目的 探讨基于临床及影像学特征的反向传播神经网络模型在预测前哨淋巴结活检1 ~ 2枚阳性乳腺癌患者腋窝淋巴结负荷水平中的准确性。 方法 回顾性分析2021年1月至2024年12月在3家医疗机构接受腋窝淋巴结清扫的386例女性乳腺癌患者临床及影像资料。根据病理检查结果将纳入患者分为腋窝淋巴结高负荷组（n = 155）和腋窝淋巴结低负荷组（n = 231）。将中心1和中心2（广东医科大学附属医院和广东医科大学附属阳江医院）共295例患者随机分为训练集（n = 207）与验证集（n = 88），将中心3（广东医科大学附属第二医院）的患者（n = 91）作为外部验证集。在训练集上采用单因素、多因素逻辑回归筛选危险因素，并在此基础上应用逻辑回归、支持向量机、随机森林和BPNN四种算法构建风险预测模型，在内部验证集和外部验证集上评估模型的性能。结合Shapley可解释性算法对模型进行特征贡献度分析和可视化。 结果 单因素和多因素逻辑回归分析显示中性粒细胞-淋巴细胞比值（neutrophil-to-lymphocyte ratio，NLR）、瘤周水肿及腋窝淋巴结皮质增厚为淋巴结负荷的独立危险因素。基于BPNN算法构建的预测模型显示良好预测性能，模型的曲线下面积为0.793。Shapley可解释性分析显示瘤周水肿具有最高贡献，其次为淋巴结皮质增厚和中性粒细胞-淋巴细胞比值。 结论 整合临床及影像学特征的可解释BPNN模型能较准确预测腋窝淋巴结负荷水平，为乳腺癌腋窝管理和个体化治疗提供辅助决策。

Objective To evaluate the accuracy of a backpropagation neural network （BPNN） model incorporating clinical and imaging features in predicting axillary lymph node burden among breast cancer patients with one to two positive sentinel lymph nodes. Methods We retrospectively analyzed clinical and imaging data from 386 female breast cancer patients who underwent axillary lymph node dissection at three medical centers between January 2021 and December 2024. Based on pathological findings， patients were categorized into a high axillary lymph node burden group （n = 155） and a low burden group （n = 231）. Data from Center 1 and Center 2 （Affiliated Hospital of Guangdong Medical University and Affiliated Yangjiang Hospital of Guangdong Medical University； n = 295） were randomly divided into a training set （n = 207） and an internal validation set （n = 88）， while data from Center 3 （Second Affiliated Hospital of Guangdong Medical University； n=91） served as the external validation cohort. Univariate and multivariate logistic regression analyses were conducted in the training cohort to identify independent risk factors. Four machine learning algorithms， including logistic regression， support vector machine （SVM）， random forest and BPNN were then used to construct predictive models， which were subsequently evaluated in the internal and external validation cohorts. The Shapley additive explanation （SHAP） method was applied to assess and visualize feature importance. Results Univariate and multivariate logistic regression analyses identified the neutrophil-to-lymphocyte ratio （NLR）， peritumoral edema， and axillary lymph node cortical thickening as independent predictors of nodal burden. The BPNN-based model demonstrated superior predictive performance， achieving an area under the receiver operating characteristic curve （AUC） of 0.793. SHAP analysis revealed that peritumoral edema contributed most significantly to model predictions， followed by lymph node cortical thickening and NLR. Conclusions An interpretable BPNN model integrating clinical and imaging characteristics provides reliable prediction of axillary lymph node burden in breast cancer patients. This approach offers valuable support for axillary management and personalized treatment planning.