Parallel Imaging Meeting 2005

1
Parallel Imaging Meeting 2005

• Esin Ozturk
• UCSF/UCB Joing Graduate Group in Bioengineering

2
Parallel MRI

• Higher speed in MRI is desirable due to cost,
  time limitations, and the motion artifacts.
• It is possible to reduce the scan time by faster
  phase encoding.
• Gradient switching rate has a limit.
• Neuromuscular stimulation might be a problem.
• Acquiring multiple MR data points simultaneously
  can result in the possibility of reducing the
  acquired data, and retrieving it in the
  reconstruction phase using parallel imaging
  techniques.

3
Main Idea

• Multiple receiver coils to acquire the MR signal
  simultaneously.
• Reduced phase encode steps acquired from each
  coil to reduce the data acquisition time by
  keeping kmax same and increasing ?k by a factor
  of 2.
• Signal induced in an RF coil varies by the
  distance between the signal source and the coil.
  ? Coil Sensitivity
• Special image reconstruction techniques employing
  coil sensitivities methods like SENSE, GRAPPA
  etc.

4
Overview of Steps
DATA ACQ/PREPARATION
Simulation - Image - Spectra
DATA RECONSTRUCTION
SENSE GRAPPA AUTO-SMASH VD-AUTO-SMASH
5
Sensitivity Encoding (SENSE)
6
Principles of SENSE

• A set of receiver coils are used to acquire
  undersampled k-space data.
• These data are inverse Fourier transformed
  directly which results in half FOV images/spectra
  in the reduction direction for R2.
• Each pixel in an aliased image is a superposition
  of multiple pixels from a full unaliased image.
• Knowing the coil sensitivities of all the pixels
  contributing to the pixel, and the composite end
  value, it is possible to solve a linear equation
  for each pixel.

7
1. Combining multi-channel data R1
x

Coil sensitivities at voxel v
Data of each coil element at voxel v
8
2. Combining with SENSE for R2
x

Coil sensitivities at voxel v
Data of each coil element at voxel v
9
Simulations Original Image
Modified Shepp-Logan phantom (256256)
10
Simulations Coil Sensitivity Maps

• Coil sensitivities are simulated as Gaussian
  functions with FWHM5cm
• and max height 1 at a FOV 15cm
• Sensitivity

11
Intermediate SENSE Images

• Phase encoding step size 2
• ?k ? ?k 2
• ? FOV1/dk ? FOV / 2
• Direct 2D IFT results in reduced FOV aliased
  images.

Simulated intermediate images from 3 seperate
coils
12
Reconstructed Image
13
Empirical Coil Sensitivity Maps

• For the calibration image of a given coil
  element
• Divide by the masked combined image
• Median filtering low-pass homomorphic
  filtering
• Dilate by a 3x3 kernel to extend the image for
  edge preservation.
• Division by the body coil images or theorotical
  maps ?

14
SENSE Spectroscopy

• Spectroscopic data can be viewed as a 4
  dimensional data, where three of the dimensions
  are spatial, and the fourth is the spectral
  dimension.
• After 1D FFT of the spectral, and 3D IFFT of the
  spatial domains, we can apply SENSE
  reconstruction for each voxel at each spectral
  point on a slice by slice basis.

15
Aliasing of spectra

• SIMULATION
• 3232 Full FOV
• 1616 1/4th FOV
• Fourier reconstruction applied
• Peaks of A, B, C, D
• A is the superimposed pixel.

Figure 1 in Dydak et al.s paper
16
SENSE reconstruction for spectra

• Phantom filled with Cre of 10 mmol/l, Glass
  spheres NAA 10 mmol/l, Cre 20 mmol/l, Lac 10
  mmol/l
• Creatine map with full FOV
• Aliased creatine map acquired with half FOV,
  voxel A is a weighted sum of the 4 marked voxels.
• SENSE reconstructed

Figure 4 of Dydak et al.s paper
17
Spectral Simulation
phase encode direction
18
Lipid Unaliasing using SENSE

• For a given voxel within the FOV
• Possible aliasing voxels (FOV away from the
  voxel) were determined.
• Voxels within the head region were selected.
• Original spectral array (FOVxFOV) was treated as
  a half FOV representation of an extended spectral
  array (2FOVx2FOV).
• Using
• 8 sets of coil sensitivities, and
• 8 spectral values of the given voxel from the
  different channels
• the spectra in the original and the aliasing
  voxels were resolved.

19
Example Grade 4 Glioma Patient
20
Problems of SENSE

• SENSE requires the inversion of matrices that
  might result in computational errors.
• SENSE is computationally expensive because
  unfolding is performed for each pixel.
• Noise can cause reconstruction errors.
• Reduction factor (R) can not exceed the number of
  coils.
• SNR is reduced by a factor of at least ? R