Authors: Jennifer Lauren Quon, MD; Lily Kim, BA; Michelle Han, BS; Edward Lee, PhD; Samuel Cheshier, MD, PhD; John Kestle, MD, PhD; Robert Lober, MD, PhD; Kristen Yeom, MD; Michael Taylor, MD, PhD; Gerald Grant, MD; Michael Edwards, MD (Stanford, CA)

Introduction: This year marks the centennial of the development of ventriculography by Walter Dandy in 1918. Significant strides have since been made in neuroimaging, and visualization of the ventricles has become routine in hydrocephalus management. However, we lack a readily available method to measure ventricular volume in clinical settings. Management of hydrocephalus is further complicated by the paucity of data on normal ventricular development and volumetric changes that occur with intracranial pathologies. We conducted a multi-institutional study to develop a deep-learning model designed to automatically detect ventricular volume. Methods: Six hundred pediatric brain magnetic resonance images (MRIs) (400 normal, 200 hydrocephalus) were divided into training (n=420), validation (n=90), and testing (n=90) datasets. All normal scans were obtained from patients between the ages of 0 and 18 evaluated at our institution between 2011-2017. The hydrocephalus cohort was selected among patients with posterior fossa tumors who had radiologic evidence of ventriculomegaly. Scans from this cohort included 50 MRIs from our institution and 150 from three outside academic centers. Expert segmentation of the ventricles was performed on all T2-weighted and thin-slice T1-weighted spoiled gradient recalled acquisition in the steady state (SPGR) scans. Our encoder-decoder convolutional neural network architecture consisted of a UNet with a pre-trained ResNet50 encoder. Results: Manual segmentation served as the ground truth for ventricle delineation. “True” ventricular volume was based on SPGR sequences. Dice score for model performance was 0.8699 for automatic segmentation of ventricle volume for T2-weighted sequences. Predicted ventricular volume was, on average, within 8% of the manually determined volume. Conclusion: In this multi-institutional study, we present a deep learning model that automatically segments ventricles and outputs volumetric information. This clinically applicable and externally validated tool may enhance our current understanding of ventricular development and facilitate accurate ventricular volumetric measurement in the clinical setting.