Document resource
Introduction Basal ganglia and thalamic hemorrhages (bgICH) have no proven therapies to improve functional outcome. Improved hematoma expansion (HE) prediction in bgICH may identify candidates for early surgical or intensive medical therapy.Objective We developed a computed tomography (CT)-based automated model for bgICH expansion prediction using radiomics and three-dimensional convolutional neural networks (3DCNNs).Methods A single-center retrospective cohort of patients with bgICH (n=267) was combined with the subcortical ICH sub-group of the MISTIE-III trial (n=305). HE was defined as a 10-mL increase in baseline bgICH volume on repeat CT. A late-fusion multimodal 3D-CNN was trained on pre-processed CT images, automatically segmented bgICH/peri-hematoma masks, and LASSO (Least Absolute Shrinkage and Selection Operator)-selected radiomics features. Model evaluation was compared to the BAT (1 point - blend sign; 2 points - any hypodensity; 2 points - presentation <2.5 hours) score using test set AUCs (Area under the receiver operating characteristic curve) from 25 sets of 60/20/20 randomly generated training/validation/test set splits.Results Five hundred seventy-two patients (median age 60 years, 62% male, 78% basal ganglia location, median bgICH volume 28-mL) were included. HE occurred in 124 (22%) patients at a median time of 17 hours (IQR 7.1-32) following baseline CT. Median BAT score was 2 (IQR 1-4) with modest-to-moderate inter-rater reliability for radiographic markers (two raters; Cohen’s kappa 0.46 for hypodensity, 0.62 for blend sign). Automatic bgICH segmentation demonstrated excellent correlation with manual ABC/2 measurements (R 2=0.90, median absolute difference 3-mL) Mean multimodal model test AUC performance for HE was significantly higher than that of the BAT score (0.72 vs. 0.61, p<0.001). The most frequently LASSO-selected radiomic feature was the normalized volumetric CT density (mean bgICH CT density normalized to that of a 5-mm surrounding rim of parenchyma; n=25 runs, 100%; mean LASSO coefficient -0.12).Conclusion CT-based automated computer vision methods can improve upon clinical paradigms for bgICH expansion prediction for real-time clinical decision-making support.Disclosures A. Kashkoush: None. S. Li: None. S. Ghodsara: None. M.E. El-Abtah: None. D. Lilly: None. N. Walborn: None. D. Hanley: None. J. Luo: None. M. Bain: 2; C; Stryker, Medtronic, Microvention, Cerenovus, Integra, Route 92. 4; C; CIT, Borvo, Algo. S. Raymond: 4; C; VonVascular, Magnendo, Kannact.Abstract E-326 Figure 1