基于MRI影像组学预测乳腺癌新辅助化疗后肿瘤退缩模式的研究

MRI-based radiomics for prediction of tumor regression pattern to neoadjuvant chemotherapy in breast cancer

目的基于治疗前乳腺MRI影像组学定量特征,并融合MRI定性影像特征及临床病理信息建立联合模型用于早期预测乳腺癌新辅助化疗后肿瘤退缩模式。材料与方法回顾性分析广东省人民医院2012年2月至2020年8月接受新辅助化疗并进行手术的420例乳腺癌患者临床资料。以手术标本的病理结果为金标准,将肿瘤退缩模式分为向心性和非向心性退缩。根据MRI检查时间顺序以7∶3的比例分为训练组(n=294)、验证组(n=126)。在动态增强MRI的第2期增强图像中对原发灶进行感兴趣区勾画,并提取影像组学特征。采用两独立样本t检验或Mann-Whitney U检验、相关性分析及最小绝对收缩和选择算子-logistic回归分析对影像组学特征进行降维筛选,然后基于人工神经网络建立影像组学标签。通过单因素、多因素logistic筛选显著相关的临床病理特征建立临床预测模型,并联合定性影像学特征和影像组学标签构建联合预测模型。使用受试者工作特征(receiver operating characteristic,ROC)曲线和校准曲线评估模型性能,并使用决策曲线分析(decision curve analysis,DCA)评价预测模型的临床实用性。结果本研究共筛选出8个与肿瘤退缩模式显著相关的影像组学特征。在训练组和验证组中,影像组学标签的曲线下面积(area under the curve,AUC)分别为0.738(95%CI:0.705~0.754)和0.696(95%CI:0.585~0.712);临床预测模型AUC值分别为0.676(95%CI:0.636~0.741)和0.619(95%CI:0.601~0.716);联合预测模型的AUC值分别为0.802(95%CI:0.753~0.824)和0.764(95%CI:0.685~0.820)。DCA显示联合模型具有临床应用价值。结论融合乳腺癌新辅助治疗前MRI的影像组学定量特征和定性影像学特征及临床病理信息所构建的联合模型有助于预测肿瘤退缩模式,有望协助临床早期识别可降期保乳的患者,以优化个体化诊疗方案,改善患者预后。

Objectives:To develop a model by combining pretreatment MRI-based quantitative radiomics and qualitative image features and clinicopathologic information for early prediction of tumor regression pattern to neoadjuvant chemotherapy(NAC)in breast cancer.Materials and Methods:Clinical data of 420 patients with breast cancer who received neoadjuvant chemotherapy and surgery from Guangdong Provincial People's Hospital from February 2012 to August 2020 were retrospectively analyzed.Pathologic findings of surgical specimens were used as the gold standard to classify the tumor regression patterns into concentric and non-concentric shrinkage.The training cohort(n=294)and the validation cohort(n=126)were divided into 7∶3 according to the chronological order of MRI examinations.In the 2nd phase images of dynamic contrast-enhanced MRI,the regions of interest(ROI)were delineated and the radiomics features of the ROI were extracted.Two independent-samples t test or Mann-Whitney U test,correlation analysis,least absolute shrinkage and selection operator(LASSO)-logistic regression were used for dimension reduction of radiomics features and artificial neural networks were used to establish a radiomics signature.Clinical prediction models were constructed by screening the significant clinicopathological features by univariate and multifactorial logistic regression.In addition,a predictive model combining qualitative image features,clinicopathologic features and radiomics signatures was constructed.The performance of the model was assessed using the receiver operating characteristic(ROC)curves and calibration curves.The decision curve analysis(DCA)was conducted to assess the clinical use of these predictive models.Results:Eight radiomics signatures significantly correlated with tumor regression patterns were selected.In the training cohort and validation cohort,the radiomics signature yielded an area under curve(AUC)value of 0.738(95%CI:0.705-0.754)and 0.696(95%CI:0.585-0.712),respectively;the clinical predictive model yielded an AUC value of 0.676(95%CI:0.636-0.741)and 0.619(95%CI:0.601-0.716),respectively;the combined predictive model yielded an AUC value of 0.802(95%CI:0.753-0.824)and 0.764(95%CI:0.685-0.820),respectively.DCA showed the clinical use of the combined predictive models.Conclusions:Prediction models combining pretreatment MRI-based quantitative radiomics and qualitative MRI image features and clinicopathologic information are useful for predicting tumor regression pattern in breast cancer,which can assist in selecting patients who can benefit from NAC for de-escalation of breast surgery,in order to optimize the individualized diagnosis as well as treatment plan,and improve the prognosis of patients.