Download as pdf or txt
Download as pdf or txt
You are on page 1of 2

117      Image Interpolation (Zero Filling)

For discussion purposes, consider an image display monitor with a 512 3 512 pixel
resolution. The image acquisition matrix in MRI is often lower, and some interpola-
tion is typically performed automatically. A straightforward approach is to place an
empty pixel between pixels (step 1 in Fig. 117.1) and assign the amplitude of this pixel
to be the average value of the adjacent pixels (linear interpolation, step 2 in Fig. 117.1).

Fig. 117.1

An advanced solution is the use of a bicubic spline interpolation. In this case, the value
of not only the adjacent pixel is taken into account, but also that of the next pixel. The
result is smoother, masking the truly measured spatial resolution. For Fourier ­encoding,
a low spatial resolution ma-
trix ­corresponds to a small
k space (limited dimension
in the readout and phase
­encoding directions). In any k
space matrix, the outer data
points (the high spatial fre-
quency components) contain
the information concerning
the detailed structure in the
image. Consider a homoge-
neous phantom that fills the
entire field of view. There
is no detailed structure, re-
sulting in zero values for
the outer data points in the
k space matrix. Reflecting
upon this ­situation, doubling
the raw data matrix size and
filling the missing (outer)
data points with a value of
zero will not improve spatial
resolution, but will mimic
in some ways the measure-
ment of a higher ­resolution
Fig. 117.2 ­matrix. Zero ­filling is simply
246

Runge_CH117.indd 246 9/29/13 11:43 AM


247

the substitution of zeroes for unmea-


sured data points for the purpose of in-
creasing the matrix size of the data prior
to Fourier transformation. The result
is an image with an increased display
­matrix that mathematically has pixels
that are ­sinc-interpolated. Fig.  117.2
illustrates Fourier interpolation as ap-
plied in MRI. In step 1, the k space data
and ­corresponding image for a 128 3
128 acquisition matrix is illustrated. In
step  2, zero filling is used to expand k
space. The resulting interpolated 256 3
a 256 image is compared with a true 256
3 256 acquisition in step 3. In Fig. 117.3,
a T2-weighted axial image of the brain
at the level of the pons acquired using a
256 3 256 matrix (Fig. 117.3a) is com-
pared with the same image interpo-
lated to 512 3 512 using zero filling in
k space (Fig. 117.3b) and a true 512 3
512  ­acquisition (Fig. 117.3c).
So what then is the advantage of
zero filling in k space over interpola-
tion in the image domain? In the case
of small structures, for example, a small
vessel in a time-of-flight MRA of a size
close to the spatial resolution of the
b
measurement, the intensity of the rep-
resenting pixel depends on where the
vessel is located relative to the measure-
ment grid. If the vessel is within a single
voxel, the signal will be very bright.
If  the vessel lies between two  voxels,
these two voxels will share the inten-
sity, resulting in a less dominant vessel
appearance. Theoretically, the appear-
ance will be improved by shifting the
measurement grid, which is the same
as shifting the position of the voxel,
eliminating the above-mentioned par-
tial ­volume effect. Zero filling in k space
c corresponds mathematically to a voxel
Fig. 117.3 shift. In other words, spatial resolution
is not improved, but partial volume
­artifacts are significantly reduced.

Runge_CH117.indd 247 9/29/13 11:43 AM

You might also like