Re-stratification of BI-RADS 3/4A breast lesions based on dynamic contrast-enhanced MRI: integrating cardiometabolic risk and breast microvascular remodeling phenotypes to improve diagnostic accuracy and reduce unnecessary biopsies

Introduction MRI BI-RADS 3 and 4A breast lesions occupy a narrow management boundary between imaging surveillance and biopsy. This study evaluated whether cardiometabolic context and MRI-derived perilesional vascular features improved re-stratification beyond conventional lesion-level MRI information. Methods This retrospective diagnostic prediction-model study included 420 women with one MRI-detected lesion, comprising 160 BI-RADS 3 lesions and 260 BI-RADS 4A lesions. Malignancy was established by histopathology or, for non-biopsied lesions, by at least 24 months of imaging stability. Four prespecified ridge logistic regression models were evaluated: MRI-only, MRI plus cardiometabolic burden, MRI plus an MRI-derived vascular remodeling score, and a fully integrated model. Class weighting was not used. The complete modeling procedure, including preprocessing and penalty selection, was internally validated using 2,000 bootstrap resamples. Results Twenty-seven lesions were malignant, including 3 BI-RADS 3 lesions and 24 BI-RADS 4A lesions. The optimism-corrected AUC was 0.760 for the MRI-only model, 0.775 for the MRI-plus-cardiometabolic model, 0.812 for the MRI-plus-vascular model, and 0.826 for the fully integrated model. The paired optimism-corrected AUC difference between the fully integrated and MRI-only models was 0.066 (95% CI 0.019–0.113; p = 0.006). The integrated model had a bootstrap-corrected calibration intercept of −0.08, calibration slope of 0.94, and Brier score of 0.052. Using aggregated out-of-bootstrap predictions, a locked risk threshold below 0.020 reassigned 91 BI-RADS 4A lesions to 6-month MRI follow-up; 90 were benign and 1 was malignant, representing a 38.1% reduction in immediate biopsy recommendations among benign BI-RADS 4A lesions. A risk threshold of at least 0.050 flagged 6 BI-RADS 3 lesions for intensified assessment, including 2 of the 3 malignant BI-RADS 3 lesions. Sensitivity, specificity, positive predictive value, and negative predictive value were 92.6%, 61.8%, 14.3%, and 99.2%, respectively. Discussion Cardiometabolic context and MRI-derived vascular information improved internally validated risk estimation. However, the malignant downgrade and the limited number of malignant outcomes preclude clinical implementation before prospective external validation.