Cross-Modal Artificial Intelligence Integrating Dynamic MRI and Clinical Data for Breast Cancer Diagnosis and Nodal Metastasis Prediction

Abstract

Accurate preoperative assessment of breast lesion malignancy and axillary lymph node (ALN) status is important for individualized breast cancer management, yet current approaches are limited by diagnostic variability and invasive nodal staging. We aimed to develop and validate a multicenter artificial intelligence framework for breast lesion classification and ALN metastasis prediction using dynamic contrast-enhanced magnetic resonance imaging (MRI) and preoperative clinical information. In this multicenter study, 3320 patients from four hospitals were assigned to a training cohort, an internal testing cohort, and two independent external validation cohorts. Lesion classification was based on MRI alone, whereas ALN prediction additionally incorporated clinical characteristics. MRI sequences were analyzed using a Video Swin Transformer Tiny backbone with low-rank adaptation and attention-based multiple-instance learning, with gated multimodal fusion used for ALN prediction. The model achieved AUCs of 1.000, 1.000, 1.000, and 0.994 for lesion classification and 0.991, 0.974, 0.987, and 0.960 for ALN prediction across the four cohorts. Calibration and decision curve analyses further characterized model performance, while exploratory survival analysis showed that AI-predicted ALN status was associated with disease-free survival. These findings support the potential of AI-assisted breast MRI to provide quantitative information for breast lesion characterization and preoperative assessment of ALN involvement.