MUSE-Net: a SegFormer-B2-based multi-expert ensemble for multicenter carotid plaque segmentation in ultrasound images

Accurate carotid atherosclerotic plaque segmentation in B-mode ultrasound is essential for quantitative vascular assessment, but multicenter variability in scanners, views, image quality, and annotation practice limits model robustness. We propose MUSE-Net, a SegFormer-B2-based multi-expert ensemble. Each expert contains a long-axis binary plaque branch and a short-axis three-class lumen/plaque branch. Each branch uses a MiT-B2 encoder and SegFormer all-MLP decoder. MUSE-Net combines five-fold expert training, long-axis test-time augmentation, probability-level aggregation, and connected-component refinement. On a four-center dataset, MUSE-Net achieved a mean Dice similarity coefficient of 0.834 and mean intersection-over-union of 0.748. Paired Wilcoxon testing showed significant case-wise gains over all baselines (win rate, 57.6-81.8%; all [Formula: see text]; rank-biserial effect size, 0.274-0.731). Component-wise ablation identified backbone capacity and five-expert aggregation as the principal contributors to performance. Center-wise and leave-one-center-out analyses indicated robustness under heterogeneous acquisition, but do not establish broad patient-level generalizability. Because strict patient-level separation could not be retrospectively verified for all historical centers, performance may be optimistically biased. The current evidence supports segmentation and pixel-based plaque-area quantification, which indicate that our MUSE-Net has the potential in clinical outcome prediction or routine deployment in the future.