为了开发基于 MRI 的深度学习特征，用于预测乳腺癌 (BC) 患者新辅助化疗 (NAC) 后的腋窝反应。我们招募了 327 名患有腋窝淋巴结 (ALN) 转移的 BC 患者，在 NAC 后接受腋窝手术。深度学习特征由 ResNet34 提取，该特征由来自 ImageNet 的大型且注释良好的数据集进行预训练。然后，我们确定了动态对比增强磁共振成像 (DCE-MRI) 上的深度学习放射组学用于预测 BC 患者 NAC 后腋窝反应。128 个深度学习放射组学 (DLR) 特征的提取依赖于每位患者的 DCE-MRI。经过最小绝对收缩和选择算子回归分析后，治疗前、治疗后和组合 DCE-MRI 分别保留了 13、8 和 21 个特征。基于 DCE-MRI 组合建立的 DLR 特征在 NAC 后的 ALN 响应中取得了良好的能力。支持向量机在训练和测试集中实现了最佳性能，曲线下面积 (AUC) 为 0.99（95% 置信区间 (CI)，0.98-1.00）和 0.83（95% CI，0.73-0.92），分别。根据临床参数建立的 LR 模型表现最佳，测试集中的 AUC 为 0.73（95% CI，0.62-0.84），敏感性为 0.73，特异性为 0.73，PPV 为 0.63，NPV 为 0.81。最后，放射组学特征和临床特征的整合导致建立了预测放射组学列线图，AUC为0.99（95％CI，0.99-1.00）。总之，我们当前的研究通过深度学习方法构建了预测列线图，证明了在训练和测试队列中表现良好。目前的预后模型为指导接受 NAC 的 BC 患者进行 ALN 管理的手术策略提供了精确和客观的基础。版权所有 © 2023 大学放射科医生协会。由爱思唯尔公司出版。保留所有权利。

To develop a MRI-based deep learning signature for predicting axillary response after neoadjuvant chemotherapy (NAC) in breast cancer (BC) patients.We enrolled 327 BC patients with axillary lymph node (ALN) metastases receiving axillary operations after NAC. The deep learning features were extracted by ResNet34, which was pretrained by a large, well-annotated dataset from ImageNet. Then we identified deep learning radiomics on magnetic resonance imaging with dynamic contrast enhancement (DCE-MRI) in predicting axillary response after NAC in BC patients.The extraction of 128 deep learning radiomics (DLR) features relied on the DCE-MRI for each patient. After the least absolute shrinkage and selection operator regression analysis, 13, 8, and 21 features remained from the pre-treatment, post-treatment, and combined DCE-MRI, respectively. The DLR signature established based on the combined DCE-MRI achieved good capacity in ALN response after NAC. The support vector machine achieved the best performance with an 0.99 area under the curve (AUC) of (95% confidence interval (CI), 0.98-1.00) and 0.83 (95% CI, 0.73-0.92) in the training and test sets, respectively. The LR model established with clinical parameters represented the best performance with 0.73 AUC (95% CI, 0.62-0.84), 0.73 sensitivity, 0.73 specificity, 0.63 PPV, and 0.81 NPV in the test set, respectively. Finally, the integration of radiomic signature and clinical signature resulted in establishing a predictive radiomic nomogram, with an AUC of 0.99 (95%CI, 0.99-1.00).In conclusion, our current study constructed a predictive nomogram through the deep learning method, demonstrating favorable performance in the training and test cohort. The present prognostic model furnishes a precise and objective foundation for directing the surgical strategy toward ALN management in BC patients receiving NAC.Copyright © 2023 The Association of University Radiologists. Published by Elsevier Inc. All rights reserved.