DeepEddy: high-quality fast eddy current and bulk motion correction using deep learning-based image synthesis and co-registration

Motivation: FSL's "Eddy" function accurately corrects eddy currents and bulk motion in diffusion data but requires 16 diffusion directions or more. Goal(s): Develop a deep learning-based correction method with Eddy-level performance without the diffusion direction sampling requirement. Approach: Our proposed DeepEddy pipeline 1) converts each diffusion-weighted image (DWI) into a b=0 image; 2) nonlinearly co-registers the synthesized and empirical b=0 images; 3) applies derived warp fields to original correspondence DWIs. Results: DeepEddy reduces diffusion volumes variance, improves diffusion metrics, and achieves Eddy-level performance without the diffusion direction sampling requirement. Impact: DeepEddy enables eddy current and bulk motion correction for diffusion data with any number of diffusion directions, showing the promise to benefit clinical applications where scan time is extremely limited.