Radiomic Feature-Based Stratification of Stroke Patients Using CT Imaging: A Multi-Institutional Study
Abstract
Purpose
Radiomics enables the extraction of high-dimensional quantitative features from medical images, capturing information related to lesion shape, intensity distribution, and texture that extends beyond conventional visual assessment. Unlike deep learning approaches, which typically require large datasets, radiomic analysis can be effectively applied to smaller cohorts, making it particularly suitable for clinical studies with limited samples. This study aimed to evaluate the ability of computed tomography (CT)–based radiomic features to differentiate between hemorrhagic and infarct strokes.
Methods
A total of 131 patients with acute stroke were retrospectively enrolled from two independent institutions. Non-contrast CT scans were acquired using a Siemens 128-slice multidetector CT scanner at Institute 1 and a Siemens SOMATOM Definition True64 scanner at Institute 2. Image preprocessing included voxel resampling, skull stripping, and intensity normalization to ensure feature reproducibility. Radiomic features encompassing shape descriptors, first-order statistics, and texture characteristics were extracted using 3D Slicer software. Statistical significance between hemorrhagic and infarct stroke groups was assessed using Student’s t-test with a significance threshold of p < 0.05. The discriminative performance of significant features was evaluated using the area under the receiver operating characteristic curve (AUC).
Results
The results demonstrated significant differences in radiomic features between hemorrhagic and infarct strokes across both institutions. At Institute 1, mean AUC values reached 0.9453 for hemorrhagic stroke and 0.9203 for infarct stroke. At Institute 2, even higher diagnostic performance was observed, with mean AUCs of 0.9696 and 0.9623, respectively. Shape- and texture-based features showed particularly strong discriminatory capability.
Conclusion
CT-based radiomic analysis exhibits excellent potential for non-invasive differentiation between hemorrhagic and infarct strokes. Although the findings are promising, further studies incorporating larger, multi-center datasets, standardized imaging protocols, and external validation are necessary to support clinical translation and routine implementation.