Reliability-aware dual-representation DCE-MRI segmentation with disagreement-guided quality control for breast tumor delineation

Abstract

Automated breast tumor segmentation on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) may support quantitative imaging analysis, but case-level segmentation failures remain a concern in heterogeneous multicenter data. We evaluated a reliability-aware dual-representation framework using the MAMA-MIA dataset. Five-fold out-of-fold (OOF) predictions from the 1,200-case development cohort were used exclusively to calibrate low-confidence burden (LCB)-based quality-control (QC) rules, which were then applied unchanged to final predictions from a separate 306-case held-out cohort. Raw3P used the pre-contrast and two post-contrast phases, whereas Sub513 used the pre-contrast phase and two subtraction maps. Separate nnU-Net v2 models were trained, and the binary-intersection ensemble was evaluated as a conservative consensus segmentation strategy. The resulting binary-intersection mask achieved mean Dice 0.763, precision 0.786, recall 0.814, 95th-percentile Hausdorff distance 40.16 mm, average symmetric surface distance 11.70 mm, normalized surface Dice at 2 mm 0.769, and absolute relative volume error 74.9%. Compared with the probability-averaging ensemble, the mean Dice gain was modest (0.0044; paired bootstrap 95% CI, 0.0013 to 0.0075), while recall decreased by 0.0166. At the frozen 40%-OOF-targeted point, 159/306 cases (52.0%) were referred for review, including 33/35 poor-quality masks (94.3%); the two missed poor-quality masks were false-negative QC cases among the 147 cases not referred. Dual-representation outputs therefore provided a useful case-level QC signal, but not a fail-safe or fully autonomous segmentation system.