Deep learning microstructure estimation of developing brains from diffusion MRI: a newborn and fetal study.

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Computational Neuroscience

Background:

• Diffusion-weighted magnetic resonance imaging (dMRI) is crucial for evaluating brain white matter structure.
• Estimating fiber orientation distribution functions (FODs) typically requires extensive dMRI measurements, posing challenges for vulnerable populations like newborns and fetuses.
• Existing methods struggle with limited dMRI data common in pediatric neuroimaging.

Approach:

• A novel deep learning model is proposed to generate accurate FODs from a significantly reduced set of dMRI measurements (as few as six).
• The model is trained using high-fidelity FODs derived from multi-shell high angular resolution diffusion imaging (HARDI) data.
• Generalizability is assessed across different scanners, protocols, and anatomical variations using clinical datasets from newborns and fetuses.

Key Points:

• The deep learning method achieves comparable or superior results to traditional techniques like Constrained Spherical Deconvolution with substantially fewer dMRI acquisitions.
• Validation includes agreement metrics within a HARDI newborn dataset and comparison of fetal FODs with post-mortem histological data.
• The study demonstrates the efficacy of deep learning for inferring developing brain microstructure from limited in-vivo dMRI data.

Conclusions:

• Deep learning offers a powerful solution for overcoming dMRI data limitations in pediatric neuroimaging, particularly for studying the developing brain.
• Despite advancements, inherent limitations of dMRI persist in analyzing early brain development.
• Further development of specialized methods is recommended to enhance the study of early human brain development using neuroimaging techniques.