Abstract WP298: Scanner-Dependent Variability in CT Clot Radiomics: A Bottleneck to Clinical Translation

Introduction: Radiomics enables quantitative analysis of clot texture, which serves as a proxy for clot biology and may guide thrombectomy and thrombolysis outcomes. However, radiomic features are highly sensitive to scanner and protocol variability, a barrier that has not been systematically evaluated in clot radiomics. We experimentally tested scanner-dependent effects using standardized clot models. Methods: Synthetic clots (40% red blood cells), with and without neutrophil extracellular traps (NETs), were fabricated (n=8) and imaged on two clinical CT systems: GE LightSpeed VCT and Canon Aquilion One (4D CT), using stroke protocols. Clots were segmented, and paired comparisons of volume, mean Hounsfield Units (HU), and 107 radiomic features (RFs) were performed with Wilcoxon tests, FDR correction, log-fold change, and Cohen’s d. Results: Volumes and mean HU were comparable across scanners (GE 405.5±71.2 mm3 vs. Canon 407.6±80.2 mm3, p=0.74; GE 11.3±6.7 HU vs. Canon 9.2±7.5 HU, p=0.35). In contrast, texture RFs diverged sharply: 64 of 107 RFs differed at p1, and 11 demonstrated |Cohen’s d|>1.2. Features most affected included First-order variance and 10th percentile; GLCM cluster tendency, cluster prominence, cluster shade; GLDM gray-level variance, large-dependence low-gray-level emphasis; GLSZM gray-level variance, large-area low-gray-level emphasis; NGTDM complexity and strength. Histogram PDFs confirmed scanner-specific distributions: GE images showed broader, diffuse intensity patterns, while Canon images appeared more condensed near the mean. Qualitatively, GE scans exhibited higher GLCM cluster tendency, reflecting more homogeneous appearance. Conclusions: Gross clot measures (volume, HU) are scanner-robust, but radiomic texture features vary substantially, exposing a critical bottleneck for translation. This is the first study to empirically demonstrate scanner-dependence in clot RFs. Impact: Recognizing and correcting scanner-dependent biases is essential. Solving this problem through harmonization approaches such as ComBat will enable trustworthy, reproducible clot biomarkers across sites. Addressing this bottleneck opens the door for radiomics to reach its full potential as a powerful adjunct to vessel morphometrics, demographics, and comorbidity data in predicting thrombectomy and tPA outcomes, ultimately paving the way for personalized stroke management at scale.