Gauge equivariant convolutional neural networks for diffusion MRI.

Area of Science:

• Neuroimaging
• Medical Imaging
• Computational Neuroscience

Background:

• Diffusion MRI (dMRI) provides directional information on neuronal fibres via water molecule diffusivity.
• Acquiring sufficient dMRI images for reliable angular resolution requires long scan times and high costs.
• Current methods face challenges due to the spherical and non-euclidean nature of dMRI data.

Purpose of the Study:

• To introduce gauge equivariant convolutional neural network (gCNN) layers for dMRI analysis.
• To address the challenges of processing spherical dMRI data with identified antipodal points.
• To improve angular resolution and enable prediction of diffusion tensor imaging (DTI) parameters with fewer gradient directions.

Main Methods:

• Development of gCNN layers specifically designed for the real projective plane manifold, a non-euclidean domain for dMRI.
• Application of gCNNs to upsample angular resolution in dMRI data.
• Prediction of DTI parameters using a model trained on limited diffusion gradient directions (six directions).

Main Results:

• The developed gCNNs effectively handle the unique geometric properties of dMRI data.
• The method allows for accurate prediction of DTI parameters from a reduced number of diffusion gradient directions.
• gCNNs demonstrated the ability to train with fewer subjects compared to standard 3D convolutional models.

Conclusions:

• gCNNs offer a novel approach to overcome limitations in dMRI data acquisition.
• This technique can significantly reduce scan times, costs, and barriers to clinical adoption of dMRI.
• The use of gCNNs in dMRI facilitates more efficient neuroimaging research and clinical applications.