目的

探讨基于超声的深度学习列线图预测乳腺癌新辅助化疗（NAC）后腋窝淋巴结（ALN）状态的价值。

方法

回顾性选取2020 年3 月至2023 年6 月在南京医科大学第一附属医院（训练集，n=257）和解放军东部战区总医院（外部测试集，n=157）接受NAC 的414 例ALN 转移的乳腺癌患者，并根据NAC 后ALN 手术病理结果分为病理完全缓解（pCR）组和非病理完全缓解（npCR）组。使用NAC 前的乳腺肿瘤二维超声图像训练并构建基于ResNet50 架构的深度学习模型；通过单因素和多因素Logistic 回归分析临床病理特征，筛选出与NAC 后ALN 病理状态有关的独立危险因素，构建临床模型；联合独立危险因素与深度学习预测概率构建深度学习列线图。利用受试者工作特征（ROC）曲线、校准曲线、决策曲线分析、临床影响曲线评估模型性能。由2 位不同年资超声医师根据超声图像及NAC 前免疫组化结果对外部测试集进行独立预测，并在深度学习列线图的辅助下进行第二次预测，比较分析两次预测结果。

结果

临床病理资料中，雌激素受体（ER）和人表皮生长因子受体2（Her-2）是预测NAC 后ALN 病理状态的独立危险因素。临床模型、深度学习模型和深度学习列线图的曲线下面积（AUC）在训练集中分别为0.724、0.872、0.878，在外部测试集中分别为0.698、0.831、0.859。深度学习列线图的预测效能优于临床模型（训练集、外部测试集中P 值均＜0.001），且在外部测试集中其优于深度学习模型（P=0.024）。医师1（低年资）和医师2（高年资）独立判断的AUC 值分别为0.570、0.606，均低于深度学习模型和深度学习列线图（P 均＜0.001）。在深度学习列线图的帮助下，医师1 和医师2 的诊断能力AUC 分别提升至0.796 和0.807，与独立判断比较，差异均有统计学意义（P 均＜0.001）。

结论

基于NAC 前超声图像的深度学习列线图可以在治疗前有效预测乳腺癌NAC 后ALN 病理状态，为个性化治疗方案的制定提供更多依据。

Objective

To assess the value of a deep learning-based nomogram in predicting axillary lymph node （ALN） status following neoadjuvant chemotherapy （NAC） in breast cancer patients.

Methods

Four hundred and fourteen ALN-positive breast cancer patients who received NAC between March 2020 and June 2023 were enrolled in this retrospective study and divided into a training set and an external test set.The training set consisted of 257 patients from the First Affiliated Hospital of Nanjing Medical University， while the external test set included 157 patients from the General Hospital of Eastern Theater Command.All patients were divided into pathologically complete response （pCR） and non-pCR（npCR） groups based on the pathology results of ALN surgery post-NAC.A deep learning model based on the ResNet50 architecture was trained and established using pre-NAC ultrasound images of breast tumors.Univariate and multivariate logistic regression analyses were performed on the training set to identify independent risk factors for post-NAC ALN status.These independent risk factors were then used to construct a clinical model.A deep learning-based nomogram was constructed by combining independent risk factors and deep learning predictive probabilities.The performance of the models was evaluated using receiver operating characteristic （ROC） curve， calibration curve， decision curve analysis， and clinical impact curve.Two radiologists with different experience levels independently predicted ALN status in the external test set based on ultrasound images and pre-NAC immunohistochemical results， and performed a second prediction with the assistance of the deep learning-based nomogram.The two prediction results were compared.

Results

Estrogen receptor （ER） and human epidermal growth factor receptor 2 （Her-2） were identifled as independent risk factors for predicting post-NAC ALN status.The area under the curve （AUC） values of the clinical model， deep learning model， and deep learning-based nomogram were 0.724， 0.872， and 0.878 in the training set， and 0.698， 0.831， and 0.859 in the external test set， respectively.The deep learning-based nomogram outperformed the clinical model （both P＜0.001 in training and external test sets） and showed superior performance to the deep learning model in the external test set （P=0.024）.The AUC values of radiologist 1 （low-experience） and radiologist 2 （high-experience） for independent judgment were 0.570 and 0.606， respectively， both signiflcantly lower than those of the deep learning model and the deep learningbased nomogram （all P＜0.001）.With the assistance of the deep learning-based nomogram， the AUC values of radiologist 1 and radiologist 2 improved to 0.796 and 0.807， respectively， showing statistically signiflcant differences compared to independent judgment （both P＜0.001）.

Conclusion

The deep learning-based nomogram based on pre-NAC ultrasound images can effectively predict the pathological status of ALN in breast cancer patients after NAC treatment， providing more evidence for the development of personalized treatment plans.