Automated Ventricular System Segmentation in CT Images of Deformed Brains Due to Ischemic and Subarachnoid Hemorrhagic Stroke.

Accurate ventricle segmentation is important for reliable automated infarct localization, detection of early ischemic changes, and localization of hemorrhages. The purpose of this study was to develop a robust and accurate ventricle segmentation method in image data of ischemic and hemorrhagic stroke patients. Early follow-up non-contrast CT image data of 35 patients with a clinical diagnosis of ischemic stroke or subarachnoid hemorrhage were collected. We proposed a ventricle segmentation method based on a combination of active contours and an atlas-based segmentation. Ground truth was obtained by manual delineation of the ventricles by 4 observers with corrections by 2 experienced radiologists. Accuracy of the automated method was evaluated by calculation of the intraclass correlation coefficients, Dice coefficients, and by Bland-Altman analysis. The intraclass correlation coefficient for the automated method compared with the reference standard was excellent (0.93). The Dice coefficients was 0.79 [IQR: 0.72-0.84]. Bland-Altman analysis showed a mean difference of 2 mL between the automatic and manual measurements, with broad limits of agreement ranging from -18 to 15 mL. The automated ventricle segmentation showed an excellent correlation and high accuracy compared to the manual reference measurement. This approach is suitable for reliable ventricle segmentation even in stroke patients with a severely deformed brain.