MR scanner (General Electric. Medical. Systems. Milwaukee,. WI). The remote teleradiology computer used was a Macintosh. 7100. PowerPC. (Apple Computer,.
Computers WaveletTransform-Based Transmission of MR J. A. Maldjia&,
was
W. C. Liu, 0. Hirschorn,
OBJECTIVE. to develop
teleradiology transfer form
The
an
image-transmission
of MR image
studies
wavelet
for trans-
system
We
for
describe
implementing
an
effi-
teleradiology
capability for transmission of diagnostic MR images. The system uses the wavelet transform to achieve sion, with
greater standard
of less than method using
than 90% image compresmodem transmission times
5 sec per compressed
is inexpensive
The
and can be implemented
commonly
MR scanner
image.
available
and
capabilities.
ily
W. Semanczuk
more
take
compression three
tion
of teleradiology
systems.
of image compression gories: lossless and retain
the fidelity
limited
to two
of the original to three
factors
Although within
ments
by
connected
Transmitting a standard
Received
modem
line (28.8
with
image
Materials System
been
described
Wavelet
with
transform,
fully
operational
functions).
Similar
ever,
unlike
tions
are very
com-
transform
sines
present We cient
of representing
data
this repre-
sines
transform,
cosines,
cosines, in space allows
such
teleradiology
imaging
study
over
transform image compression. can transmit diagnostic-quality
kBps)
would
eas-
in less than
5 sec per
selection,
and
All
basic
compression,
for
was
Supercomputing
Development.
residing
system
data
functions
Applica-
on the remote
designed
IL) and
Champaign,
computer.
specifically
to
work
The with
MR images. eliminating the need for a film digitizer. Figure 1 diagrams the system’s components and overall design. The backbone of the system consists of a SPARC2O workstation (Sun Microsystems, Mountam View, CA) on-site connected via the conventional
1.5-T Signa Horizon MR scanner (General Electric Medical Systems. Milwaukee, WI). The remote
do not
hospital
teleradiology
Ethernet
network
computer
used
(10
was
Mbps)
a
installed
effi-
on wavelet The method MR images
image.
and a 28.8-kBps
modem.
Both
to a
Macintosh
PowerPC (Apple Computer, Cupertino, running System 7.5.3 with Open Transport 7100
typically
highly
these
Software
software
and
data
func-
at characterizing
based
tions
transmission,
at the
has four
How-
the wavelet
as the data
system
and funcwavelet
wavelet
in an image. describe an inexpensive,
21, 1996; accepted after revision January 8, 1997. Department of Radiology (UH-C320), University of Medicine
and cosines
in frequency.
to be very efficient
discontinuous
system
Center
wavelet
This
image
The
(National
and
localized
to infinity.
end).
the remote
the
and
the need for any intervention
(sending
meth-
are more complicated wavelets and mother
localized
to be
(receiving
compression
using
to sines
are
site
decompression and are controlled from site using a standard Telnet connection
to the fast
data
was designed remote
viewing.
transform,
For
the basis functions have been called
site
the
[1-4].
is a method
is performed
host
system from
methods
similar
For a fast Fourier
of image
for lines.
of the original of 20-30: 1 have
these
transform-based
continue
high-speed
features ratios
ods fall into this category. The wavelet transform, data.
Methods
size
components:
functions
methods environ-
and
Design
CA) 1.1
computers
had Interactive Data Language (IDL) version 4.0.1 installed (Research Systems, Boulder, CO). The other software programs used were all shareware including
Fetch
(Dartmouth
College,
Hanover,
NH), Telnet, and Gnuzip (GNU, Free Software Foundation. Cambridge. MA). Gnuzip is a pro-
November
tAll authors: spondence
MR
each
(assuming
The teleradiology
important Compression
but are
institutional
of 30 images
lossless
used
taming image.
image,
lossless contained
if only
were
end) without
tions.
methods
such methods are impractical across conventional phone an entire
in
fall into two broad catelossy. Lossless methods
pression [1, 2]. may be workable direct lines, transmission
These
an hour
Lossy compression methods offer the benefit of higher compression ratios while main-
sentation
mage compression technologies are of paramount importance the development and implementa-
than methods
series
(basis
I
for
Data
Fourier
workstation
Compression
of 137 kB).
compression.
