Cross-cohort reliability and failure attribution in AI-based breast DCE-MRI tumor segmentation

Purpose Reliable breast tumor segmentation in dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) requires performance that generalizes beyond the development cohort. We evaluated cross-cohort reliability and patient-level failure mechanisms. Methods We used 982 I-SPY2 cases for patient-level five-fold internal validation and 526 leakage-controlled non-I-SPY2 MAMA-MIA cases for external testing. Five DCE-derived input volumes were used to compare 2D nnU-Net, 3D nnU-Net, residual encoder nnU-Net, probability-map ensembles, and SwinUNETR. The strongest model was evaluated under direct zero-shot transfer, timing-aware phase selection, and largest connected component (LCC) postprocessing. Results The residual encoder nnU-Net achieved the strongest internal performance (mean Dice 0.786; IoU 0.667; median HD95 3.74 mm; median ASSD 0.95 mm). Zero-shot external Dice was 0.547, increasing to 0.555 with timing-aware phase selection and 0.574 with LCC postprocessing. LCC-collapse occurred in 43/526 external cases (8.2%). Opposite-image-half component selection was strongly associated with collapse (odds ratio 44.1, 95% CI 20.4–95.3; Fisher exact p \< 0.001 ). Volume-bound analysis showed that external degradation reflected both volume mismatch and additional spatial or boundary error. Conclusion Favorable internal and cohort-average segmentation metrics can obscure substantial cross-cohort and postprocessing-related failures. Leakage-controlled external validation and explicit case-level failure attribution provide evidence concerning cross-cohort transportability and failure mechanisms but do not establish prospective clinical utility.