Region-specific lesion size and low-HU burden as prognostic markers in intracerebral haemorrhage: an exploratory CT analysis

Abstract Background Spontaneous intracerebral haemorrhage (ICH) is associated with high mortality and poor functional outcomes. The prognostic role of density-based CT measurements is controversial. This study aimed to investigate the prognostic value of region-specific Hounsfield Unit (HU)-derived features, focusing on the lower tail of the HU distribution and low-HU burden. Methods This retrospective study included 132 patients with spontaneous ICH. Three-month functional outcomes (modified Rankin Scale, mRS) defined good (mRS 0–3, n = 69) and poor (mRS 4–6, n = 63) outcome groups. Admission CT images were segmented semi-automatically into haematoma core, perihematomal oedema, and core+oedema regions. Density-based features, including lower-tail HU metrics ( p01 , ratio_low10 , and mean_low20 ) and morphological features were extracted. Group differences were assessed using the Mann–Whitney U test and Cohen’s d. Benjamini–Hochberg false discovery rate (FDR) correction was applied within each region across ten pre-specified features, and intra-observer reproducibility was assessed in two independent repeat-segmentation samples (ICC). Results Poor outcomes were associated with larger lesion sizes and higher low HU burdens. Morphological markers of lesion extent (minor_axis_mm and lesion area, mm²) were significantly higher in the poor-outcome group across all regions (p < 0.05). Lower-tail HU features showed univariate associations with poor outcome; after within-region FDR correction, five oedema-region associations remained significant ( minor_axis_mm , minHU , p01 , ratio_low10 , and mean_low20 ; adjusted p = 0.045–0.046). Intra-observer reproducibility was excellent for morphological size features but poor or inconsistent for ratio_low10 and mean_low20 . Median HU showed no consistent association with outcome. On multivariable logistic regression adjusting for age, admission GCS, and hematoma location, lesion minor axis length in the perihematomal oedema region remained independently associated with poor outcome (OR = 1.02, 95% CI 1.00–1.04, p = 0.030), improving model discrimination beyond clinical predictors alone (AUC 0.82 to 0.84; likelihood ratio p = 0.025). Conclusions Lesion size was the only feature supported by both an independent association with outcome and consistently high intra-observer reproducibility. Lower-tail HU features showed univariate associations, but their independent incremental value was not established, and the reproducibility of ratio_low10 and mean_low20 was limited or inconsistent. These exploratory, single-centre, single-slice findings are conditional on the selected slice and require confirmation in larger, preferably volumetric, externally validated cohorts before clinical application.