Abstract 019: Radiomics analysis to predict etiology in large vessel occlusions.

Introduction Large vessel occlusions (LVO) are responsible of 40% of acute ischemic strokes (AIS). Nonetheless, nearly 30% of cases lack a clear source. Radiomics has emerged as a non‐invasive imaging tool that analyzes voxel‐by‐voxel signal intensity. We aim to correlate the histology of clots retrieved from LVO with radiomic features (RFs) to assess stroke etiology. Methods Patients diagnosed with AIS due to LVO between 2019 to 2024 were analyzed. Ten clots retrieved from mechanical thrombectomy were imaged using micro‐computed tomography (micro‐CT) and histologically analyzed. The three main clot components were red blood cells (RBCs), fibrin and calcium. 3D Slicer was used for 3D segmentations of each component. One hundred and thirty RFs were extracted using Py‐radiomics add‐in. NCCT and computed tomography angiography images were co‐registered with the corresponding slices obtained from histology. A cohort of 426 patient with NCCTs obtained at the time of presentation were then analyzed. Only clots with an identifiable hyperdense sign in NCCT were included. Stroke etiology was adjudicated based on TOAST, in cardioembolic, large artery atherosclerosis (LAA) and cryptogenic. Results The RFs total energy (TE), joint average (JA) and large dependence high gray level emphasis (LDHGLE) were specific for RBCs in micro‐CT (p<0.001, p 0.003, and p<0.002, respectively). Corresponding RFs of fibrin included minimum (MI) and 10 percentile (p .005, and p<0.001, respectively), while calcium RFs included difference variance (DV, p 0.05). TE, JA and LDHGLE were strongly correlated with clots that had at least 70% RBCs (Rho 0.654 and 0.652, respectively). MI was strongly correlated with clots that had at least >80% of fibrin (Rho 0.795). DV was not correlated to calcium composition in clots (Rho 0.400). The RFs of the NCCT segmentations of these clots had strong correlation with corresponding micro‐CT values of TE (Rho 0.687) JA (Rho 0.809) and LDHGLE (Rho 0.657). MI was negatively correlated (Rho ‐0.851). RFs analysis thresholds in the identification of clot components showed that TE (AUC 0.800, sensitivity 0.750, specificity 0.800, cutoff 38244.3089) and LDHGLE (AUC 0.750, sensitivity 0.750, specificity 0.800, cutoff: 52.64) had significant accuracy for determining clots with higher RBC composition in NCCT. MI (AUC: 0.350, sensitivity: 0.500, specificity: 0.400, cutoff: 32.3) did not have accuracy in the identification of fibrin. A total of 145 of 426 patients were included in the final analysis due to image quality purposes. Fifty patients have a stroke of cardioembolic origin, 45 due to LAA and 50 were cryptogenic. Applying TE and LDHGLE thresholds, thirty (60%) cardioembolic, 12 (27%) of LAA and 21 (42%) of cryptogenic strokes were mainly composed of RBCs (> 70% of the clot). Conclusion RFs are sensitive and specific to determine clots rich in RBC composition in NCCT. Cardioembolic clots have higher RBCs compared with LAA and cryptogenic. Radiomic analysis is a promising non‐invasive tool to determine stroke etiology.