High Resolution Ultrasound Radiomics Differences in Carotid Plaques between Patients with and without Ischemic Stroke

Background High-resolution ultrasound (HRUS) provides improved spatial resolution for superficial vascular structures, while radiomics quantifies image intensity and texture patterns. However, the robustness of HRUS radiomics differences between carotid plaques from patients with and without ischemic stroke (IS) remains uncertain. This study aimed to identify radiomics features robustly associated with IS. Methods This retrospective study included a primary cohort of 1,148 patients (361 with IS and 787 without IS) and an additional population of 83 patients with IS and moderate or severe carotid plaque. We extracted 474 radiomics features in the primary cohort. Features meeting the reliability criterion were analyzed primarily using propensity score overlap weighting for eight clinical covariates. Associations were also assessed using covariate adjusted logistic regression, with false discovery rate correction applied to the overlap-weighted and covariate-adjusted analyses. LASSO selection stability was evaluated in bootstrap samples after penalty selection by the one standard error rule. A separate 1,000 iteration bootstrap of the full modeling pipeline provided internal validation. An expanded clinical model, interaction analyses, and a comparison of the two IS case cohorts were also performed. Results A total of 412 of 474 features (86.9%) met the ICC threshold of 0.80. After overlap weighting, all eight clinical covariates had absolute standardized mean differences below 0.01 (effective sample size = 817.2). Of the 403 estimable features, 73 were associated with IS after false discovery rate correction. Seven features in the primary cohort met all robustness criteria. Wavelet-LH_glrlm_RunEntropy showed the strongest inverse association in the weighted analysis (OR per SD, 0.60; 95% CI, 0.49 to 0.74; q < 0.001). In the expanded clinical model, all seven features retained the same association direction, and one remained significant after false discovery rate correction. Conclusion HRUS radiomics profiles differed between patients with and without IS in the overlap-weighted analysis. Seven features met the prespecified robustness criteria. These findings suggest that plaque echogenicity and spatial texture are associated with IS. Prospective multicenter validation is needed.