Federated AI Reads Neonatal Brain Scans Without Moving Patient Data

An artificial intelligence framework that learns to diagnose brain injuries in newborns without hospitals ever sharing a single scan has been reported in Multimedia Tools and Applications. The system, called NeuroSono-Net, was developed by Sultan A. Alotaibi and uses federated learning, a training strategy in which multiple institutions jointly build a machine learning model while every patient image stays inside the hospital that collected it. Instead of moving data, the participating sites exchange only model updates, which are combined by a secure aggregation process so that no individual hospital's contribution can be reconstructed.

The motivation comes from a stubborn bottleneck in medical AI. Neonatal brain ultrasound is a first-line tool for detecting neurological disorders in premature and critically ill infants, but automated diagnostic systems trained at a single center tend to be fragile, because one hospital's scanner settings, patient mix and imaging protocols produce a narrow dataset. Data privacy rules and institutional silos have traditionally prevented the pooling of larger, more diverse collections, leaving many promising algorithms without the breadth of evidence needed for clinical use.

NeuroSono-Net tackles the recognition problem with a multi-scale feature extraction architecture built on attention-guided convolutional neural networks. The model is designed to classify several key conditions seen on neonatal brain ultrasound: intraventricular hemorrhage, periventricular leukomalacia, and hydrocephalus, alongside normal findings. To test the framework, the study constructed five simulated institutional data partitions from publicly available ultrasound datasets, mimicking the distributed conditions under which such a system would realistically operate.

Under federated proof-of-concept simulation conditions, with strong differential privacy guarantees set at epsilon equal to 1.0 and delta equal to 10 to the power of minus 5, the model reached 94.8% accuracy, with a 95% confidence interval of 93.8 to 95.8%. Sensitivity was 94.8%, specificity 94.6%, the weighted F1-score 95.0%, and the area under the ROC curve 0.947. Performance varied across diagnostic categories, and the reported per-class recall figures capture that spread.

The model correctly identified 97.2% of normal cases, 91.9% of severe intraventricular hemorrhage of Grade III to IV, and 93.1% of hydrocephalus cases, while recall was 89.7% for milder Grade I to II hemorrhage and 87.6% for periventricular leukomalacia. The framework was also benchmarked against alternative training strategies. NeuroSono-Net outperformed the strongest federated baseline, MOON paired with a ConvNeXt backbone, by 5.2 percentage points in accuracy, a difference the author reports as statistically significant with p less than 0.001 using McNemar's test.

It also surpassed centralized pooled training, in which all data would be gathered in one place, by 8.4 percentage points, an outcome the study attributes to the benefits of its aggregation approach under the simulated partitioning conditions. Communication efficiency, often the practical weak point of federated systems, received dedicated attention. A novel federated aggregation algorithm reduced the total data transferred during training from 29.7 gigabytes to 13.1 gigabytes over 100 rounds, a 55.9% reduction in total transfer volume, achieved through a 46.8% compression of per-round bandwidth.

Speed at the bedside is another reported strength. Inference latency was measured at 12.3 milliseconds on an NVIDIA A100 GPU and 34.2 milliseconds on an NVIDIA T4, timings the author presents as compatible with real-time workflows in a neonatal intensive care unit. The study's scope deserves careful reading.

The evaluation relied on simulated institutional partitions of publicly available datasets rather than live deployments across real hospitals, so the results demonstrate technical feasibility and proof of concept rather than validated clinical performance in routine care. Even with those caveats, the work addresses a recognized gap in neonatal care: the need for scalable, privacy-preserving diagnostic tools that can learn from distributed data without breaching institutional boundaries. By combining attention-guided multi-scale networks, secure aggregation, an efficient communication protocol and formal privacy guarantees in a single framework, NeuroSono-Net offers a template for how collaborative medical imaging AI might be built in fields where patient data cannot simply be uploaded to a central server.

Subject of Research: Federated deep learning for privacy-preserving multi-institution neonatal brain ultrasound diagnosis

Article Title: NeuroSono-Net federated learning for privacy preserving multi institution neonatal brain ultrasound diagnosis

Article References: Alotaibi, S. A. (2026). NeuroSono-Net federated learning for privacy preserving multi institution neonatal brain ultrasound diagnosis. Multimedia Tools and Applications, 85(11), Article 824. https://doi.org/10.1007/s11042-026-21987-z

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21987-z

Keywords: federated learning, neonatal brain ultrasound, privacy-preserving AI, deep learning, intraventricular hemorrhage, hydrocephalus, periventricular leukomalacia, differential privacy, medical imaging, convolutional neural networks, NICU diagnostics, secure aggregation