探讨基于增强MRI影像组学构建的预测模型对乳腺癌腋窝淋巴结转移的诊断效能。

根据纳入及排除标准，回顾性分析2016年1月至2020年6月在吉林省肿瘤医院接受手术治疗的376例女性乳腺癌患者的临床、病理及影像资料。患者术前接受乳腺增强MRI检查。用随机数字表法从所有患者中选择20例，由2名放射科医师对其影像独立分割病灶，勾画感兴趣区域(ROI)。采用组内相关系数(ICC)检验对2名医师勾画的ROI进行一致性分析。采用A.K.(Version 3.3.0)软件提取病灶区域的三维纹理特征参数，通过IPMs(Version 2.0.2)软件及最小绝对收缩和选择算子(LASSO)筛选出最佳影像组学特征，用于构建乳腺癌淋巴结转移预测模型。按7∶3的比例将患者随机分为训练组(n＝263)和验证组(n＝113)。采用6种机器学习模型，包括Logistic回归、随机森林(RF)、贝叶斯算法(NB)、决策树(DT)、邻近算法(KNN)和支持向量机(SVM)模型，对2组数据进行处理，采用受试者操作特征(ROC)曲线分析各机器学习模型对乳腺癌腋窝淋巴结转移的诊断效能，根据训练组ROC曲线下面积(AUC)大小，选择最佳模型。用决策曲线分析(DCA)评价最佳模型的临床获益。

共有腋窝淋巴结阳性患者114例，腋窝淋巴结阴性患者262例。操作者一致性检验结果显示：2名医师具有很强的操作者一致性(ICC＝0.915，95%CI: 0.894~0.932, P<0.001)。每个病灶提取得到396个三维纹理特征参数，筛选保留7个淋巴结阳性组与淋巴结阴性组间差异明显的影像组学特征，并将其与2个临床指标(年龄和绝经状态)和5个影像学指标(病灶位置、有无钙化、病灶数量、病灶长径和造影剂时间-信号曲线类型)一起，共14个参数，构建预测模型。6种机器学习模型中，Logistic回归模型在训练组中诊断乳腺癌腋窝淋巴结转移的AUC最高(0.798)，准确率为73.0%(192/263)、特异度为75.4%(138/183)、敏感度为67.5%(54/80)，在验证组的AUC为0.767、准确率为73.5%(83/113)、特异度为77.2%(61/79)、敏感度为64.7%(22/34)，为最佳机器学习模型。基于Logistic回归模型的决策曲线分析结果显示：训练组中阈值为0.15~1.00，验证组中阈值为0.10~0.60，有临床获益。

基于增强MRI影像组学结合机器学习模型建立的预测模型能够鉴别乳腺癌淋巴结转移，其中以Logistic回归模型的诊断效能更优。

To investigate diagnosis efficiency of the predictive model based on enhanced MRI radiomics for axillary lymph node metastasis of breast cancer.

According to the inclusion and exclusion criteria, the clinical, pathological and imaging data of 376 female breast cancer patients who received surgical treatment in the Jilin Cancer Hospital from January 2016 to June 2020 were retrospectively analyzed. The patient underwent enhanced MRI of the breast before surgery. Twenty patients were selected from all patients by random number table method, and 2 radiologists independently segmented the lesions on their images and delineated the region of interest (ROI). The intraclass correlation coeficient (ICC) was used to analyze the consistency of the ROI delineated by the two radiologists. The A. K. (Version 3.3.0) software was used to extract the three-dimensional imaging parameters of the lesion area, and the optimal radiomic parameters were screened out using the IPMs (Version 2.0.2) software and the least absolute shrinkage and selection operator (LASSO) to construct a predictive model for lymph node metastasis in breast cancer. Patients were randomly divided into the training group (n=263) and validation group (n=113) at a ratio of 7∶3. Six machine learning algorithms including the Logistic regression, random forest (RF), Bayesian algorithm (NB), decision tree (DT), K-neighborhood algorithm (KNN) and support vector machine (SVM) were used to process the data of the training group and validation group. The receiver operating characteristic (ROC) curve was used to investigate the diagnosis efficiency of the above-mentioned 6 models for axillary lymph node metastasis of breast cancer, and the optimal model was selected based on AUC of the training group. The clinical benefit of the optimal model was evaluated by the decision curve analysis (DCA).

There were 114 patients with positive axillary lymph nodes and 262 patients with negative axillary lymph nodes. The ICC was 0.915, indicating high consistency between the two radiologists (95%CI: 0.894-0.932, P<0.001). Totally 396 three-dimensional imaging parameters were extracted from each lesion. Among them, 7 radiomic parameters with significant differences between the lymph node-positive group and the lymph node-negative group were screened out. The 7 radiomic parameters, 2 clinical parameters (age and menopausal status) and 5 imaging parameters (lesion location, calcification, number of lesions, long diameter of lesions, and time-signal curve type of contrast agent) were employed to construct a predictive model. Among six machine learning algorithms, the Logistic regression model had the highest AUC (0.798) in the diagnosis of axillary lymph node metastasis in breast cancer patients of training group, with the accuracy of 73.0%(192/263), specificity 75.4%(138/183) and sensitivity 67.5%(54/80). In the validation group, the Logistic regression model had the AUC of 0.767, accuracy 73.5%(83/113), specificity 77.2%(61/79) and sensitivity 64.7%(22/34), indicating that the Logistic regression model was the optimal. The decision curve analysis of the Logistic regression model showed that the threshold was 0.15-1.00 in the training group, and 0.10-0.60 in the validation group, indicating obvious clinical benefit.

The predictive models based on enhanced MRI radiomics combined with machine learning algorithm can diagnose lymph node metastasis of breast cancer, and among them the Logistic regression model shows the optimal diagnostic efficiency.