Multi-expert AI network sharpens carotid plaque segmentation across ultrasound centers

Researchers have developed MUSE-Net, a deep learning model built on the SegFormer-B2 architecture that combines multiple expert networks to segment carotid atherosclerotic plaques in B-mode ultrasound images. In a study published in Medical & Biological Engineering & Computing, the team reports that the ensemble achieved a mean Dice similarity coefficient of 0.834 and a mean intersection-over-union of 0.748 on a dataset drawn from four centers, outperforming all comparison models in paired statistical testing.

Accurate segmentation of plaques in ultrasound is a prerequisite for quantitative vascular assessment, but models trained at a single institution often falter when faced with images from other hospitals. Differences in scanner hardware, imaging views, image quality and annotation practice among centers can degrade performance. MUSE-Net was designed specifically to address this multicenter variability. Each expert within the ensemble contains two branches: a long-axis branch that performs binary plaque segmentation and a short-axis branch that classifies pixels into three categories covering the lumen and plaque. Both branches use a MiT-B2 encoder paired with a SegFormer all-MLP decoder.

The full pipeline stacks several techniques on top of the expert networks. The model is trained with five-fold expert training, and at inference time long-axis test-time augmentation generates multiple predictions per case. These are aggregated at the probability level, and a connected-component refinement step cleans up the final segmentation masks. According to the authors’ ablation experiments, backbone capacity and the aggregation of five experts were the principal contributors to the observed performance.

The statistical evaluation was deliberately strict. Paired Wilcoxon tests showed significant case-wise gains over all baselines, with win rates ranging from 57.6 to 81.8 percent, all p-values below 0.05, and rank-biserial effect sizes between 0.274 and 0.731. Center-wise analyses and a leave-one-center-out scheme, in which one center is held out entirely during evaluation, indicated that the model remains robust under heterogeneous acquisition conditions. The authors are explicit, however, that these analyses do not establish broad patient-level generalizability, and the leave-one-center-out results should not be read as proof of performance in unseen clinical populations.

The study also carries an important caveat about its own data. Because strict patient-level separation could not be retrospectively verified for all of the historical centers included in the dataset, the reported performance may be optimistically biased. The authors state that the current evidence supports the model for segmentation and pixel-based plaque-area quantification, and that MUSE-Net has potential for clinical outcome prediction or routine deployment in the future, rather than demonstrated readiness for either today. For now, the work offers a documented approach to building segmentation models that hold up across the scanner and protocol differences that routinely undermine medical image analysis in multi-center settings.

Subject of Research: A deep learning ensemble for multicenter carotid plaque segmentation in ultrasound images

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

Article References: Hao, M., Huang, J., Song, J., Chen, M., Liu, C., Song, X., Gong, K., Fan, L., Ge, H., Li, Y., & Qian, C. (2026). MUSE-Net: a SegFormer-B2-based multi-expert ensemble for multicenter carotid plaque segmentation in ultrasound images. Medical & Biological Engineering & Computing. https://doi.org/10.1007/s11517-026-03695-w

Image Credits: AI Generated

DOI: 10.1007/s11517-026-03695-w

Keywords: carotid plaque, ultrasound, image segmentation, deep learning, SegFormer, ensemble learning, multicenter study, medical imaging, atherosclerosis, Dice coefficient, MUSE-Net, SegFormer-B2-based