Objective:To investigate the application value of deep learning ultrasound radiomics in predicting axillary lymph node (ALN) metastasis of breast cancer.MethodsThe ultrasound images of breast cancer patients pathologically confirmed were retrospectively analyzed in Gongli Hospital, Shanghai Pudong New Area, from January 2021 to December 2023. Based on whether ALN had metastasized, patients were divided into two groups: those without ALN metastasis and those with ALN metastasis. The research focused on correlating ultrasound characteristics of primary breast cancer lesions with ALN metastasis and evaluating their predictive efficacy. The dataset was randomly split into training and testing sets at an 8∶2 ratio. Nine deep learning models, ResNet50, EfficientNet, MobileNetV3, DenseNet121, DenseNet201, Vision Transforme, VGG16, MobileViT, and Mamba Transformer, were used to predict ALN metastasis. Through five-fold cross-validation, the best-performing model was selected, and decision curve analysis (DCA) was conducted to assess the clinical net benefit of each model. The study also compared the predictive performance of deep learning models against traditional ultrasound features in identifying ALN metastasis in breast cancer patients.ResultsA total of 324 breast cancer patients were included in the study, with a total of 324 breast lesions. Among them, 198 cases had no ALN metastasis, and 126 cases had ALN metastasis. Univariate analysis revealed statistically significant differences (P<0.05) between the non-ALN metastasis and ALN metastasis groups in terms of primary lesion characteristics, including size, shape, orientation, margin, calcification, echogenic halo, spiculation, and lobulation. Multivariate logistic regression identified larger lesion size, non-parallel orientation, presence of an echogenic halo, and spiculation as independent risk factors for ALN metastasis. The combined diagnostic performance of these four features yielded an area under curve (AUC) of 0.805. Among the nine deep learning models evaluated, DenseNet201 demonstrated the highest performance, with AUCs of 0.964 (training set) and 0.861 (testing set). The deep learning models outperformed traditional ultrasound features in predicting ALN metastasis. DCA of DenseNet model indicated a significant net benefits within a risk threshold range of 0.170 6 to 0.605 2.ConclusionDeep learning ultrasound has high clinical value in non-invasive evaluation of axillary lymph node metastasis of breast cancer before surgery, and can provide a basis for the selection of preoperative diagnosis and treatment.