基于X线及超声乳腺影像报告和数据系统构建机器学习模型预测乳腺癌分子分型

目的 观察基于X线及超声的乳腺影像报告和数据系统（BI-RADS）选取影像学特征构建的机器学习模型预测乳腺癌分子分型的可行性。方法 回顾性分析200例经病理的浸润性乳腺癌，根据免疫组织化学结果分为Luminal组（n=109）与非Luminal组（n=91），组内按7 ：3比例随机分为训练亚组及测试亚组。采集11个临床信息，并提取24个影像学特征，建立4种机器学习模型，通过受试者工作特征（ROC）曲线评价各模型预测不同分子分型乳腺癌的效能，比较各模型曲线下面积（AUC）的差异。结果 测试组随机森林（RF）、极端梯度提升（XGBoost）、逻辑回归（LR）及支持向量机（SVC）模型判断不同分子分型乳腺癌的敏感度分别为74.10%、74.10%、77.80%和70.40%，特异度分别为63.60%、51.50%、57.60%和60.60%，准确率分别为68.30%、61.70%、66.70%和65.00%；其中RF模型判断Luminal型与非Luminal型乳腺癌的AUC值最大（AUC=0.70，P<0.05），但与其他模型间差异均无统计学意义（P均>0.05）。结论 RF模型预测不同分子分型乳腺癌的效能较好。

Machine learning models for predicting molecular types of breast cancer based on X-ray and ultrasound breast imaging reporting and data system

Objective To observe the feasibility of predicting molecular type of breast cancer based on machine learning models established using X-ray and ultrasound breast imaging reporting and data system (BI-RADS). Methods A total of 200 patients of invasive breast cancer confirmed by pathology were retrospectively analyzed. According to immunohistochemical results, the patients were divided into Luminal group(n=109)and non-Luminal group(n=91), and were randomly further divided into training subgroups and test subgroups according to the ratio of 7: 3. Totally 11 clinical information were collected, and 24 imaging features were extracted, and then 4 machine learning models were established, respectively. The receiver operating characteristic (ROC) curve was used to evaluate the effectiveness of each model for predicting different molecular types of breast cancer, and the area under the curves (AUC) were compared among models. Results The sensitivity of random forest (RF), extreme gradient boosting (XGBoost), logistics regression (LR) and support vector machine (SVC) models for predicting Luminal and non-Luminal breast cancer was 74.10%, 74.10%, 77.80% and 70.40%, specificity was 63.60%, 51.50%, 57.60% and 60.60%, accuracy was 68.30%, 61.70%, 66.70% and 65.00%, respectively. AUC of RF model in predicting Luminal and non-Luminal breast cancer was the highest (AUC=0.70, P<0.05), but there was no statistical significant difference among 4 models (all P>0.05). Conclusion RF had good overall performance in predicting molecular type of breast cancer.