Quantitative Perihematomal Imaging Analysis for the Prediction of Intracerebral Hematoma Expansion.

Abstract

Background and purpose: Hematoma expansion (HE) is a major cause of early neurologic deterioration and poor outcome in spontaneous intracerebral hemorrhage. The perihematomal region (PHR), a potential site of secondary vascular injury, remains poorly characterized. We evaluated whether quantitative PHR features provide predictive information for HE beyond hematoma features and insight into HE dynamics.

Materials and methods: In this retrospective 2-center study, 149 patients with acute intracerebral hemorrhage underwent baseline CT within 24 hours of symptom onset and follow-up CT within 72 hours. HE was defined as >6 mL or >33% hematoma growth on follow-up CT. Hematoma regions and PHRs were segmented using relative (half and full radius) and absolute (2, 5, and 10 mm) definitions. Radiomic features were extracted and modeled using multivariable logistic regression. Clinical predictors were added to create combined clinical radiomic models, and performance was evaluated on an independent test set.

Results: HE occurred in 29% (44/149) of patients. The hematoma only radiomic model showed modest discrimination (area under the receiver operating characteristic curve [AUROC] = 0.71; 95% CI, 0.62-0.80). Incorporating PHR features improved performance, with the absolute 5-mm PHR model achieving the highest discrimination (AUROC = 0.91; 95% CI, 0.86-0.95). Adding clinical predictors further enhanced performance (AUROC = 0.94; 95% CI, 0.90-0.97). Key PHR features associated with HE included reduced sphericity, increased texture coarseness, and clustered heterogeneity.

Conclusions: Quantitative PHR characterization was strongly associated with HE and provided incremental value beyond hematoma derived radiomics, with further improvement when combined with clinical predictors. This contrast-free approach may support objective risk stratification, guide targeted interventions, and reduce expansion related morbidity.