【摘要】 目的 基于超声影像组学特征构建机器学习模型，探讨其预测乳腺癌人类表皮生长因子受体2(HER2)低表达的临床价值。方法 选取我院经病理证实的乳腺癌患者197例，根据HER2表达状态分为HER2低表达72例、阳性表达61例和阴性表达64例。本研究设计2个二分类预测任务：任务(1)为鉴别乳腺癌HER2低表达与阳性表达，任务(2)为鉴别乳腺癌HER2低表达与阴性表达。采用随机抽样法按7∶3的比例将纳入患者分为训练集93例、验证集40例（任务(1)）和训练集95例、验证集41例（任务(2)）。模型构建前，基于对应任务的乳腺癌二维超声图像分别提取1316个影像组学特征，采用Wilcoxon秩和检验、Pearson相关系数（PCC）及最小绝对收缩和选择算子（LASSO）回归筛选与HER2低表达相关的特征，并评估其重要性。基于筛选出的特征分别构建逻辑回归、支持向量机、轻度梯度提升机（LightGBM）、自适应增强算法、多层感知机5种机器学习模型，根据受试者工作特征曲线结果，将训练集和验证集中综合表现最好者确定为最优模型，并在模型构建后采用SHAP方法分析最优模型中与乳腺癌HER2低表达相关特征的重要性。结果 针对2个预测任务分别进行影像组学特征筛选：任务(1)中，经Wilcoxon秩和检验、PCC、LASSO回归筛选获得5个与乳腺癌HER2低表达相关的关键特征；任务(2)中，经Wilcoxon秩和检验、PCC、LASSO回归筛选获得9个与乳腺癌HER2低表达相关的关键特征。基于上述筛选出的特征采用5种机器学习算法分别构建任务(1)和任务(2)的预测模型，结果显示LightGBM模型在2个任务中均表现最优：任务(1)中，该模型在训练集和验证集的曲线下面积（AUC）分别为0.79和0.80；任务(2)中，该模型在训练集和验证集的AUC分别为0.81和0.78。任务(1)中采用LASSO回归和SHAP分析显示，LightGBM模型中original＿shape＿Elongation的贡献分别位于第1、2位；任务(2)中采用LASSO回归和SHAP分析显示，LightGBM模型中wavelet-HHH＿glszm＿LowGrayLevelZoneEmphasis的贡献均位于首位，表现出较高的稳定性。结论 基于超声影像组学特征构建的机器学习模型在预测乳腺癌HER2低表达方面有一定的临床价值。更多还原

【Abstract】 Objective To construct a machine learning model based on ultrasound radiomics features,and to investigate its clinical value in predicting human epidermal growth factor receptor 2(HER2)low expression in breast cancer.Methods A total of 197 patients with breast cancer confirmed by pathology in our hospital were selected.According to HER2 expression status,they were divided into HER2-low expression(n=72),HER2-positive expression(n=61),and HER2-negative expression(n=64).Two binary classification prediction tasks were designed in this study:Task(1) was to distinguish HER2-low expression from HER2-positive expression,Task(2) was to distinguish HER2-low expression from HER2-negative expression.The enrolled patients were divided into a training set of 93 cases and a validation set of 40 cases(Task(1)),and a training set of95 cases and a validation set of 41 cases(Task(2))by a random sampling method with a ratio of 7∶3.Before model construction,1316 radiomics features were extracted from two-dimensional ultrasound images of breast cancer for the corresponding tasks.Features associated with HER2-low expression were screened by the Wilcoxon rank-sum test,Pearson correlation coefficient(PCC),and LASSO regression,and their importance was assessed.Five machine learning models,including Logistic regression,support vector machine,light gradient boosting machine(LightGBM),adaptive boosting algorithm,and multilayer perceptron,were constructed based on the selected features.The optimal model was determined in the training and validation sets based on the results of the receiver operating characteristic curve.After the optimal moder constructed,the SHAP method was used to analyze the features’ importance associated with HER2-low expression in the model.Results Radiomic feature selection was conducted separately for two predictive tasks:in Task(1),5 key features associated with HER2-low expression were selected by the Wilcoxon rank-sum test,PCC,and LASSO regression.In Task(2),9 key features associated with HER2-low expression were selected by the Wilcoxon rank-sum test,PCC,and LASSO regression. Based on the features selected from the above procedures,5 machine learning algorithms were respectively employed to construct prediction models for Task(1) and Task(2).The results showed that the LightGBM model performed optimally in both tasks:in Task(1),the model achieved area under the curves(AUCs)of 0.79 and 0.80 in the training and validation sets,respectively.In Task(2),the model achieved AUCs of 0.81 and 0.78 in the training and validation sets,respectively.In Task(1),both the LASSO and SHAP methods revealed that the contribution of original＿shape＿Elongation ranked first and second,respectively.In Task(2),both LASSO and SHAP methods indicated that the contribution of wavelet-HHH＿glszm＿LowGrayLevelZoneEmphasis ranked first in both,demonstrating high stability.Conclusion The machine learning model based on ultrasound radiomics features has certain clinical value for predicting HER2-low expression in breast cancer.更多还原