A multicenter clinicopathologic and MRI-based multimodal fusion model for predicting axillary lymph node metastasis in breast cancer

Background Preoperative prediction of axillary lymph node (ALN) metastasis remains challenging. We developed and externally validated a DCE-MRI-based multimodal fusion model integrating clinical, radiomics, habitat, and deep learning features. Methods A total of 404 patients from two centers were included. Center A ( n = 284) was split into training ( n = 198) and validation ( n = 86) sets; Center B ( n = 120) served as an independent external test cohort. Four sub-models (random forest for clinical variables, ExtraTrees for whole-tumor and habitat radiomics, ResNet50) were combined by weighted late fusion. Progressive fusion, a sequential ablation introducing modalities one at a time, tested whether each stream contributed non-redundant information; all 15 combinations were additionally evaluated. SHAP was applied to the feature-based sub-models and Grad-CAM to the CNN. Results On external testing, the fusion model achieved an AUC of 0.847 (95% CI 0.771–0.921), significantly higher than the clinical model (0.754; DeLong P = 0.017). Differences versus the radiomics (0.797), deep learning (0.784), and habitat (0.817) sub-models were numerical rather than statistically significant ( P = 0.074, 0.066, and 0.248). Progressive fusion showed stepwise gains (0.754, 0.807, 0.812, 0.847), and exhaustive combination analysis indicated that each modality contributed non-redundant information. SHAP identified T stage as the dominant clinical predictor, and the most influential habitat features originated from three different subregions. Grad-CAM showed tumor-focused attention. Conclusion This multimodal fusion model incorporating habitat radiomics achieved good discrimination for preoperative ALN prediction with external validation, and ablation confirmed each feature stream contributed non-redundant information.