Deep Learning Can Now Predict Brain Outcomes in Preterm Infants from Routine Ultrasound

Predicting neurodevelopmental outcomes in very preterm infants currently relies on waiting. Waiting for developmental milestones. Waiting for assessments at 6 months, 12 months, sometimes 3 years before a definitive diagnosis of impairment is made. By then, the window for the earliest and most effective interventions has narrowed significantly.

A new study funded by the RSNA R&E Foundation is trying to change that timeline by using deep learning to extract predictive information from images that are already being acquired: routine cranial ultrasound.

The Problem with Current Approaches

Cranial ultrasound (CUS) is standard practice for very preterm infants (born before 31 weeks). It is performed at three time points: within the first week, at six weeks, and near term-equivalent age. Radiologists use it to screen for brain injury like intraventricular hemorrhage.

But here is the gap: while severe abnormalities on CUS are known predictors of neurodevelopmental impairment (NDI), outcomes for apparently "normal" scans are highly variable. Some infants with clean ultrasounds still develop cerebral palsy, cognitive delay, or other impairments. Traditional logistic regression models struggle with this complexity.

What the AI Models Found

Dr. Tahani Ahmad from Dalhousie University and her team built three different AI models using data from very preterm infants born in Nova Scotia between 2004 and 2016.

The first model used a convolutional neural network (EfficientNetB0) to automatically classify CUS images as normal or abnormal. It performed well enough to standardize and accelerate ultrasound interpretation, flagging when it was confident versus uncertain about a result.

The more ambitious models combined imaging data with clinical variables to predict NDI outcomes directly. Deep learning outperformed traditional statistical approaches, capturing nonlinear relationships between imaging features and clinical data that conventional methods miss.

Why Radiologists Should Care

This research repositions cranial ultrasound from a screening tool that detects obvious injury to a predictive tool that can stratify risk for outcomes that will not manifest for years. If validated at scale, it means the neonatal CUS exam you are already reading could carry far more clinical weight than the current report conveys.

For pediatric radiologists, this is particularly relevant. The interpretation of "normal" neonatal cranial ultrasound may need to evolve beyond binary normal/abnormal reads toward quantitative risk assessments powered by AI.

Implications for Indian NICUs

India has one of the highest burdens of preterm births globally. NICUs across the country perform cranial ultrasound routinely, but follow-up developmental assessments are inconsistent, especially outside major academic centers. An AI tool that can flag high-risk infants from imaging data already being collected could bridge this gap, enabling earlier referral to developmental intervention programs even in resource-limited settings.

This is still early-stage research, but the direction is clear. Routine imaging data contains predictive information we are not yet extracting. Deep learning is starting to change that.