CEREBLEED: Automated Quantification and Severity Scoring of Intracranial Hemorrhage on Noncontrast CT

Standardized interpretation of intracranial hemorrhage (ICH) severity on noncontrast computed tomography (NCCT) is limited by the absence of objective, reproducible tools for quantifying lesion burden and its anatomic impact. We developed and externally validated a deep learning-based framework for automatic segmentation and volumetric quantification of ICH, deriving a quantitative Severity Index based on volumetric relations among hemorrhage subtypes and brain structures, and prospectively assessed its clinical applicability.

A total of 2112 NCCT scans were analyzed: 1110 for training and internal evaluation (900 retrospective, 200 prospective) and 1002 from external data sets (503 hospital cohort, 499 public database). Three no-new U-Net segmentation models addressed total hemorrhage, subtype differentiation, and brain structure delineation. Segmentation performance was evaluated in internal and external cohorts using overlap and volumetric similarity metrics. The Severity Index was prospectively correlated with expert visual grading and its ability to predict urgent neurosurgical intervention.

The total hemorrhage model achieved median Dice scores of 0.90 (95% CI 0.89-0.91) internally and 0.70 (0.69-0.71) externally, with volumetric similarity of 0.96 (0.95-0.97) and 0.83 (0.82-0.84), respectively. The Severity Index correlated with expert-rated severity (H = 39.6, P < .001; ε2 = 0.39, 95% CI 0.24-0.55) and predicted the need for neurosurgical intervention (area under the curve = 0.83, 95% CI 0.74-0.92). A threshold of ∼300 yielded sensitivity 0.87 (0.69-0.96) and specificity 0.83 (0.72-0.91).

This framework provides standardized, interpretable quantification of ICH severity. The Severity Index may support surgical triage and improve interdisciplinary communication in acute neurocritical care.