Non-parametric prediction of brain MRI microstructure using transfer learning.

Area of Science:

• Neuroimaging
• Biophysics
• Machine Learning

Background:

• Magnetic Resonance Imaging (MRI) is sensitive to tissue microstructure.
• Traditional biophysical models use simplified representations of brain tissue.
• Machine learning (ML) offers a data-driven approach for microstructural feature extraction.

Purpose of the Study:

• To develop a novel ML approach for reliable brain tissue microstructure estimation using MRI.
• To minimize the requirement for paired MRI-histology datasets in ML model training.
• To predict non-parameterized joint distributions of g-ratio and axon diameters from MRI data.

Main Methods:

• Pretraining a conditional normalizing flow model on synthetic MRI data generated from unpaired histology and MRI physics.
• Utilizing transfer learning to fine-tune the model with experimental MRI/Electron Microscopy (EM) data.
• Generating synthetic MRI data through segmentation, feature extraction, and MRI simulators.

Main Results:

• Achieved close agreement between predicted and EM ground-truth histograms for axon diameter and g-ratios.
• Demonstrated up to 4% decreased mean percent errors compared to biophysical model fitting.
• Showed significant differences in g-ratio predictions in mice after myelin remodeling seizures.

Conclusions:

• Pretraining on synthetic MRI data combined with transfer learning effectively addresses the scarcity of paired MRI/histology data.
• This approach enables accurate prediction of microstructural features, advancing the development of foundation models for MRI-based microstructure analysis.
• The method shows promise for applications in neuroscience research, including the study of neurological disorders.