Unsupervised abnormality detection in neonatal MRI brain scans using deep learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging Analysis

Background:

• Neonatal brain MRI analysis lacks robust unsupervised methods.
• Neonatal encephalopathy (NE) and other early-onset conditions require better diagnostic biomarkers.
• Early identification of neonatal brain anomalies is crucial for lifelong outcomes.

Purpose of the Study:

• To develop and assess an AI framework for neonatal brain MRI analysis.
• To improve the identification and prognostication of neonatal brain conditions.
• To provide an assistive tool for neuroradiologists in interpreting neonatal brain scans.

Main Methods:

• Utilized deep convolutional Autoencoder (AE) unsupervised learning architectures.
• Developed a framework to learn the normal structure of neonatal brains.
• Tested the framework on the Developing Human Connectome Project (dHCP) dataset (97 patients).

Main Results:

• The framework effectively identified subtle morphological signatures in neonatal brain structures.
• Normal and abnormal neonatal brain scans were distinguished with up to 83% accuracy.
• The AI identified previously missed anomalies, later corroborated by a neuroradiologist.

Conclusions:

• The proposed AI framework enhances the analysis and assessment of neonatal brain MRI scans.
• This approach shows promise in improving prognostication for neonatal brain conditions.
• The framework can effectively localize novel brain anomalies in neonatal MRI data.