Design team members: Subhash Ramanathan, Shingo Yuki
Supervisor: Ed Jernigan
Background
Film-less radiology promises to allow physicians to extract relevant information from medical images with the help of image-processing techniques. For instance, patient progress over time can be tracked by noting quantitative changes in the images. The challenge is to incorporate image-processing techniques so that the repetitive aspects of image analysis are automated as much as possible without sacrificing accuracy.
Inter-operator consistency and intra-operator consistency have always been a concern in the area of clinical radiology since the measurements taken by one trained operator inevitably differ from another’s. On the other hand, research-related activities, which involve working with large amounts of data, require some sort of automation to free the resources allocated to repetitive tasks. Incorporating image processing techniques into both a clinical and research environment introduces a certain level of consistency and intelligence that can help solve these problems.
EasyPax, a newly formed company based in Toronto, is currently developing a PACS (picture archiving and communication system) software package. In order to move to the forefront of PACS technology, EasyPax is looking to incorporate image processing based quantification techniques into their structured reporting software.
Project description
In this project we investigate liver ultrasound images. The liver is a fairly homogenous organ, in the sense that any given cross-section tends to be fairly uniform. This makes identification of abnormalities somewhat easier for non-experts who are not necessarily familiar with medical image analysis. Ultrasound liver images are convenient to work with due to their high availability and use in the medical community.
The objective of the project is to implement a methodology to measure the quantifiable characteristics (e.g. area, brightness, etc.) on a liver ultrasound image given the specific type of abnormality and seed points identifying the region of interest by the radiologist. Our goal is to create a software module that determines the extent/boundaries of a particular cyst or tumour given a priori information such as the type of abnormality and seed points that identify the location and shape characteristics of the abnormality.
Design methodology
Noise
in
ultrasound
images
Ultrasound
images
are
relatively
noisy
and
contain
speckle
noise.
Although
the
imaging
capabilities
of
modern
ultrasound
machines
are
increasing,
noise
remains
a
problem
when
trying
to
devise
image
processing
techniques.
We
found
that
median
filtering
and
homomorphic
filtering
using
a
Gaussian
Low
Pass
filter
in
the
density
domain
reduced
noise
as
well
as
preserving
edges.
Data
characteristics
In
discussions
with
EasyPax
we
found
that
tumours
appear
as
ring
shaped
objects.
Therefore
in
any
tumour
boundary
identification
method
we
could
check
to
see
whether
the
boundary
formed
is
circular.
Cysts,
on
the
other
hand,
form
distinct
dark
bounded
region
in
ultrasound
images.
Thresholding
to
identify
cysts
Cysts
can
be
handled
relatively
well
with
thresholding
techniques.
First
a
grey
level
histogram
of
the
affected
area
is
obtained.
Within
the
histogram
we
identify
the
mode
grey
level
value.
We
identify
as
a
cyst
any
pixel
that
has
a
grey
level
value
equal
to
or
below
the
threshold.
This
method,
after
being
run
through
a
median
filter
to
remove
outlying
points,
produced
a
distinct
binary
image
showing
the
cyst.
Model
based
and
region
growing
We
are
attempting
to
take
1-dimensional
intensity
profiles
of
the
tumour
at
different
angles
about
a
common
centre
co-ordinate
(most
likely
a
radiologist
defined
seed
point),
and
generating
scalograms
for
each
profile
using
the
continuous
wavelet
transform.
By
viewing
the
scalograms
we
can
decompose
the
profile
into
signal
and
noise
components.
We
presume
that,
at
the
intersection
of
the
profile
and
the
hypoechoic
ring,
there
is
a
significant
dip
in
the
signal
that
can
be
captured
as
a
local
minimum
of
the
profile.
We
then
choose
the
two
minima
closest
to
the
common
centre
co-ordinate
and
log
the
corresponding
co-ordinates
as
datum
points.
Since
the
centre
point
is
presumed
to
be
known,
we
can
attempt
to
fit
a
circle
to
the
data
points
by
choosing
a
radius
that
minimizes
the
least
square
error
between
the
predicted
radius
and
the
actual
distance
between
the
centre
point
and
a
datum
point.
Using one seed point as an initial position, we employed a simple region growing technique to try to isolate the pixels occupied by the ring (tumour). On a median filtered image we located a point within a tumour and searched its eight neighbours for the pixel that had the closest grey level value. The region of interest of the ring was well identified for 100 iterations. When we use a large number of iterations, to further expand our region we found that the region that is grown moves out of the tumour boundary, and into other parts of the image. This method needs to be enhanced to take into account directionality and the fact that tumours generally appear as rings.
Preliminary conclusions and future directions
Identifying cysts is a fairly straightforward method using histogram analysis and thresholding along the dominant grey levels. Passing the thresholded image through a median filter results in a well-bounded cyst. We feel this approach is adequate in identifying cyst-related metastases.
The model-based method of fitting the data to a circle needs to be improved to account for irregularities in the tumours shape. One possibility is to perform principal component analysis in order to set up a linear regression on an ellipse. The principal components would help define the orientation and approximate lengths of the axes, which in turn would help generate a radial distance versus angle plot. The aim then is to minimize the error between this plot and the radial distances of the datum points. Another possible strategy is to move from a regression-based approach to an adaptive deformable-model-based approach that would allow us to more accurately fit the tumour contour.
Region growing proves to be a viable method in identifying the tumour. In the future we hope to incorporate a model-based approach to correct for errors in the region formed by growing. A model-based approach will also help in directing our search, especially through parts of the tumour which are not well defined.