【摘要】 目的 基于MRI图像的深度学习技术预测乳腺癌脉管浸润状态。方法 回顾性收集2010年1月至2021年12月在广州市红十字会医院病理证实457例乳腺癌患者的MMRI图像（FS-T₂WI）。根据病理结果将脉管浸润分为阳性和阴性,将患者按9:1随机分为训练集（n=411）及验证集（n=46）。训练集采用深度学习中的2D VanillaCNN和ResNet10、3DResNet10算法训练模型,并在验证集中评价其效能。结果457例患者中,脉管浸润阴性304例,阳性153例。基于FS-T₂WI序列的2D VanillaCNN和ResNet10、3D ResNet10算法模型预测脉管浸润的曲线下面积（AUC）分别为0.70、0.77、0.69;基于DCE序列的2D VanillaCN N和ResN et 10、3D ResNet1 0算法模型预测脉管浸润的AUC值分别为0.76、0.72、0.80。结论 基于深度学习技术能有效预测乳腺癌脉管浸润状态,且DCE-T1C序列的预测效能优于FS-T₂WI序列。更多还原

【Abstract】 Objective To predict the status of lymphovascular invasion（LVI） in breast cancer with MRI based on deep learning.Methods The MR images（FS-T₂WI） of 457 breast cancer confirmed by pathology in Guangzhou Red Cross Hospital from January 2010 to December 2021 were retrospectively enrolled.According to the pathological results,LVI divided into positive and negative.The patients were randomly divided into a training set（n=411） and a validation set（n=46） at a ratio of 9:1.The training set uses 2D Vanilla CNN and ResNet 10,3D ResNet10 algorithms from deep learning to train the model,and evaluate the prediction performances in the validation cohort.Results Among the 457 patients,a total of 304 were LVI negative and 153 were positive. The area under the curve（AUC） of 2D-Vanilla CNN、ResNet10,3DResNet10 algorithm models based on FS-T₂WI sequence for predicting are 0.70,0.77,and 0.69,respectively;The AUC of 2D Vanilla CNN and ResNet10,3D ResNet10 algorithm models based on DCE sequence for predicting are 0.76,0.72,and0.80,respectively.Conclusion The deep learning technology can effe ctively predict LVI status of breast cancer,and the prediction performance of DCE sequence is better than that of FS-T₂WI sequence.更多还原