Appearance-aware robustness probes for strict external breast ultrasound segmentation under multi-source joint...

Breast ultrasound segmentation is sensitive to external-domain shift caused by scanner, acquisition, annotation and appearance variation. Shadowing, gain and contrast variation are particularly relevant given that ultrasound is not an optical imaging modality, yet many segmentation reports still rely mainly on randomly split or source-overlapping validation. We reformulated the study as a strict external validation analysis. Six model configurations-U-Net, CMU-Net, RTCMUNet, PMix, CGate and PMix+CGate-were evaluated under three source-composition protocols: all-source joint training, leave-BUS-UCLM-out training and leave-BUS-BRA-out training. BUS () and BUSI-WHU () were held out from training and validation under all protocols and used as fixed strict external out-of-distribution cohorts. The primary endpoint was image-level intersection over union (IoU), with paired image-level bootstrap (10 000 resamples) used to estimate 95% confidence intervals. A targeted source-inclusion sensitivity analysis additionally compared two training-pool compositions for RTCMUNet and PMix+CGate that differed only in whether a small set of previously excluded malignant-red lesion images from one source was added. PMix+CGate under all-source joint training achieved the highest average strict external out-of-distribution IoU across the two held-out cohorts (equal-domain mean 72.19). Against RTCMUNet trained without BUS-UCLM, the cross-protocol delta was +1.32 percentage points; the 95% CI [0.33, 2.98] crossed zero and did not support a cross-protocol advantage. In the source-inclusion sensitivity analysis, RTCMUNet and PMix+CGate responded in opposite directions to the same compositional change, producing a model-by-composition interaction in IoU (4.31 percentage points, 95% CI [6.66,2.15]) whose image-sampling interval excluded zero. PMix+CGate achieved the highest average strict external out-of-distribution IoU under all-source joint training, but its cross-protocol advantage over RTCMUNet was not supported by an interval that crossed zero. The source-inclusion analysis further indicates model- and cohort-dependent sensitivity to training-pool composition, rather than a universal benefit from adding data or a universally preferred configuration.