Imaging the brain at ultra-high resolution using 3D FatNavs

Transcription

1 Imaging the brain at ultra-high resolution using 3D FatNavs Daniel Gallichan Centre d Imagerie BioMédicale EPFL, Lausanne, Switzerland

2 Overview Introduction How motion affects MRI scans Ways we can track motion Methods to correct for motion Our approach: 3D FatNavs Results

3 Motion in MRI If the object moves during the scan the image will be affected! The effect is different to motion blur in photography Head-motion is a simpler problem as can be considered rigid-body Photograph MRI Zaitsev et al, NeuroImage 2006

4 The kind of motion is important for MRI No motion Photograph MRI Large, smooth motion

5 The kind of motion is important for MRI No motion Photograph MRI Small, rough motion

6 So how do real subjects move? MRI Zaitsev et al, NeuroImage 2006 Not the most realistic motion Most motion-correction studies ask volunteers to make large deliberate motion

7 So how do real subjects move? Preliminary data from 31 subjects

8 The kind of image is also important for MRI No motion MRI No motion MRI Small, rough motion

9 The kind of image is also important for MRI

10 The kind of image is also important for MRI Similar motion during two scans Without Fat Suppression Howarth et al, NeuroImage 2006 With Fat Suppression

11 Clearly, motion is an issue for MRI So how can it be corrected for?

12 Motion-correction in MRI For 3D structural neuroimaging Measurement of motion: MR Navigators vs External markers Welch et al, MRM 2001 Tisdall et al, MRM 2012 Zaitsev et al, MRM 2006 Maclaren et al, PLOS ONE 2012 Application of correction: Prospective vs Retrospective Scanner coordinates updated during scan Correct 3D k-space offline after the acquisition

13 Example of high-resolution motion-correction 0.6 mm isotropic MP-RAGE at 7T, from Maclaren et al, PloS One (2012) MoCo off MoCo on

14 Example of high-resolution motion-correction 0.6 mm isotropic MP-RAGE at 7T, from Maclaren et al, PloS One (2012) - Requires line-of-sight between camera and marker - Existing MR navigator-based methods yet to be demonstrated at this resolution take too long to acquire - Our aim: a high-resolution navigator as quickly as possible, with minimal influence on host sequence MoCo on

15 A fat-based MR navigator for motion-correction Water Image (32 s GRE at 2x2x2mm) Advantages of a FatNav - Minimal impact to spin history of host sequence 1,2 - Inherently sparse image suitable for very high acceleration 3,4 Fat Image (32 s GRE at 2x2x2mm) 1. Ari and Kraft, Proc ISMRM van der Kouwe et al, Proc ISMRM Gallichan et al, Proc ISMRM Skare et al, MRM 2015

16 Natural sparsity of fat images allows GRAPPA at very high acceleration factors Gallichan et al, proc ISMRM 2014 Water image R = 2x2 = 4 R = 6x6 = 36 R = 10x10 = 100 R = 14x14 = 196 R = 18x18 = 324 Fat image

17 Implementation All data from 7T head-only system (Siemens) with 32ch RF receive array (Nova Medical Inc.) MP2RAGE T 1 -weighted TSE T 2 -weighted GRE T 2 *-weighted

18 Tracking of motion with FatNavs Small deliberate motion

19 Retrospective correction of k-space Translations can be accounted for by simple multiplication of each k-space plane by a phase ramp Rotations require rotation of the k-space planes

20 Retrospective correction of k-space Need to interpolate back onto Cartesian grid before FFT Implemented using NUFFT (from Jeffrey Fessler s Matlab toolbox)

21 Example: MP2RAGE 330 µm x 330 µm x 1 mm, acquired in 37 mins Gallichan et al, MRM 2016 RMS RMS motion: motion: mm, mm, degrees degrees

22 Example: MP2RAGE 330 µm x 330 µm x 1 mm, acquired in 37 mins Zoom from same dataset as previous slide Improvement subtle, but could be critical for quantitative measurements (e.g. cortical thickness) or identifying very small features No MoCo Gallichan et al, MRM 2016

23 Example: MP2RAGE 330 µm x 330 µm x 1 mm, acquired in 37 mins Zoom from same dataset as previous slide Improvement subtle, but could be critical for quantitative measurements (e.g. cortical thickness) or identifying very small features With MoCo Gallichan et al, MRM 2016

24 Example: MP2RAGE 330 µm x 330 µm x 1 mm, acquired in 37 mins Vessel contrasts also available from the same MP2RAGE dataset No Moco Minimum Intensity Projection over 15 mm slab of INV 2 contrast Maximum Intensity Projection over 80 mm slab of INV 2 contrast

25 Example: MP2RAGE 330 µm x 330 µm x 1 mm, acquired in 37 mins Vessel contrasts also available from the same MP2RAGE dataset With Moco Minimum Intensity Projection over 15 mm slab of INV 2 contrast Maximum Intensity Projection over 80 mm slab of INV 2 contrast

26 So how do real subjects move? Preliminary data from 31 subjects

27 Usefulness in standard cases Subject who moved the least No MoCo

28 Usefulness in standard cases Subject who moved the least With MoCo

29 Usefulness in standard cases Subject who moved the most No MoCo

30 Usefulness in standard cases Subject who moved the most With MoCo

31 Usefulness in standard cases Small deliberate motion No MoCo

32 Usefulness in standard cases Small deliberate motion With MoCo

33 Usefulness in standard cases Small deliberate motion Correction not complete May be improved by: Fixing motion-corrupted GRAPPA calibration data Iterative NUFFT to account for k- space gaps and overlaps Autofocusing (Atkinson 1997, Loktyushin 2013)

34 Latest work: motion-correction enabled ultra-high resolution imaging

35

36

37 Motion-corrected GRE with 3D FatNavs, 380 μm isotropic resolution

38 The hippocampus From Neurocomic M. Farinella and H. Ros

39 1 mm resolution (1000 nl) 500 µm resolution (125 nl) 380 µm resolution (55 nl) Manually segmented hippocampus

40 1 mm resolution (1000 nl) 500 µm resolution (125 nl) 350 µm resolution (43 nl)

41 Hot off the press Gretsch et al, ISMRM 2016 Integration of dual-echo FatNavs with FIDnavs 1 Sensitive to respiratory and cardiac (!) changes in motion-parameters and B0-field changes [1] Kober et al, MRM 2011 T2*-weighted GRE 250 µm x 250 µm x 1.00 mm Minimum-intensity projection over 9 mm slab

42 Outlook Make more widely available! Distinguish brain regions by cortical contrast on a single-subject basis

43 Outlook Make more widely available! Distinguish brain regions by cortical contrast on a single-subject basis Identify pathologies where ultra-high resolution may be beneficial Cryptogenic epilepsy Hippocampal abnormalities in early Alzheimer s Substantia nigra changes in Parkinson s

44 Outlook Make more widely available! Distinguish brain regions by cortical contrast on a single-subject basis Identify pathologies where ultra-high resolution may be beneficial Analyse many datasets to assess impact of motion-correction Centre d Imagerie BioMédicale