A Multi-Contrast Translation-Based Registration Approach for Distortion Correction in DTI.

Correcting eddy currents and motion artifacts is crucial in Diffusion Tensor Imaging (DTI) preprocessing, traditionally managed through affine registration to an undistorted reference. However, the contrast variation across diffusion-weighted images complicates direct registration. To surmount this challenge, our study introduces a translation-based registration approach, utilizing 312 DTI datasets from the Human Connectome Project (HCP), including both b=0 and b=2000 volumes. We employed an advanced 3D Self-Attention Conditional Generative Adversarial Network (SC-GAN) for the synthesis of imaging data. This method allowed for the generation of synthesized b=2000 volumes, enhancing the distortion-correction process by facilitating efficient registration of distorted images. The results showed the translation network's effectiveness in synthesizing b=2000 volumes from real data, with these volumes serving as stable registration targets, particularly in limited directional data scenarios. The approach also effectively corrected eddy current and motion artifacts, aligning FA and FOD maps with gold standard results from FSL's eddy method, confirming the translation-based registration's efficacy in addressing cross-contrast registration challenges in DTI preprocessing.