David Atkinson1, Muhammad Usman1, Lebina Kakkar1, and Simon Arridge1
1University College London, London, United Kingdom
Synopsis
Keeping
track of positional information, from planning the position of a field of view,
through to the final reconstructed image can be challenging. This educational
poster highlights some of the issues and pitfalls in the acquisition and
reconstruction process. Use of DICOM-style spatial referencing is encouraged
and strategies for validation are presented.
Introduction
Maintaining
correct geometrical information during reconstruction is important for surgery and
radiotherapy applications, for consistent alignment when using multiple images,
e.g. multi-parametric MRI, PET-MR, B0
distortion correction, for clarity in standards e.g. ISMRMRD1 and
when comparing new reconstructions to clinical DICOMs. Correct handling can be challenging
due to: conventions used in the Fast Fourier Transform, radiological image orientations,
a requirement for square pixels and images and the ‘centre point’ falling on a
voxel boundary for even-length arrays. We provide some clarity, guidance and
suggestions for validation.DICOM
Figure 1
shows the DICOM convention for spatially referencing an image using the ImagePositionPatient (IPP) vector to a point, and ImageOrientationPatient (IOP) vectors along a row and column. This point and two vectors convention
allows
for easy manipulations of image data following the Fourier Transform, e.g. to display in a radiological orientation.
In
multi-slice 2D imaging, the temporal order of slice acquisition can vary. When
assembling a volume, slice positional order can be determined from the scalar
product of the IPP vector with the slice normal (formed
from the cross product of the row and column vectors).
Scanning Process
The initial
laser positioning sets a point in the body and after planning, table movements
mean this point may not coincide with the magnet isocentre (Figure 2). The
position of the excited slice (2D) or centre of a slab (3D) is set by the RF
excitation frequency and the slice select gradient which can have any angulation,
achieved by combining physical X,Y and Z gradients. An entire slice/slab is
excited, but filtering the received frequencies restricts the spatial extent in
the frequency encode direction (M), and by changing the centre frequency, a
desired offset in M is achieved. In the phase encode (PE) direction, linearly changing
the acquired phase provides a PE offset of the image after Fourier Transform. These
offsets may differ from the user-specified offset if a) a half-pixel
shift is added in acquisition (to move the image ‘centre’ off a voxel
boundary), or b) in-plane offcentres are not applied in acquisition, but achieved
by shifting in post-processing, e.g. in the case of readout ramp sampling. The
k-space origin is the point acquired when the net gradient area was zero. Zero-padding
of k-space may be applied to adjust the reconstructed voxel size, typically to
make it equal in M and PE directions (Figure 3). Shim fields may be defined in
terms of X, Y and Z directions.Fourier Transform
K-space to
image space conversion uses the forward
Fourier Transform ( –ve sign in the exponential). By convention, the FAST FT
algorithm expects both input and output data to be half-shifted so that the origin
is at the start of the array (not the centre). Shifting and inverse shifting
operations are not identical for odd-sized arrays so should be applied in the
correct order. In MATLAB the operation is img = fftshift( fft( ifftshift(
kspace ))). Using 1-based indexing and after shifting, the origin is at array
index: ceil((N+1)/2), see Figure 4.
Following
FFT, any required image cropping, zero-padding, reflection or 90 degree rotations do not
change the voxel size and spatial referencing is easily updated using the
DICOM-style point and two vectors representation. For example, following a 90
degree counter-clockwise rotation, the new row vector is the old column vector,
the new column vector is the negative of the old row vector and the new top
left pixel is the old bottom left (whose 3D coordinates are easily calculated before the rotation using IPP and IOP vectors and voxel sizes).
Validation
Geometrical
handling can be checked by comparing reconstructions from scans acquired at
multiple offsets, angulations, fat shift and phase encode directions. For test data,
use spin echo sequences with high bandwidth (to reduce distortions), a power of
2 FOV, e.g. 256mm with 2mm voxel size and 128 acquisition and reconstruction
matrices (if zero-padding is to be avoided), image near isocentre (to reduce
the effects of gradient non-linearities) and ignore alignment differences near
to susceptibility boundaries. Small structures within a phantom are useful. Comparison
requires reslicing one reconstruction to the geometry of another reconstruction
chosen as a reference, possibly from an alternative algorithm, e.g. a scanner
DICOM. Residual distortion and intensity changes mean subtraction images need
visual interpretation. Comparison with errors from an intentional sub-pixel
offset can help to guide interpretation. Interpretation of X, Y and Z
directions can be tested by acquiring a field map with zero shim and comparing
to one after shims specified in X, Y, Z have been subtracted (Figure 5). Acknowledgements
Cancer Research UK A21099, UK EPSRC EP/P022200/1, EP/M022587/1. References
1. Inati et al MRM 77 p411 (2017). 10.1002/mrm.26089