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

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Eddy current and motion artifacts are significant challenges in Diffusion Tensor Imaging (DTI) preprocessing.
• Traditional affine registration methods struggle with contrast variations across diffusion-weighted images.
• Accurate registration is essential for reliable DTI analysis.

Purpose of the Study:

• To develop and validate a novel translation-based registration approach for DTI preprocessing.
• To address the challenge of cross-contrast registration in diffusion-weighted images.
• To improve the correction of eddy current and motion artifacts in DTI data.

Main Methods:

• Utilized 312 DTI datasets from the Human Connectome Project (HCP).
• Employed a 3D Self-Attention Conditional Generative Adversarial Network (SC-GAN) for synthesizing b=2000 volumes.
• Implemented a translation-based registration strategy for distorted DTI images.

Main Results:

• The SC-GAN effectively synthesized stable b=2000 registration targets from real DTI data.
• The translation-based registration successfully corrected eddy current and motion artifacts.
• Fractional Anisotropy (FA) and Fiber Orientation Distribution (FOD) maps aligned with gold standard results.

Conclusions:

• The proposed translation-based registration is effective for DTI preprocessing, especially in challenging cross-contrast scenarios.
• This method enhances artifact correction and improves registration accuracy in DTI.
• The approach shows promise for improving the reliability of DTI analysis, particularly with limited directional data.