A comparative analysis of deep learning-based quantitative hemorrhagic CT parameters versus traditional...

Delayed cerebral ischemia (DCI) is the leading cause of poor outcomes after aneurysmal subarachnoid hemorrhage (aSAH). Hemorrhage extent assessment on pre-treatment CT is critical for DCI prediction, but traditional semi-quantitative CT scores remain subjective and imprecise despite their widespread use. This study aimed to assess whether deep learning (DL)-based quantitative hemorrhagic CT parameters outperform traditional scores in predicting DCI.

Patients with aSAH from a prospectively maintained observational registry trial database (the *BLIND* study) were stratified into retrospective (2021.01 to 2023.12), prospective (2024.01 to 2024.12), and external validation (2018.07 to 2024.11) cohort. Hemorrhage was quantified using 3D-UNet. The primary outcome was DCI. Multivariate analyses explored the association between DL-based parameters and DCI. Receiver operating characteristic (ROC) analysis and decision curve analysis (DCA) were used to compare the predictive accuracy and clinical benefit of DL-based parameters (based on SAH volume with the combination of intraventricular hemorrhage, intraparenchymal hematoma, and subdural hematoma) with traditional scores.

The three cohorts were comparable in baseline characteristics. Multivariate binary logistic analysis showed that DL-based parameters significantly correlated with DCI (all p < 0.05). ROC analysis revealed superior predictive accuracy for DL-based parameters compared to traditional scores across all cohorts: AUCs ranged 0.735-0.816 vs. 0.635-0.698 for traditional scores (all adjusted p < 0.05 by multiple DeLong tests). DCA demonstrated more favorable clinical benefit for DL-based parameters.

DL-based quantitative hemorrhagic CT parameters, particularly SAH volume-related parameters, provide significantly better predictive accuracy and clinical benefit for DCI in aSAH patients compared to traditional semi-quantitative CT scores, and may offer a more objective and precise method for future risk stratification.