Where is the plaque?

Carotid artery atherosclerosis—an asymptomatic yet principal driver of stroke and other cardiovascular events—is routinely monitored with B-mode ultrasound, where the total plaque area (TPA) offers a sensitive, non-invasive marker of risk progression. Precise TPA measurement, however, still depends on labor-intensive manual tracing that suffers from significant inter- and intra-observer variability, limiting throughput in both clinical practice and large-scale epidemiological studies. Over the past decade, researchers have sought to streamline this task: early semi-automated pipelines based on fuzzy region growing, optical-flow tracking, or active-contour “snake” models reduced – but did not eliminate – human input, while first-generation fully automated systems adopted U-Net architectures yet were typically trained on a few hundred images and leveraged ensembles or model stacking to reach competitive accuracy. Building on this foundation, we present a systematic comparison of four U-Net variants that pair a standard decoder with encoder backbones of increasing representational power—VGG19, MobileNet-v2, Inception-ResNet-v2, and Xception—trained and validated on 851 images drawn from 214 patients. All models were optimized end-to-end using the Dice Similarity Coefficient (DSC) loss, and their generalisation was assessed on an unseen 10% test split. The Xception-U-Net achieved the best performance, reaching a DSC of 0.86 without any ensemble boosting, and outperformed a comparably sized Inception-ResNet-v2 model, despite having ~40% fewer parameters. Transfer learning from ImageNet conferred only marginal gains (< 0.02 DSC), highlighting the limited relevance of natural-image features for the speckled and low-contrast visual patterns characteristic of ultrasound imaging. Although overall accuracy is bounded by annotation variability and the modest cohort size, our results demonstrate that a single, computationally efficient model trained on a medium-scale, domain-specific dataset can match or exceed stacked architectures that rely on smaller corpora. By delivering expert-level segmentation in a fraction of a second, the proposed framework enables automated, point-of-care plaque quantification, supports standardised multi-site trials, and contributes to more equitable cardiovascular risk stratification in resource-constrained settings.