Geometric Deep Learning for Unsupervised Registration of Diffusion Magnetic Resonance Images.

Area of Science:

• Medical Imaging
• Neuroimaging
• Computer Vision

Background:

• Deep learning models excel at medical image registration, offering speed and accuracy for scalar modalities like MRI and CT.
• Existing deep learning methods have limitations in handling complex diffusion MRI data.
• Non-rigid registration of diffusion MRI is crucial for analyzing white matter architecture.

Purpose of the Study:

• To present the first end-to-end geometric deep learning model for non-rigid registration of diffusion MRI-derived fiber orientation distribution fields (fODF).
• To enable fully-unsupervised training using only input fODF image pairs.
• To develop accurate and computationally efficient registration algorithms for diffusion MRI.

Main Methods:

• Developed a novel end-to-end geometric deep learning model for fODF registration.
• Introduced differentiable layers for local Jacobian estimation and reorientation.
• Integrated these layers into a manifold-valued convolutional network architecture.
• Enabled fully-unsupervised training without ground truth deformation fields.

Main Results:

• Achieved accurate deformable registration of diffusion MRI data.
• Demonstrated the model's ability to handle fiber orientation distribution fields (fODF).
• Reduced registration time from hours to seconds compared to classical methods.

Conclusions:

• The proposed deep learning model offers a significant advancement in diffusion MRI registration.
• This method provides accurate and rapid non-rigid registration for complex dMRI data.
• The approach paves the way for faster and more precise analysis of brain white matter structure.