Automated Evans index measurement using deep learning in acute subarachnoid hemorrhage: reliability, agreement with experts, and association with external ventricular drainage.
Authors
Affiliations (3)
Affiliations (3)
- Department of Neurosurgery, Weifang People's Hospital, Shandong Second Medical University, Weifang, China.
- Department of Neurosurgery, Laixi People's Hospital, Qingdao, China.
- Department of Neurological Rehabilitation, Weifang Hospital of Traditional Chinese Medicine, Shandong Second Medical University, Weifang, China.
Abstract
The Evans index (EI) is widely used to assess ventricular enlargement and support external ventricular drainage (EVD) decision-making after subarachnoid hemorrhage (SAH), but manual measurement is time-consuming and subject to inter-reader variability. The reliability of automated EI measurement in acute SAH remains insufficiently validated. In this retrospective cohort study, admission non-contrast CT scans from 364 patients with spontaneous SAH were analyzed using TotalSegmentator (TS), an open-source nnU-Net-based deep learning pipeline, to generate automated EI measurements. Each scan underwent two independent inference runs, and two neurosurgical experts independently performed manual measurements. Agreement between TS and expert measurements was evaluated for continuous EI values and EI > 0.30 classification. External ventricular drainage (EVD) placement was used as a pragmatic endpoint reflecting contemporaneous clinical decision-making. Prespecified subgroup analyses excluded frontal horn hematoma or frontal horn periventricular edema. Multivariable logistic regression assessed the association between EI and EVD placement after adjustment for key clinical covariates. TS demonstrated excellent reproducibility between repeated inference runs (ICC = 0.996, 95% CI 0.996-0.997). Agreement between expert readers was high (ICC = 0.983, 95% CI 0.978-0.988). Between-method agreement between TS and expert EI measurements was good in the overall cohort (ICC = 0.76, 95% CI 0.73-0.81) and improved after exclusion of frontal horn hematoma (ICC = 0.87, 95% CI 0.85-0.89). For EI > 0.30 classification, TS identified more positive cases than expert assessment (29% vs. 17%), with moderate Cohen's kappa (0.57, 95% CI 0.54-0.60). TS-derived EI demonstrated discrimination for EVD placement (AUC = 0.75, 95% CI 0.73-0.79), approaching expert-derived EI (AUC = 0.80, 95% CI 0.78-0.83). After covariate adjustment, TS-derived EI remained independently associated with EVD placement (adjusted OR = 1.09, 95% CI 1.03-1.17; <i>p</i> = 0.009). Automated EI measurement using TS provides reproducible and clinically informative assessment of ventricular enlargement on CT in acute SAH. Although threshold-sensitive disagreement occurred near EI = 0.30, automated EI showed meaningful agreement with expert assessment and remained independently associated with contemporaneous EVD decision-making. Further SAH-specific refinement may improve robustness in hemorrhage-related ventricular distortion.