DWIDTI QC Martin Styner UNC Thanks to Guido Gerig NYU And many many folks Content Todays topic Diffusion MRI QC Still underappreciated Artifact correction with FSL Motion amp eddy current ID: 1048496
Download Presentation The PPT/PDF document "Diffusion MRI - Lesson 3" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1. Diffusion MRI - Lesson 3DWI/DTI QCMartin Styner UNCThanks to Guido Gerig, NYUAnd many, many folks
2. ContentToday’s topic: Diffusion MRI QCStill under-appreciatedArtifact correction with FSL:Motion & eddy currentSusceptibility artifactsNRRD – NIFTI conversionRemove bad DWI data with DTIPrepInvestigative tractography in Slicer
3. DWI ArtifactsDiffusion images are sensitive to a number of artifactsEddy-current distortionsSusceptibility artifactsNoise/SNR issuesVibrational artifactsMotionVenetian blind artifacts“unknown”…DTIPrep: Remove bad DWI’s
4. Bad DWI: Slicewise ArtifactBad DWI detection => 3 choicesRemove bad DWI data (DTIPrep)Simple and clean, but removes partially ok dataDWI model correctionRESTORE: DTI model, during estimationHARDIPrep (NYU/Gerig): Correction of corrupt slices (identified by DTIPrep) via spherical harmonic model FSL: Correction of corrupt drop-out slices (different identification) via Gaussian process modelBoth: Need motion & eddy current correction done Don’t care…Estimated data significantly biased => incorrect resultsBerl et al (2015). Investigation of vibration-induced artifact in clinical diffusion-weighted imaging of pediatric subjects. Hum Brain Mapp, 36(12), 4745–57.
5. Within DWI Motion ArtifactVenetian blind artifactsDTIPrep: Detects and removes DWI with motionDifferences between interleaved volumes FSL: Detects motion between slices and correct them (only new version 5.12 or 6.1, not 6.0)Slice motion model as part of eddy current/motion correction, performs slice to volume registrationVery slowHigh error setting for registrationI would not use (at this time)
6. Modern dMRI QC WorkflowDICOM conversion(better) Remove bad DWIs via DTIPrepSusceptibility artifact computation FSLConcatenation of corrected AP & PABrainmaskMotion & Eddy current correctionAlso applies susceptibility un-deformationAlternatively correct bad DWIs(simpler) Remove bad DWIs via DTIPrepDTI estimation – tractography QC
7. DICOM convertionDiffusion tags in DICOM are a mess!No field permeating standard yetNo single great converter toolCurrently Best: DWIConvert (Slicer)Captures many vendors and special cases wellCan write both NRRD and NIFTI imagesCan convert to/from NRRD vs NIFTIUses B-matrix for actual gradient and b-value (option)Alternative: dcm2nixConverts to NIFTI, uses prescribed information
8. NIFTI vs NRRD DWINIFTI4D – Image without diffusion informationGradients (b-vectors) & b-values in separate text filesConfusing: often bvec & bval files not name linkedNo measurement frameb-vectors need to be specified in image coordinate spaceVendors use different spaces (patient vs scanner coordinate space) => Adaption of bvecs needed
9. NIFTI vs NRRD DWINRRDSingle file format (.nrrd) or raw+header (.raw +.nhdr)All image information in header including diffusion dataNo confusionsMeasurement frame infoInconvenient: b-vector and b-value info combined into single value
10. Susceptibility CorrectionCorrection of susceptibility artifactsGeometric distortion due to air/tissue interface or other sources (e.g. metal plates)Anatomy/Object induced distortionAt first approximation: constant for all DWIAlternative: twice refocused sequence approach APPACorr
11. Susceptibility CorrectionNeeds two acquisitions with opposing polarities of the phase-encode direction AP vs PA, or LR vs RLAlternatively use structural T2w image or field mapFSL TopupComputes geometric un-distortion fieldWorks for both opposing polarity or with structural T2wTime-consumingApplication best with joint motion corr (only new FSL)Not crucial for “Quick” QCOnly used for minimal processing, not at QC”Quick” QC separate for AP and PA
