Lin Ma
To determine how auxiliary-source selection, annotation compatibility, sampling strategy, and evaluation role affect breast ultrasound lesion segmentation across heterogeneous public cohorts. We conducted a retrospective five-cohort audit comprising 3821 raw cases and 3808 identity-clustered analysis units. A five-fold grouped outer split with shared BUSI inner validation was used for 125 completed formal runs spanning core architectures, corrected-versus-legacy BUS-UCLM retraining, unweighted and source-balanced multi-source training, an all-five multicohort sensitivity, a parameter-count-matched no-Retinex comparator, and targeted PMix/CGate controls. The analysis plan contained 896 executed contrasts in 13 observed Holm families; confidence intervals used 2000-replicate identity-clustered, outer-fold-stratified paired bootstrap sampling. Corrected all-lesion BUS-UCLM training improved RTCMUNet Dice relative to legacy green-only training by 0.1088 (95% CI, 0.0748-0.1436; Holm-adjusted p < 0.001 ) on included-source-held-out BUS-UCLM. Relative to BUSI-only training, corrected BUS-UCLM produced a direct BUS-UCLM Dice gain of 0.5296 but reduced Dice on non-included BUS_UC by 0.0908 and BUS-BRA by 0.1412. Unweighted all-source stacking improved strict-OOD BUS Dice by 0.1522, whereas source balancing showed mixed or null effects. Architecture comparisons were heterogeneous, and the no-Retinex, PMix, and CGate controls did not support universal mechanism superiority after multiplicity adjustment. Auxiliary-source value was conditional on annotation compatibility, sampling strategy, and evaluation role. Role-labelled data-governance audits are therefore necessary before pooled breast-ultrasound training results are interpreted as evidence of external generalization.