Advancements in Cranial Ultrasound for Early Detection of Neurodevelopmental Impairment in Preterm Infants

Recent advancements in the use of cranial ultrasound (CUS) may offer significant improvements in the early identification of neurodevelopmental impairment (NDI) in very preterm infants (VPI). Born before 31 weeks of gestation, these infants face a high risk of NDI, including cerebral palsy and cognitive delays. Current predictive methods often result in a definitive diagnosis only by age three, potentially missing opportunities for early intervention.

Dr. Tahani Ahmad, a full professor in the Department of Radiology at Dalhousie University, used her R&E Foundation Philips/RSNA Research Seed Grant to investigate the prognostic value of CUS performed in the neonatal period. The goal of her research was to facilitate prompt clinical decision-making and early access to necessary interventions by predicting developmental outcomes much earlier in life.

Traditional prediction using logistic regression faces limitations in handling complex medical imaging data. To address this, Dr. Ahmad’s team investigated the use of elastic net regression and deep learning models to process and analyze CUS images. They found that deep learning, in particular, could efficiently identify meaningful patterns that are challenging to pinpoint through conventional methods.

The study analyzed a cohort of infants born at 22-30 weeks of gestation in Nova Scotia, Canada, excluding those with congenital anomalies or who were receiving palliative care. The research used data from the Nova Scotia Provincial Perinatal Follow-Up Program database, along with radiological data from PACS, focusing on routine CUS images obtained at various neonatal stages.

Dr. Ahmad's team developed three artificial intelligence (AI) models. The first model used CUS images to automatically detect abnormal findings, demonstrating effectiveness with a deep learning model known as EfficientNetB0. The second model combined CUS images with clinical variables, significantly improving predictive performance compared to using clinical data alone. The third model relied solely on clinical variables and traditional radiology reports, where machine-learning methods provided more accurate predictions than logistic regression.

The findings underscore the potential for AI-enhanced CUS analysis to improve early risk stratification and guide intervention strategies for high-risk preterm infants. Dr. Ahmad highlights the need for further validation across diverse populations and clinical settings. Plans include developing a user-friendly application for neonatologists to access predictive insights at the point of care.

These initiatives, supported by the R&E Foundation Philips/RSNA Research Seed Grant, set the foundation for broader studies, aiming to enhance both patient outcomes and research in neonatal care.