85, those of the clinical prediction model ranged from 0.87 to 0.96, and those of the IDRF model ranged from 0.80 to 0.81. The analysis of DCA curves and correction curves shows that the combined application of the radiomics model and the IDRF model has significantly improved in terms of stability and clinical benefits. CONCLUSION The radiomics model developed in this study shows good predictive ability for NB risk stratification, offering a key reference for personalized NB drug therapy.