Assessing the risk of breast cancer recurrence with pre-treatment MRI: A transfer learning study on multicenter data

Background: MRI features have demonstrated prognostic value in predicting future breast cancer recurrence. However, deep learning studies remain limited particularly those evaluating performance across tumor subtypes and different time horizons in large multicenter datasets. Materials and Methods: We used pretreatment DCE-MRI exams from the multicenter MAMA-MIA dataset (433 breast cancer patients who underwent NAT; 115 with recurrence events, 318 recurrence-free) to evaluate a transfer-learning framework based on DenseNet121 pretrained on ImageNet. Middle tumor-containing slices were bias-field corrected, resampled, cropped to the tumor regions, and resized to 224×224 pixels. The top 10% of DenseNet121 layers were unfrozen for model fine-tuning for recurrence prediction. Five-fold stratified cross-validation preserved site distribution across data splits. Model performance was evaluated using Harrell's C-index and time-dependent AUCs at 3 and 5 year horizons. Correlated C-indices were compared using the two-sided Kang et al. test. We also assessed the added value by our deep-learning risk score to a baseline prognostic model based on the established clinical factors HR and HER2. Model performance was also assessed across tumor subtypes. Results: Our deep learning model achieved a C-index of 0.67±0.03 with 3 and 5 year AUCs of 0.69±0.04 and 0.67±0.10, respectively. Adding our deep-learning risk score to the baseline model significantly improved performance from 0.64 to 0.71 (p=0.007). Subtype-specific evaluations (Table 1) showed variable performances with the highest performance in HER2-pure and TNBC patients. Conclusion: Our findings highlight the potential of a transfer-learning-based DenseNet121 MRI model to predict 3 and 5 year recurrences in breast cancer patients, providing added value beyond standard clinical factors. Future optimizations will aim at improving subtype-specific performance in large multicenter datasets. Table 1. Recurrence-free survival analysis results in multi-center MAMA-MIA dataset (N = 433). Added prognostic value by our deep learning model Model C-index p-value* DL model 0.67 ± 0.03 0.01 Baseline model (HR, HER2) 0.64 ± 0.05 N/A Baseline + DL 0.71 ± 0.05 0.007 * p-value for C-index differences with respect to the baseline model. Performance of our deep learning model by time horizon 3-year AUC 0.69 ± 0.04 5-year AUC 0.67 ± 0.10 Performance of our deep learning model by breast cancer subtype C-index 3-year AUC 5-year AUC Luminal A (N = 121) 0.61±0.02 0.61±0.11 0.64±0.08 Luminal B (N = 50) 0.67±0.13 0.62±0.09 0.57±0.07 TNBC (N = 120) 0.70±0.04 0.72±0.03 0.72±0.15 HER2-enriched (N = 65) 0.65±0.16 0.59±0.19 0.68±0.28 Her2-pure (N = 56) 0.82±0.11 0.86±0.13 0.90±0.20