基于MRI时空异质性模型预测三阴性乳腺癌病理完全缓解

Value of radiomics model based on spatiotemporal heterogeneity of MRI to predict pathological complete response in triple-negative breast cancer

目的 基于MRI时空异质性构建影像组学模型以早期预测三阴性乳腺癌(triple-negative breast cancer, TNBC)的病理完全缓解(pathological complete response, pCR)。材料与方法 回顾性分析我院2017年9月至2022年3月接受新辅助化疗(neoadjuvant chemotherapy, NAC)的173名TNBC患者资料,收集每位患者NAC前(Pre-)和NAC两疗程后(During-)的MRI图像。55名DUKE大学的患者构成外部验证队列。从瘤内亚区域和瘤周区域提取影像组学特征来表征空间异质性,计算NAC前后特征值的变化(Delta-)来表征时间异质性。分别使用Pre-、During-和Delta-的特征,应用最小绝对收缩选择算子(least absolute shrinkage and selection operator, LASSO)回归构建影像组学模型。采用多因素逻辑回归对单模态模型进行集成,构建纵向融合(Stacking)模型。通过受试者工作特征(receiver operating characteristic, ROC)曲线、校准曲线和决策曲线分析(decision curve analysis, DCA)评价模型的诊断效能和临床应用价值。结果 分别从Pre-、During-和Delta-特征集中选取8个、4个和10个特征构建模型。基于空间异质性的Pre-模型具有预测pCR的性能,在训练集、内部验证集和外部验证集中的ROC曲线下面积(area under the curve, AUC)分别为0.74、0.71和0.71。在训练集和验证集中,纵向融合模型预测pCR的性能最佳,AUC均为0.86。DCA结果显示纵向融合模型在临床应用中价值是最高的。结论 基于MRI空间异质性特征可以有效预测TNBC的pCR,整合时空异质性构建的纵向融合模型可以进一步提高预测性能。

Objective:To develop a spatiotemporal heterogeneity based radiomics model for the early prediction of pathological complete response(pCR)in triple-negative breast cancer(TNBC).Materials and Methods:The data of 173 TNBC patients who received neoadjuvant chemotherapy(NAC)in our hospital from September 2017 to March 2022 were retrospectively analyzed.MRI images of each patient were collected at pretreatment(Pre-)and after two cycle of NAC(During-).The 55 patients from the DUKE university constituted the external validation cohort.Radiomics features were extracted from the intratumoral subregions and peritumoral region to characterize spatial heterogeneity,and the changes of features before and during NAC(Delta-)were calculated to characterize temporal heterogeneity.The radiomics models were developed by least absolute shrinkage and selection operator(LASSO)regression using the Pre-,During-,and Delta-features.Multi-factor logistic regression was used to integrate single-mode models to develop the longitudinal fusion(Stacking)model.The diagnostic performance and clinical application value of models were evaluated by the receiver operating characteristic(ROC)curves,calibration curves,and decision curve analysis(DCA).Results:Finally,8,4 and 10 features were respectively selected from the Pre-,During-and Delta-feature sets to construct the models.The Pre-model based on spatial heterogeneity could predict pCR,with area under the curve(AUC)of 0.74,0.71 and 0.71 in the training set,validation set and external validation set,respectively.In the training and validation sets,the Stacking model achieved the best performance to predict pCR,and the AUC was 0.86 in both sets.DCA indicated that the value of Stacking model was highest in clinical practice.Conclusions:Features based on MRI spatial heterogeneity can effectively predict the pCR of TNBC.The longitudinal fusion model integrated spatiotemporal heterogeneity has the potential to further improve the prediction performance.