Accurate segmentation of neonatal brain MRI with deep learning

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence

Background:

• Accurate segmentation of 3D Magnetic Resonance Imaging (MRI) scans is crucial for mapping the human brain's connectome.
• Segmentation of perinatal brain MRI presents unique challenges due to developmental variability.
• Deep learning methods require substantial labeled data, which is often scarce for neonatal brain datasets.

Purpose of the Study:

• To develop an automated deep learning pipeline for accurate segmentation of neonatal brain MRI tissues.
• To introduce an age prediction pathway within the segmentation pipeline.
• To investigate strategies for mitigating label scarcity in deep learning for perinatal neuroimaging.

Main Methods:

• Developed an automated deep learning pipeline for neonatal brain MRI segmentation.
• Incorporated an age prediction pathway into the pipeline.
• Evaluated strategies for label distribution (annotated 2D slices across 3D images) and fine-tuning pre-trained models on limited preterm infant data.
• Utilized T1- and T2-weighted MRI scans from the Developing Human Connectome Project (dHCP) cohort (n=709).

Main Results:

• Distributing limited annotated data across a larger number of 3D brain scans enhances segmentation performance.
• Fine-tuning pre-trained models, even partially, outperforms models trained from scratch under label scarcity.
• The proposed pipeline demonstrates effective segmentation and age prediction for neonatal brain MRI.

Conclusions:

• The developed deep learning pipeline offers an automated solution for neonatal brain MRI segmentation.
• Strategies for efficient data annotation and transfer learning are vital for overcoming label scarcity in developmental neuroimaging.
• This work contributes to building accurate human brain connectomes from perinatal MRI data.