12. TopupBest: only use b=0 images (high SNR)topup --imain=<B0s> --out=<fieldcoef> --fout=<field> --iout=<B0corr> --datain=<topupParam> --config=<fslregParam> Needs 2 parameter filesfslregParam: describes registration settings, defaults from FSL work well for most datatopupParam: lists phase encoding & readout timeReadout time is commonly consistent => bogus value okAbsolute sign irrelevant, relative sign is importantE.g. 2 B0 AP & PA (orientation RAS) :0 1 0 0.10 1 0 0.10 -1 0 0.10 -1 0 0.1
13. TopupTopup computes the un-deformation fieldIncorporate b0 image motion correctionapplytopup (old style) to apply to non b0Does NOT correct for motion, assumes no motionAvoidUse eddy_openmp (new style)Jointly estimates motion & eddy currents, as well as applies the susceptibility correctionCompute on full concatenated dataVery slow (use longleaf)
14. Brain Masking in DWIOptions depending on availabilityPropagating sMRI brain mask to geometry corrected average b0 (affine reg)FSL bet to IDWI imageGeometric average of non-b0 imagesContains little non-brain tissueMost will need manual correction QuickCheck on FA for non-inclusion of skullI suggest to use ITK-Snap
15. DWI Motion CorrectionQuantifies and corrects motion between DWIsBoth DTIPrep and FSLReporting of detected/corrected motionAnalysis of motion presence wrt clinical variables necessary nowadaysOld FSL (pre 5.12 and 6.0) Registration to first b0DTIPrepRegistration to unbiased b0 averageEstimate motion better for later acquired DWIsBetter than old FSL (Kreilkamp 2015 JMRI)
16. Eddy Current ArtifactsDifferent for every DWIShearing geometric distortionNo/little artifacts for b0 imagesFirst order approximationCorrect via affine registrationIncorporates motion correctionNearly all tools do it this way
17. New FSL EddyOnly in 5.12/5.13 and 6.1 (and higher)eddy_openmp does everythingEddy current correctionMotion correctionSusceptibility field applicationWithin DWI motion correction (do not use)Within DWI drop out correction (prefer removal)Needs brain mask
18. New FSL Eddyeddy_openmpVery slow (> 1 day on single core)Can use multiple cores (do so)Environment variable: OMP_NUM_THREADS 4bvals, bvecs, topup para, “--data_is_shelled”Needs new configuration file identifying registration target per DWI by index (base 1)Simple: all 1 => old FSL, register to first b0Better: index of last acquired b0Example: 5 b0, equally spaced, 35 total1 1 1 1 1 1 1 2 2 2 2 2 2 2 3 3 3 3 3 3 3 4 4 4 4 4 4 4 5 5 5 5 5 5 5
19. DTIPrepDTIPrep – automatic diffusion QC toolProtocol based QC and correctionImage dimensions checkGradients directions checkIntensity changes across slicesExcessive motion across slicesMotion and eddy current correctionResidual motion detectionVibration/direction DTI artifact detectionGUI and command lineSummary reporting, including motion
20. DWI infoDetailedDWIInfo
21. Gradient info in 3DDisplays gradient scheme on unit sphere(F = File)Check for uniformity
22. Visual DTI QCVisual in 3D Slicer:Remove DWI’s manually with significant residual motionHappens rarelyReconstruct tensor and perform investigative tractographyEnsure appropriate tractography Major tracts, spatially diverseCC genu, CC splenium, CC tapetum, uncinate, fornix, inferior longitudinal, cerebro-spinal, anterior cingulate, (arcuate)
23. ConclusionDiffusion QC is a must Otherwise you get crap dataFSL finally has detection of bad DWI!Hurray!FSL does not remove, but correct bad DWI and within-DWI motionEhhh, that’s not what I wanted…Best: Combo of FSL/HARDIPrep and DTIPrepDWI QC is now very very time consuming
24. Slicer Demo