Developing a Multi-Phase DCE-MRI Radiomics Signature for Breast Cancer Recurrence Risk Prediction: A Multicenter Study

Recurrence is the major life-threatening event of breast cancer (BC) survivors, yet recurrence risk prediction remains challenging, leading to unnecessary adjuvant chemotherapy. This study aimed to develop and validate a non-invasive multi-phase DCE-MRI-based radiomics signature for BC recurrence risk prediction. In this retrospective multi-center cohort study, we included 524 patients with invasive breast cancer who received neoadjuvant chemotherapy, from three cohorts harmonized through the MAMA-MIA consortium: a discovery cohort (DUKE, n = 289) and two validation cohorts (I-SPY1, n = 171; NACT, n = 64). Radiomics, delta radiomics, and kinetics features were extracted from pre-treatment DCE-MRI. Signatures were established using a multi-step feature selection integrating variance-based and correlation-based filtering, univariate Cox regression, and LASSO-penalized multivariate Cox regression. Radio-transcriptomic analysis explored biological basis underlying radiomics signatures. Feature selection identified 25 features for Delta Radiomics signature and 16 features for Single-Phase Radiomics signature. Delta Radiomics and Single-Phase Radiomics signatures significantly stratified recurrence risk across all three cohorts (all P ≤ 0.001) and remained independent prognostic factors after clinical-covariate adjustment. Delta Radiomics achieved the highest predictive performance, with integrated time-dependent area under the receiver operating characteristic curve (iAUC) of 0.804 (95% CI = 0.764–0.861), 0.680 (95% CI = 0.630–0.751), and 0.818 (95% CI = 0.749–0.884) in DUKE, I-SPY1, and NACT, respectively. Radio-transcriptomics analysis showed that Delta Radiomics and Single-Phase Radiomics signatures captured biologically connected but non-redundant information. Combined models integrating both radiomics signatures with clinical and kinetics variables further improved predictive performance and model stability. Specifically, Delta Radiomics + Single-Phase Radiomics + Kinetics + Clinical model displayed iAUC of 0. 0.868 (95% CI = 0.830–0.929), 0.846 (95% CI = 0.791–0.900), and 0.988 (95% CI = 0.970–0.988) in DUKE, I-SPY1, and NACT, correspondingly. DCE-MRI radiomics, particularly delta radiomics that exploit multi-phase enhancement dynamics, provide robust, externally validated, non-invasive prognostic information for BC recurrence and could complement existing prognostic tools. Prospective validation in larger and more diverse cohorts is warranted.