Monitoring the lateral ventricles in the presence of intracranial hemorrhage using automated dual segmentation

Segmentation of intracranial hemorrhage (ICH) alongside the brain's ventricles can provide crucial information in the management stroke or traumatic brain injury (TBI). Automated computed tomography (CT) based segmentation is time-efficient and desirable when manual annotation is not feasible.

To develop a fully automated and simultaneous segmentation method capable of quantifying ICH and the lateral ventricle volumes. Segmentation results are evaluated in patients with a range of TBI conditions and patients with hemorrhage in the ventricles, known as intraventricular hemorrhage (IVH).

Five deep learning models were trained using 154 non-contrast CT images with manual annotations of the ICH and lateral ventricles. Model performance was evaluated using Dice and Hausdorff Distance metrics. A top performing model was applied to three clinical samples: 1) N = 591 patients with mild TBI without ICH; 2) N = 91 moderate-to-severe TBI with baseline and follow-up CT; and 3) N = 5 patients with IVH and repeat CT. Statistical tests assessed model performance, and one model was selected to: test for relationship between lateral ventricle volume and age and sex, and investigate volume estimates from baseline to follow-up.

Significant segmentation differences were observed between models for the Dice score (F = 3.3, p = 0.012), but not Hausdorff Distance (F = 0.58, p = 0.68). The 3D nnU-Net model had the highest performance with a mean Dice score of 0.92 ± 0.05 for the lateral ventricles and 0.88 ± 0.05 for ICH. In mild TBI, a quadratic association between lateral ventricle volume and age was found for the whole sample and after stratifying by sex. There were six ICH false positives out of 591 with mild TBI. In moderate-to-severe TBI patients, volume changes in both lateral ventricle and ICH were observed from baseline and follow-up (p < 0.04, Wilcoxon signed rank test). Follow-up lateral ventricle volume was significantly associated with the baseline estimates, but this was not the case for serial ICH estimates. The IVH case series demonstrated the feasibility of measuring lateral ventricle volume changes despite the presence of IVH.

Automated dual segmentation of lateral ventricles and ICH using deep learning provides a reliable method to monitor TBI severity over time. The approach yielded clinically relevant information, including ICH and lateral ventricle volume changes across sessions. This approach offers an efficient analysis of CT images in ICH and particularly patients that develop IVH.