Project Details

Quantifying Ventricular Volume from Neonatal Head Ultrasound Using Deep Learning

Kee Chung, MD MBA

Summary

BACKGROUND: Neonatal hydrocephalus is a serious condition that can lead to brain injury if worsening ventricular enlargement is not recognized and treated in time. Cranial ultrasound is the standard bedside tool for follow-up, but assessment still depends largely on visual judgment or simple 2D measurements rather than a true volume measurement. GAP: There is currently no validated bedside way to measure ventricular numerical volume from routine neonatal cranial ultrasound, so clinicians still rely on limited 2D measurements, holding back more objective monitoring to further progress in care and research. HYPOTHESIS: Routine neonatal cranial ultrasound can be used to estimate ventricular volume accurately using a deep learning model using MRI as the reference standard. The deep model-derived ventricular volume will detect progression better than standard 2D measures. METHODS: We will retrospectively study 200 neonates with standard cranial ultrasound and brain MRI obtained within 24 hours across a range of ventricular enlargement, using AI to generate numerical estimates of ventricular volume from the ultrasound images and compare these with MRI-derived ventricular volume as the reference standard. A smaller longitudinal cohort with serial imaging will then be used to compare AI-derived volume with standard 2D ultrasound measures for detecting progression. RESULTS: Pending. IMPACT: This approach would improve serial monitoring of neonatal hydrocephalus by providing an objective quantitative measure with routine cranial ultrasound, resulting in better bedside monitoring to centers without advanced subspecialty expertise or ready access to MRI. Additionally, it could provide a more standardized quantitative marker for clinical guideline formation and clinical research.