Abstract Body: Introduction: Identifying patients with basal ganglia intracranial hemorrhage (bgICH) at risk for hematoma expansion (HE) may help to define selection criteria for early surgical evacuation. The most effective radiographic predictor of HE is the spot sign on computed tomography angiography; however, it lacks sensitivity and thus cannot independently rule out patients at-risk for HE. We hypothesized that automated quantification of bgICH CT-density (CTD) was associated with hemorrhage age, could identify patients at-risk for HE, and could augment spot-sign diagnostic accuracy.

Methods: We performed a single-center retrospective review of patients admitted with bgICH between 2013-2024. Patients with a structural bgICH etiology were excluded. We also excluded high intraventricular hemorrhage severity (modified Graeb scale >10), thalamic location, and small bgICH volume (<5-mL) as these patients are poor surgical candidates. Using 678 images from 63 patients, we trained a convolutional neural network to automate bgICH segmentation. Normalized volumetric CT-density (NVCTD) was calculated as the mean bgICH CT-density normalized to the mean CTD of a 1-cm rim of surrounding parenchyma (Image 1). Multivariate logistic regression with area under the receiver operating characteristic curve (AUC) was utilized to evaluate the combined diagnostic accuracy of spot-sign presence and NVCTD. HE was defined as a 10-mL bgICH volume increase within 24 hours of the first CTH.

Results: We included 108 patients. Automated and hand-measured ICH volumes were highly concordant (R²=0.88, p<0.001) (Image 2). HE occurred in 24 patients (22%) and was associated with shorter duration between symptom onset and initial CT (median 1 vs. 3 hours, p=0.006), a lower nv-CTD (median 2.0 vs. 2.2, p=0.011), and a positive spot-sign (41% vs. 5%, p<0.001). NVCTD was positively associated with time to presentation (R² = 0.13, p<0.001) and was negatively associated with HE in spot-sign negative patients (median 2.0 vs. 2.1, p=0.016). Multivariate logistic regression modeling using NVCTD and spot sign as inputs demonstrated improved diagnostic accuracy compared to that of the spot sign alone (AUC 0.80 vs. 0.68, p=0.008) (Image 3). The AUC of NVCTD alone was 0.67 (95% CI: 0.56-0.78), which was statistically similar to that of the spot sign alone (0.68, 95% CI: 0.54-0.82) (p=0.819).

Conclusions: NVCTD is a measure of bgICH acuity and can augment spot-sign basal ganglia HE risk stratification.