Prediction of the axillary lymph node (ALN) metastasis is an essential factor to guide breast cancer therapy. MRI is the useful imaging tool for prediction of ALN metastasis. The fusion of multiple modalities of MRI images can improve the prediction accuracy of ALN metastasis in breast cancer. However, existing fusion methods often neglect the specific information across different modalities and do not fully exploit the specificity of medical data with respect to patients’ physiological conditions, result in the limitation of performance improvement. To address these issues, we propose a multi-level granularity adversarial contrast learning network (MACMH-net) for ALN metastasis prediction. First, a multi-modal adversarial contrastive learning module is developed with incorporation of adversarial contrast loss and orthogonal contrast loss. It has the ability to learn the shared information and modality-specific information, and combine more relevant information and complementary information in the fusion features. Second, a multi-level granularity contrast learning module is developed with incorporation of the multi-level granularity supervised contrast loss to make full use of specific physiological information of each patient, further improve the discriminative ability of features. Experimental results demonstrate the effectiveness of the proposed method.