DCE-MRI影像组学特征在预测乳腺癌腋窝淋巴结转移中的价值

Value of DCE-MRI based radiomics features for prediction of axillary lymph node metastasis in breast carcinoma

目的探讨动态对比增强MRI(dynamic contrast enhanced MRI,DCE-MRI)影像组学特征在预测乳腺癌腋窝淋巴结(axillary lymph node,ALN)转移中的价值。材料与方法回顾性分析2017年1月至2020年12月间经河南省人民医院手术病理证实为乳腺癌患者的首次术前MRI图像及临床病理资料(包括患者年龄、病灶的位置和大小、SBR分级,ER、PR、HER-2及Ki-67的表达情况,ALN是否转移及脉管癌栓的有无),共入组356例患者,年龄26~82(49.17±10.75)岁,并按8∶2的比例将其随机分为训练集(n=284)和测试集(n=72)。提取DCE-T1WI序列第3期图像的影像组学特征,并通过Mann-Whitney U检验、Z分数归一化、方差阈值、K最佳及最小绝对值收缩和选择算子(least absolute shrinkage and selection operator,LASSO)算法筛选与ALN转移强相关的定量影像组学特征;应用多种分类器算法以排列组合方式分别构建影像组学标签,并利用受试者工作特征(receiver operating characteristic,ROC)曲线得到的曲线下面积(area under the curve,AUC)、敏感度、特异度及准确度评价模型的预测性能,根据模型效能从中确定最佳影像组学预测模型。结果356例乳腺癌患者中ALN转移组117例(32.9%,117/356),无ALN转移组239例(67.1%,239/356);HER-2阳性表达在ALN转移组和无转移组之间的差异有统计学意义(χ^(2)=5.433,P=0.020),其余临床病理指标在两组间的差异均无统计学意义(P>0.05);且临床病理指标在训练集与测试集患者中的差异均无统计学意义(P>0.05)。从初始653个影像组学特征中共筛选得到18个与ALN转移强相关的影像组学特征,包括形态特征、一阶特征及纹理特征各6个。基于绝对值最大归一化和Bagging决策树算法构建的影像组学标签是预测ALN转移的最佳模型,该模型在训练集和测试集的AUC、敏感度、特异度和准确度分别为0.929[95%置信区间(confidence interval,CI):0.897~0.960]、69.9%、96.9%、88.0%和0.803(95%CI:0.701~0.905)、75.0%、75.0%、75.0%。结论基于DCE-MRI影像组学特征构建的预测模型有助于乳腺癌术前ALN评估。

Objective:To explore the value of radiomics features extracted from dynamic contrast enhanced MRI(DCE-MRI)for preoperative prediction of axillary lymph node(ALN)metastasis in breast cancer.Materials and Methods:The first preoperative MRI images and clinicopathological data(including patient age,location and size of lesion,SBR grade,expression of ER,PR,HER-2 and Ki-67,whether ALN metastases and vascular cancer thrombus were present)of patients with breast cancer confirmed by surgical pathology in Henan Provincial People's Hospital from January 2017 to December 2020 were retrospectively analyzed.A total of 356 patients aged 26 to 82(49.17±10.75)years were enrolled,which were randomly divided into the training set(n=284)and test set(n=72)according to the ratio of 8∶2.The radiomics features of phase 3 images in the DCE-T1WI sequence were extracted,and the quantitative radiomics features having strong correlation with ALN metastasis were selected using Mann-Whitney U test,Z-score normalization,variance threshold,K-best and least absolute shrinkage and selection operator(LASSO)regression methods.A variety of classifier algorithms were used to construct radiomics labels in a permutation-combination way.The area under the curve(AUC),sensitivity,specificity and accuracy of receiver operating characteristic curve(ROC)were used to evaluate the efficiency of the model,then the optimal prediction model was selected according to the efficiency.Results:Among 356 patients with breast cancer,117 patients(32.9%,117/356)had ALN metastasis and 239 patients(67.1%,239/356)had no ALN metastasis.There was a statistically significant difference in HER-2 positive expression between the ALN metastasis group and non-metastasis group(χ^(2)=5.433,P=0.020),and there was no statistically significant differences in the other clinicopathological indicators between the two groups(P>0.05).There was no statistically significant differences in clinicopathological indicators between the training set and the test set(P>0.05).A total of 18 radiomics features having strong correlation with ALN metastasis were selected finally from the initial 643 radiomics features,including each 6 morphological features,first order features and texture features respectively.The optimal ALN prediction model was selected through radiomics signatures based on maximum absolute value normalization and Bagging decision tree algorithm,and the AUC value,sensitivity,specificity and accuracy of the model in the training set and test set were 0.929[95%confidence interval(CI):0.897-0.960],69.9%,96.9%,88.0%and 0.803(95%CI:0.701-0.905),75.0%,75.0%and 75.0%respectively.Conclusions:The prediction model based on DCE-MRI radiomics features could be helpful for preoperative predicting ALN metastasis in breast cancer.