Gauge equivariant convolutional neural networks for diffusion mri.

Area of Science:

• Neuroimaging
• Medical Imaging Analysis
• Machine Learning in Radiology

Background:

• Diffusion MRI (dMRI) is crucial for neuroimaging, providing directional information on neuronal fibers.
• Current dMRI requires numerous images for reliable angular resolution, leading to long scan times and high costs.
• The spherical data acquisition in dMRI presents challenges for standard Convolutional Neural Networks (CNNs).

Approach:

• Introduced gauge equivariant convolutional neural network (gCNN) layers specifically designed for dMRI data.
• Leveraged the mathematical properties of the real projective plane (ℝ P²) for signal processing on a sphere with identified antipodal points.
• Applied gCNNs to upsample angular resolution, enabling prediction of diffusion tensor imaging (DTI) parameters.

Key Points:

• gCNNs handle the non-Euclidean and non-orientable manifold nature of dMRI data (ℝ P²).
• The method successfully predicts DTI parameters using only six diffusion gradient directions.
• Symmetries inherent in gCNNs allow for training with fewer subjects compared to traditional 3D CNN models.

Conclusions:

• gCNNs offer a novel solution to overcome dMRI's angular resolution limitations.
• This approach reduces scan times and costs, potentially increasing clinical adoption of dMRI.
• The technique demonstrates the potential of specialized neural networks for complex neuroimaging data analysis.