Radiomics features based on automatic segmented MRI images: Prognostic biomarkers for triple-negative breast cancer treated with neoadjuvant chemotherapy
Radiomics features based on automatic segmented MRI images: Prognostic biomarkers for triple-negative breast cancer treated with neoadjuvant chemotherapy
复制标题
DOI：
10.1016/j.ejrad.2021.110095
复制
发表时间：
2022-01-01
影响因子：
3.3
通讯作者：
Wang, Xiaoying
中科院分区：
文献类型：
--
作者：
Ma, Mingming;Gan, Liangyu;Wang, Xiaoying
目的：建立基于自动分段磁共振成像（MRI）的放射组学预测模型，用于预测新辅助化疗（NAC）后三阴性乳腺癌（TNBC）患者的全身复发。材料与方法：本研究共纳入2009年1月至2018年12月期间接受NAC的147例TNBC患者。收集临床病理资料，采用单因素和多因素分析复发与非复发患者之间的差异。患者被随机分为训练集和测试集。训练集由104例患者组成（复发：22例，未复发：82例），测试集由43例患者组成（复发：9例，未复发：34例）。为了建立放射组学预测模型，我们使用深度学习分割模型自动分割NAC前后磁共振检查的动态对比增强MRI图像上的肿瘤区域。然后从肿瘤区域提取放射组学特征。在训练集中开发了三种MRI放射组学模型：基于NAC前MRI特征的放射组学模型（模型1），基于NAC后MRI特征的放射组学模型（模型2），以及基于NAC前和NAC后MRI特征的放射组学模型（模型3）。使用独立的临床预测因素在训练集中建立用于预测系统性复发的临床模型。受试者工作特征曲线分析用于评价放射组学和临床模型的性能。结果如下：在预测系统性复发方面，临床模型在训练集中产生的曲线下面积（AUC）为0.747，在测试集中为0.737。在预测系统性复发方面，模型1、2和3在训练集中的AUC分别为0.879、0.91和0.963，在测试集中的AUC分别为0.814、0.802和0.933。在测试集中，所有放射组学模型都比临床模型获得了更高的AUC。采用DeLong检验比较模型之间的AUC，结果表明模型3的预测性能优于临床模型，差异有统计学意义（p < 0.05）。总结：基于NAC前后MRI特征组合建立的放射组学模型在预测TNBC患者是否会在NAC后3年内发生全身复发方面表现出良好的性能。这可以帮助我们非侵入性地识别哪些患者是NAC后复发的高风险患者，以便我们加强对这些患者的随访和治疗。从而改善患者的预后。
Purpose: To establish radiomics prediction models based on automatic segmented magnetic resonance imaging (MRI) for predicting the systemic recurrence of triple-negative breast cancer (TNBC) in patients after neoadjuvant chemotherapy (NAC). Materials and methods: A total of 147 patients with TNBC who underwent NAC between January 2009 and December 2018 were enrolled in this study. Clinicopathologic data were collected, and the differences between the recurrent and nonrecurrent patients were analyzed by univariate and multivariate analyses. Patients were randomly divided into training and testing sets. The training set consisted of 104 patients (recurrence: 22, nonrecurrence: 82), and the testing set consisted of 43 patients (recurrence: 9, nonrecurrence: 34). To establish the radiomics prediction model, we used a deep learning segmentation model to automatically segment tumor areas on dynamic contrast-enhanced-MRI images of preand post-NAC magnetic resonance examinations. Radiomics features were then extracted from the tumor areas. Three MRI radiomics models were developed in the training set: a radiomics model based on pre-NAC MRI features (model 1), a radiomics model based on postNAC MRI features (model 2), and a radiomics model based on both preand post-NAC MRI features (model 3). A clinical model for predicting systemic recurrence was built in the training set using independent clinical prediction factors. Receiver operating characteristic curve analysis was used to evaluate the performance of the radiomics and clinical models. Results: The clinical model yielded areas under the curve (AUCs) of 0.747 in the training set and 0.737 in the testing set in terms of predicting systemic recurrence. Models 1, 2, and 3 yielded AUCs of 0.879, 0.91, and 0.963 in the training set and 0.814, 0.802, and 0.933 in the testing set, respectively, in terms of predicting systemic recurrence. All of the radiomics models had achieved higher AUCs than the clinical model in the testing set. DeLong test was used to compare the AUCs between the models and indicated that the predictive performance of model 3 was better than the clinical model, and the difference was statistically significant (p < 0.05). Conclusion: The radiomics models built based on the combination of preand post-NAC MRI features showed good performance in predicting whether patients with TNBC will have systemic recurrence within 3 years postNAC. This can help us non-invasively identify which patients are at high risk of recurrence post-NAC, so that we can strengthen follow-up and treatment of these patients. Then the prognosis of these patients might be improved.