Segmentation of Optic Nerve and Lateral Ventricles in Low-Dose Non-Contrast CT with nnU-Net: A Pilot Study

Hao Li, Barite Gutama, Erin N Abbott, Annika Coleman, Matthew Pontell, Ipek Oguz

Keywords: Deep learning, intracranial pressure (ICP), non-invasive, craniosynostosis

Abstract: In this study, we evaluate the nnU-Net, a self-configuring deep learning framework, for automated segmentation of the optic nerve and lateral ventricles from pediatric low-dose non-contrast brain CT scans. The segmentation performance is assessed using the Dice score via cross-validation. The overall mean Dice scores of 0.90 for the ventricles and 0.85 for the optic nerve indicate that nnU-Net can achieve promising segmentation performance in this context. In addition, we had human raters review the segmentations and provide revisions for even slight defects. The segmentations require few or no revisions, supporting the feasibility of structure-specific nnU-Net segmentation in low-dose CT and its potential to streamline annotation in future practical applications.

Submission Number: 70