Next: Summary
Up: Title and abstract
Previous: Atlases and Deformable
Models
Part 2 - The VHD as a Source of Test Data
Registration algorithms, whether rigid or non-rigid, require
validation. This presents a problem because no "gold standard"
perfect
registration exists for real data. In this section we will look
at this
problem and examine the possibility of using the VHD to produce
test
data for registration algorithms.
For a registration algorithm to be used clinically it is
important that
its accuracy is established in many conditions. Such validation
is
ideally based on a gold standard registration, i.e. a match where
the
transformation is known to a much higher degree of accuracy than
the
algorithm itself. For real patient data there is no such gold
standard.
One attempt at achieving this is the retrospective registration
study
in which rigid registration methods are compared to that achieved
with
markers implanted into the skull. This study goes a long way
towards
validation of registration algorithms 3.
However, the marker based system, whilst considered to be more
accurate
than the algorithms under scrutiny, is not a true gold standard.
There
are potential inaccuracies due to marker identification and
imager
distortion, though the latter is rectified in the study. The
observed
accuracy of the marker based registration is quoted as 0.6mm
whereas the best
registration methods give median errors of 0.7-1.5mm for
rectified MRI to
CT registration. To be considered as a gold standard a method
should
really approach an order of magnitude increase in accuracy.
One possible way to provide multiple images where the
transformation is
known to a very high degree of accuracy is to produce synthetic
scans.
Simple synthetic scans are possible but suffer from the fact that
they
do not produce realistic images with convincing image statistics.
Perhaps the VHD can help to provide more acceptable synthetic
scans.
This suggestion will be investigated in the next section.
The use of synthetically generated images from a high resolution
dataset to simulate a lower resolution scan has been applied to
MRI and
PET images with some success 4.
A
difficulty with this example is that PET images show function
whereas
MRI shows anatomy. Clearly, the cryosections from the VHD show
only
anatomical information. We will restrict the following discussion
to
cover production of synthetic anatomical datasets.
The VHD provides us with very high resolution information about a
particular subject. To use this information to create synthesised
anatomical images we need to establish a correspondence between
VHD
voxels and the intensity values expected from the simulated
imaging
modality. Some basic image processing techniques combined with
simple
segmentation may be sufficient to produce a reasonably convincing
effect. Some examples of this will follow.
Having established an intensity correspondence any transformation
can
be chosen and a synthetic scan for the desired region can be
created.
Since the synthesised scan will be of lower resolution each voxel
will
cover a number of VHD voxels. This will mean that any blurring
due to
interpolation will be small which is one reason why the VHD is
ideal
for this purpose. Multiple scans, of either similar or multiple
modalities, can be created. This will produce scans where the
transformation between them is known to a very high accuracy and
no
distortion is present.
Ideally synthetic images would come from a full simulation of the
imaging process itself. This would require a very complex
segmentation
in which each voxel is labelled according to the physical
properties
relevant to the imaging modality. Since many other groups are
working
on segmentation of the data we have not yet considered such a
task.
Two examples follow which demonstrate the main areas in which
such test
data may be applicable. Firstly, an MRI scan is produced which
shows
the feasibility of creating preoperative images from the
cryosection
data. Multiple such images could be used to test 3D image
registration
algorithms. The second example is an ultrasound image. Similar
images
could prove useful for testing of 2D-3D image matching for
intraoperative guidance.
MRI provides good contrast of soft tissues, but bone gives no
signal.
To produce something similar to an MRI scan, the cryosection
image is
first converted to a grey level image. In this case a simple
intensity
from RGB equation is used. The bone is then segmented and removed
from
the dataset. Finally, the data is blurred to match the resolution
of the
MR scan. This process is shown in Figure 4.
a) 
b) 
c) 
d) 
e) 
Figure 4. MRI scan created from cryosection data
a) colour cryosection, b) greyscale cryosection,
c) greyscale with bone removed, d) the previous image blurred,
and e) proton density MRI from similar slice.
The images shown in figure 4(d) and 4(e), whilst not showing
ideally
corresponding intensity variations, are of similar quality. The
fact that
blurred soft tissue cryosection data bears such resemblance to
MRI is
a testament to the quality of MRI in showing soft tissue detail.
Clearly
a better mapping of the cryosection data than simple intensity is
required
to provide a more convincing MRI scan.
Ultrasound imaging is based on the reflection of sound waves at
boundaries between tissues in which sound travels at different
speeds.
As a simple first step to creating a simulated ultrasound scan,
the
cryosection image is blurred and edge enhanced. The noise
is then increased and this image is further blurred in an arc
fashion
from the supposed probe position. This simulates the way in which
edges
that are perpendicular to the wave direction produce stronger
reflection. The result has similar quality to an ultrasound
image.
Here, however, simulation of the physics of the imaging process
will be
required to produce a convincing scan. The strength of the beam
should
decrease with each reflection and should also not penetrate bone
or
air. The results can be seen in figure 5.
a) 
b) 
c) 
d) 
Figure 5. Ultrasound scan created from cryosection data
a) abdomen cryosection, b) edge enhanced,
c) blurred in arc fashion d) real ultrasound
If a physically based model such as that proposed in part 1 can
be
created at the high resolution level of the VHD, this could
provide
detailed data about the shape and position of different tissues.
This
will produce a deformation which is potentially more accurate
than a
warp based on a CT or MRI scan alone. The MRI/CT datasets
produced from
the deformed VHD, either synthetically or by resampling of the
VHD
MRI/CT scans, may be used to assess the accuracy of a variety of
warping methodologies.
Since we have highly accurate data about the true anatomical
deformation, albeit synthetic, we can use this to examine
the performance of non-rigid registration algorithms.
Next: Summary
Up: Title and abstract
Previous: Atlases and Deformable
Models