Last week, we examined the Pulse-Coupled Neural Network (PCNN)
developed at Battelle Memorial Institute for use in
medical image analysis. Briefly, a PCNN is a software and
hardware model of the mammalian visual cortex, which has
a remarkable feature called the pulse sychrony process.
Because of this feature, the visual cortex, and the
associated PCNN can make determinations about similarities
of structures in an image. The Battelle researchers have
taken advantage of this ability of their PCNN to study two
representative imaging applications : segmentation from
Magnetic Resonance Imaging (MRI) and pulmonary studies using
scintigraphic imagery. The major focus of these projects
was to determine the ability of the PCNN to resolve body
structures from both high-quality images and those with
more "noise".
Magnetic Resonance Imaging provides physicians with a high
quality image of the body part of interest with little
background "noise", or interfering signals from such sources
as machine frequencies, atmospheric contaminants, etc.
Images produced from MRI are often used in detecting tumors
and other abnormal tissue conditions. While most diagnoses
made from such images are done manually, it is often difficult
to determine with the naked eye if a particular region of
space could be trouble or not. Therefore, the Battelle
group used PCNN's to automatically manipulate the signal
in an attempt to resolve structures in an image.
In tests on images of both the human brain and the human
mid-section, the PCNN's were able to sufficiently enhance
MRI images to more clearly delineate major organ structures.
This could potentially be an important step to making more
accurate and complete diagnoses.
While MRI generates nice, clean images, nuclear scintographic
images are often grainy, with poor uniformity. This results
from a concern over patient safety : because active nuclear
material is involved in this form of testing, the dosing
is kept as low as possible to minimize exposure. This
results in much of the resultant signal being overshadowed
by ambient signals that have no meaning to a diagnosis.
If the PCNN's could enhance such an image, therefore, it
could be a potential boon to this type of testing.
In fact, the results of the PCNN segmentation of the
scintographs were somewhat mixed. While there was resolution
of the lung boundaries with the PCNN, it wasn't much
better than what more simple methods yield. In addition,
the PCNN segmentation of one lung image indicated and airway
obstruction where it is likely that none existed. So, it
appears that, while there may be some immediate problems
with using PCNN's for structure identification from grainy
images at the present, there is potential here.
In total, the Battelle group made a very impressive
