摘要目的:探索多期动态增强磁共振成像（dynamic contrast-enhanced magnetic resonance imaging，DCE-MRI）图像中影像学特征在识别乳腺癌分子亚型中的应用价值。方法:回顾性分析2016年1月至2023年12月间于东营市人民医院接受诊察的172例患者的195例乳腺癌病变的多期DCE-MRI影像资料。样本总体包括21例三阴型、18例人类表皮生长因子受体2（human epidermal growth factor receptor 2, HER2）过表达型、53例Luminal A型、76例Luminal B型HER2阴性和27例Luminal B型HER2阳性。从DCE-MRI的每次扫描及不同后处理得到的图像中，通过手动选取的方式划分感兴趣区域并提取影像学特征。将样本按约8∶2的比例划分为主要样本和测试样本，对主要样本进行10次重复的5倍交叉验证以获得训练集和验证集，分别对基于Logistic回归、分类回归树、支持向量分类、随机森林、梯度提升树等方法构建的预测模型进行训练和验证，之后在测试样本上进行受试者工作特征测试。结果:基于Logistic回归的预测模型获得最好的识别性能，其依据各组图像特征进行识别的平均受试者工作特征（receiver operating characteristic curve,ROC）曲线下的面积（area under curve，AUC）为0.781；依据提取自对比后第三次+减影图像的特征识别乳腺癌不同分子亚型的整体AUC为0.809，其中识别Luminal A型、Luminal B型HER2阴性、Luminal B型HER2阳性、HER2过表达型和三阴型的AUC分别为0.784、0.578、0.599、0.812和0.844。使用不同扫描次数和后处理图像的特征，模型的识别效果不同，对比后第三次扫描+减影图像特征和对比后第二次扫描+减影图像特征的识别效果最好，AUC分别为0.762±0.037和0.757±0.046。结论:基于多期DCE-MRI图像影像学特征及机器学习方法优化的模型对乳腺癌分子亚型的预测识别具有潜在应用价值。使用注射对比剂后的中期增强（第二次、第三次扫描）图像特征，以应用图像减影处理，有助于达到更好的识别效果。

abstractsObjective:To explore the applicative potential of radiomic features obtained from multi-stage dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) in distinguishing molecular subtypes of breast cancer.Methods:DCE-MRI archives of 195 breast cancers of 172 patients undergoing radiology examination from Jan. 2016 to Dec. 2023 were retrospectively analyzed, in which the cases of triple-negative subtype, human epidermal growth factor receptor 2 (HER2) overexpression subtype, Luminal A subtype, Luminal B HER2 negative subtype, and Luminal B HER2 positive subtype were 21, 18, 53, 76, 27, respectively. By using the images obtained from each scan of DCE-MRI and with/without processing, regions of interest were manually segmented and marked, based on which the radiomic features were extracted. The cohort was divided into main samples and test samples at a ratio of 8∶2. The models based on Logistic regression, classification and regression trees, support vector classification, random forest, and gradient-boosting trees, were first internally trained and validated by corresponding feature sets obtained using ten times repeated five-fold cross-validation on the primary samples, and then further evaluated by receiver operating characteristic tests using the test samples.Results:The evaluation tests obtained best identification result by adopting the Logistic regression-based model, which achieved an average AUC of 0.781 by using features from various image sequences; AUCs were 0.784, 0.578, 0.599, 0.812 and 0.844 respectively for identifying Luminal A subtype, Luminal B HER2 negative subtype, Luminal B HER2 positive subtype, HER2 overexpression subtype, triple-negative subtype, and for overall identification, by using features extracted from the third post-contrast scan and subtraction-processed images. The models acquired varied identification utilities in tests when using features from images of different scans and with/without processing. The use of radiomic features of subtracted images generated by the third and the second post-contrast scans achieved better identification results with AUCs of 0.762±0.037 and 0.757±0.046, respectively, for all the models.Conclusions:The methodology in this study, aiming at identifying breast cancer molecular subtypes based on radiomic features from multi-stage DCE-MRI and models optimized by machine learning methods, has a certain applicative potential. The use of image features of scans in medium stages (second/third scan) after the administration of contrast agent and the image subtraction processing, facilitates more satisfactory identification utility.