Automated Pipelines for DTI Geometric Distortion Correction

Reducing Artifacts: Best Practices in DTI Geometric Distortion Correction

Diffusion Tensor Imaging (DTI) is highly sensitive to magnetic field inhomogeneities and susceptibility differences, which produce geometric distortions and artifacts that degrade tensor estimates and downstream analyses. This guide summarizes best practices to reduce artifacts from geometric distortion in DTI, covering acquisition strategies, preprocessing steps, quality checks, and recommendations for common pitfalls.

1. Acquisition strategies to minimize distortions

• Use multi-band or segmented EPI where possible: Reduces echo-train length and phase-encoding blurring; segmented EPI trades off motion sensitivity for reduced distortions.
• Shorten echo spacing / echo-train length: Optimize sequence parameters (higher bandwidth, shorter TE) to reduce susceptibility-induced shifts.
• Acquire reverse phase-encoding (blip-up/blip-down) images: Collect at least one pair of b0 images with opposite phase-encoding directions for robust susceptibility distortion correction (e.g., TOPUP).
• Acquire field maps: Gradient echo field maps (dual-echo) provide voxelwise B0 estimates useful for alternate correction methods and for cross-checking TOPUP results.
• Consider SE-EPI for reference images: Spin-echo EPI b0 reduces T2weighting and may produce cleaner references for distortion estimation.
• High-resolution structural (T1/T2) with similar coverage: Obtain undistorted anatomical scans for accurate registration and brain masking.

2. Recommended preprocessing pipeline

1. Denoising: Apply patch-based or PCA denoising (e.g., MRtrix dwidenoise, MP-PCA) to improve SNR before other corrections.
2. Gibbs ringing removal: Use subvoxel-shift methods to remove ringing artifacts that affect diffusion metrics.
3. Motion and eddy-current correction: Use advanced tools (FSL’s eddy with outlier replacement and slice-to-volume correction) that simultaneously correct motion and eddy-current distortions while preserving diffusion gradients.
4. Susceptibility-induced distortion correction: Prefer using reversed phase-encoding b0s with tools like FSL’s TOPUP to estimate the field map and apply correction. If reversed PE data are unavailable, use gradient-echo field maps or boundary-based registration to structural images as alternatives.
5. Apply combined transforms in single interpolation step: Concatenate motion, eddy, and susceptibility warp fields and apply once to minimize blurring from repeated resampling.
6. Bias-field (intensity nonuniformity) correction: Use tools like ANTs N4 on structural images and, if needed, on diffusion data after geometrical corrections.
7. Brain extraction and masking: Generate accurate brain masks from corrected b0 or structural images; refine masks to avoid excluding peripheral white matter.
8. Gradient reorientation: Ensure b-vectors are rotated appropriately with motion transforms to maintain correct diffusion orientations.

3. Tools and software recommendations

• FSL (TOPUP + eddy): Widely used; eddy’s slice-to-volume and outlier replacement are strong for motion-prone datasets.
• MRtrix3: Excellent denoising (dwidenoise), preprocessing wrappers (dwipreproc uses FSL TOPUP/eddy), and tensor fitting tools.
• ANTs: High-quality registration for structural alignment and bias correction (N4).
• SPM / ExploreDTI / TORTOISE: Alternatives for registration-based corrections and advanced modeling; TORTOISE offers robust B0 unwarping and motion correction pipelines.
• QSIPrep / fMRIPrep-style workflows: Provide automated, reproducible pipelines combining best practices and provenance.

4. Quality control (QC) steps

• Visual checks: Compare pre- and post-correction b0s and FA maps; inspect alignment to structural images.
• Quantitative metrics: Track residual distortion by measuring voxel displacement maps, mutual information with structural images, and variance explained in eddy outputs.
• Outlier detection: Use eddy’s outlier reports and QC tools (eddy_quad, eddy_squad, QC modules in QSIPrep) to identify bad slices/volumes.
• Reproducibility checks: Run subset test–retest or assess consistency of tensor metrics across symmetric regions.

5. Practical tips and common pitfalls

• Always collect reverse phase-encoding b0s when possible: Offers the most reliable susceptibility correction.
• Plan acquisition with motion in mind: Use shorter scans, interleaved b-values, and subject comfort measures to reduce motion-induced interactions with distortions.
• Avoid multiple interpolations: Combine transforms to a single resampling to preserve spatial fidelity.
• Check b-vector handling: Incorrect reorientation is a common source of erroneous orientations and spurious metrics.
• Be cautious with aggressive denoising or smoothing: Can bias diffusion metrics; apply conservative parameters and verify effects.
• Document preprocessing choices: For reproducibility and interpretation, record software versions, parameters, and QC results.

6. When distortions persist

• Try alternative correction strategies: Compare TOPUP vs. field-map-based correction; evaluate registration-based unwarping to structural images.
• Use advanced modeling: Robust tensor or multi-compartment models (e.g., constrained spherical deconvolution) can mitigate some artifact impact on downstream analyses.
• Exclude severely corrupted volumes: If certain diffusion volumes remain unusable, remove them and account for missing directions during tensor fitting or use robust fitting algorithms.

7. Summary recommendations

• Acquisition: collect reversed phase-encoding b0s and a high-quality structural; minimize echo-train length.
• Preprocessing: denoise → remove Gibbs ringing → motion/eddy → TOPUP (or field map) → single-step resampling → bias correction → brain mask → rotate b-vectors.
• QC: use visual and quantitative checks; track eddy/TOPUP diagnostics.
• Tools: prefer FSL (TOPUP+eddy), MRtrix3 denoising, and QSIPrep for automated reproducible workflows.

Following these practices will substantially reduce geometric distortions and related artifacts in DTI, improving tensor estimation and the validity of tractography and quantitative diffusion analyses.