Document resource
Introduction/Purpose Intracerebral hemorrhage (ICH) accounts for 10-15% of all strokes with a case fatality rate of 30-40%, underscoring the need for improved imaging biomarkers to stratify prognosis and guide clinical decision-making. Disruption of the corticospinal tracts (CSTs) predicts poor motor and functional outcomes but relies on either expert manual segmentation or costly and time-consuming diffusion tensor imaging (DTI) for detection. We leveraged machine learning to generate approximate CST labels directly from CT scans and evaluate their prognostic value in patients with deep ICH.Materials and Methods We retrospectively identified 307 patients with DTI studies, comprising raw diffusion-weighted imaging and a structural T1, acquired within two weeks of a CT or CT angiography. CST tractography was performed using TractSeg, and the resulting bilateral CST masks were nonlinearly registered onto the corresponding CT using SynthMorph. The unified CST-CT pairs were split 90/10 for training and testing with a 3D cascade full-resolution nnU-Net model. The model was subsequently applied to CT scans from patients with deep ICH enrolled in the MISTIE III trial. ICH segmentation was performed using a previously validated computer vision model, and output CST masks were divided into ips- and contralesional hemispheres. Four imaging features were derived: (1) minimum distance from the ICH centroid to the ipsilesional CST, (2) percent overlap between the ICH and ipsilesional CST, (3) presence of an axial CST interruption, and (4) ipsito-contralesional CST volume asymmetry. Univariate logistic regression evaluated the prognostic value of each feature for modified Rankin Scale (mRS) 0–3 and Glasgow Outcome Scale Extended (GOSE) 4–8 at 365 days.Results Model performance yielded a Dice similarity coefficient of 0.718, beating previously published CT-based CST tractography benchmarks. CST interruption and volume asymmetry were significant predictors of both mRS and GOSE, but not minimum ICH-to-CST distance and percent of ICH-CST overlap. For mRS 0–3, the odds ratios (ORs) for CST interruption and volume asymmetry were 0.212 (95% CI: 0.115–0.388) and 0.00768 (95% CI: 0.000757–0.0780), respectively. For GOSE 4–8, the corresponding ORs were 0.332 (95% CI: 0.184–0.598) and 0.0203 (95% CI: 0.00192–0.214).Conclusion Machine learning can generate CST approximations from CT alone with acceptable segmentation accuracy. CST interruption and volume asymmetry are promising imaging-derived prognostic biomarkers of functional outcome following deep ICH and warrant further investigation as features to inform surgical versus medical decision-making.Disclosures S. Li: None. A. Kashkoush: None. M.E. El-Abtah: None. E. Plow: None. K. Sakaie: None. M. Bain: 2; C; Stryker, Medtronic, Microvention, Cerenovus, Integra, Route 92. 4; C; CIT, Borvo, Algo. S. Raymond: 4; C; VonVascular, Magnendo, Kannact.