Inter-Scanner Harmonization of High Angular Resolution DW-MRI using Null Space Deep Learning

Area of Science:

• Neuroimaging
• Biophysics
• Machine Learning

Background:

• Diffusion-weighted magnetic resonance imaging (DW-MRI) enables non-invasive visualization of brain's microstructural architecture.
• Current methods for reconstructing local fiber orientations from DW-MRI data, such as constrained spherical deconvolution (CSD), often lack reproducibility across different MRI scanners.
• There is a need for more robust and reproducible techniques for accurate brain fiber analysis.

Purpose of the Study:

• To introduce a novel data-driven technique, the null space deep network (NSDN), for improved local fiber reconstruction from DW-MRI data.
• To enhance the accuracy, reproducibility, and generalizability of fiber orientation estimation in the brain.
• To address the limitations of existing methods in terms of cross-scanner consistency.

Main Methods:

• Developed a new neural network architecture, the null space deep network (NSDN), designed to learn from both ground-truth data (ex-vivo DW-MRI and histology) and repeated scan data.
• Trained the NSDN using ex-vivo DW-MRI and histology from squirrel monkey brains, augmented with repeated human scan data from two different MRI scanners.
• Validated the NSDN's performance on a held-out set of histology voxels, comparing its accuracy and reproducibility against traditional methods like CSD and other deep learning approaches.

Main Results:

• The NSDN demonstrated significant improvements in absolute performance compared to CSD (3.87%) and a recent deep learning method (1.42%) when evaluated against histology.
• Reproducibility was substantially enhanced by the NSDN, outperforming CSD by 21.19% and a recent deep learning approach by 10.09% on paired scan data.
• The NSDN showed improved generalizability to an unseen in vivo human scanner, outperforming CSD by 16.08% and a recent deep learning approach by 10.41%.

Conclusions:

• Data-driven approaches, exemplified by the NSDN, offer a more reproducible, informative, and precise method for local fiber reconstruction in the brain.
• The proposed NSDN provides a novel and practical solution for improving the reliability of DW-MRI-based fiber tractography.
• This study highlights the potential of advanced machine learning techniques to overcome limitations in neuroimaging reproducibility and accuracy.