Ground-truth effects in learning-based fiber orientation distribution estimation in neonatal brains.

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Computational Neuroscience

Background:

• Diffusion Magnetic Resonance Imaging (dMRI) noninvasively visualizes brain microstructure.
• Fiber Orientation Distributions (FODs) map white matter tracts, with deep learning showing recent success in estimation.
• Current deep learning models for neonatal FOD estimation are often trained on multi-shell multi-tissue constrained spherical deconvolution (MSMT-CSD) data, which may not be optimal for developing brains.

Purpose of the Study:

• To investigate the suitability of single-shell three-tissue constrained spherical deconvolution (SS3T-CSD) versus MSMT-CSD for neonatal brain microstructure analysis.
• To evaluate the impact of input gradient directions on FOD estimation accuracy in neonates.
• To assess the robustness of SS3T-CSD across different age groups in an age domain-shift scenario.

Main Methods:

• Trained a U-Net based deep neural network model using both SS3T-CSD and MSMT-CSD as ground truth.
• Compared the performance of the trained models using metrics relevant to neonatal brain development.
• Evaluated model performance under varying numbers of input gradient directions and in an age domain-shift setting.

Main Results:

• SS3T-CSD yielded a more realistic ratio of single to multiple fiber-estimated voxels compared to MSMT-CSD for neonatal brains.
• Increasing input gradient directions significantly enhanced SS3T-CSD performance more than MSMT-CSD.
• SS3T-CSD demonstrated robust performance across different age groups, indicating stability in age domain-shift conditions.

Conclusions:

• SS3T-CSD may be a more appropriate ground truth for neonatal brain FOD estimation than MSMT-CSD.
• The performance of SS3T-CSD is sensitive to the number of diffusion gradient directions.
• SS3T-CSD shows potential for more accurate and reliable dMRI-based neonatal brain imaging across various ages.