ODF based deep learning network for unsupervised deformable diffusion resonance image registration (ODDRnet).

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Deep Learning

Background:

• Diffusion MRI (dMRI) is crucial for mapping brain white matter architecture.
• Current registration methods often ignore dMRI's high-dimensional directional information, impacting white matter alignment accuracy.
• Complex fiber crossings in dMRI data pose challenges for precise anatomical correspondence.

Purpose of the Study:

• To introduce ODDRnet, an unsupervised deep learning framework for nonlinear registration of dMRI data.
• To leverage directional information in dMRI by aligning Fiber Orientation Distribution Functions (fODFs).
• To improve the accuracy of white matter pathway alignment in dMRI.

Main Methods:

• ODDRnet directly aligns high-dimensional fODFs derived from raw dMRI signals.
• The framework predicts a dense deformation field to spatially warp fODFs and reorient directional information.
• An end-to-end, unsupervised deep learning approach is employed for dMRI registration.

Main Results:

• ODDRnet achieved superior macroscopic accuracy in dMRI registration.
• Demonstrated a mean increase of 0.02 in tract dice and a reduction of 0.31 mm in tract distance.
• Showcased robust generalization across diverse datasets, ethnicities, age groups, health conditions, and imaging protocols.

Conclusions:

• ODDRnet offers a significant advancement in dMRI registration by effectively utilizing directional information.
• The framework ensures accurate alignment of white matter structures, improving subsequent analyses like tractography.
• ODDRnet's adaptability makes it suitable for a wide range of dMRI registration tasks.