CONCLUSION. cient
system
using
Image
R. Murthy,
of this study inexpensive
purpose
effective,
in Radiology
and Dentistry
of New Jersey-New
Jersey Medical
School, 150 Bergen St., Newark,
NJ 07103. Address
corre-
to J. Maldjian.
AJR 1997:169:23-26
0361-803X/97/1691-23
AJR:169, July 1997
© American
Roentgen
Ray Society
23
Maldjian
et al.
achieved.
For
threshold
is selected
example,
for a 90% such
that
compression,
90%
a
of the wavelet
fall below this cutoff and are set to 0. In sparse storage mode, these zero values are eliminated, resulting in a compact array containing only 90% of the original wavelet-transformed data. coefficients
GINX
MR Scanner Data
Base
(compress
Ethernet
& zip)
This
Fetch modem
PowerMac
concept
accurate
Home
7100
is deceptive
in terms
image
The
file.
pression
is desired,
of elements
& decompress)
by hospital Ethernet lines to scanner
system disk. Remote computer can connectto SPARC2Oworkstation bystandard lnternetTCP/IP connection. MR images are transferred to SPARC2O using GINX. MR images are then waveletand Gzip-compressed. Compressed images are transferred by modem using Fetch to PowerMac, where compression is reversed.
are
cient),
which
Electric
Medical
files
nations
currently
can
be decompressed
across
different
com-
prompts
puter platforms. The
user
to the
a remote
performs
SPARC2O
workstation
percentage logs
into
the MR
IDL
fers
the
appropriate
and invokes gram logs plays the prompted percentage
the teleradiology program. The proonto the scanner system disk and discurrent examinations. The user is for examination number and desired of compression. The appropriate exam-
ncr
is then system
disk
connected,
the user
transferred
directly
to the
SPARC2O.
initiates
from
the
The
compression
for
extracted system
files
disk,
The
transfer
Within
IDL,
the user
(down
to
selected
images.
Thus.
then
uses
Fetch
to transfer
connection
and
the compressed
already allows
some
images
the compression
of the
specifically
lion
Image
starts
the
IDL
compression
(Gzip).
on the Macintosh
the MRUNCOMPRESS widget-driven pression Image [‘he
viewing
PowerPC
program.
graphical and
The
user
user and
This
interface
invokes a
for decom-
of the images.
Selection imaging
network
at our
institution
is con-
nected via a 10-Mbps Ethernet line to an on-site SPARC2O workstation used exclusively for image processing. Images to be compressed must first be transferred to this workstation. This transfer is initiated
remotely
by
a Telnet
connection
to the
SPARC2O. In order to keep user functions to a minimum, all on-site transfer and image compression is handled through a single command, MRCOMPRESS, invoked in IDL. This program invokes an
24
a new
the
program trans-
examination
SPARC2O,
data),
doubles
routine
more
no
is relative
on images
that
and
as
coefficients
integer
arrays
greater
SPRSIN
function
one
20
on the basis of the desired percentage This thresholding is accomplished
retaining
containing
only
the
than returns the
elements
the specified a structure
surviving
wavelet
scale.
and
mode.
10-1000,
This
depending
mapping
is facilitated
is to map This
converts
sparse with
absolute
threshold. of two
the
storage The arrays,
available
or less,
on
within
by storing
the position
integer
=
into
16-bit
images
a unique
of
value
size
must
be
There are 16-bit integer
in the array.
values
in the
-32,767
to +32,767.
indexed
x 256
for
to full
we have
vector
only
position
from
the arrays
modification
because
for every
(256
in restoring
is possible
216 possible
from
65,536).
Array
ele-
0 to the size
of the
the position
val-
Thus,
can be mapped to the integer 16-bit scale by subtracting 32,767 from each position. With these modifications, the 137-kB ues for a 256 x 256 element
image
is now
wavelet
reduced written
A final lossless the
array
to
compression
files are then
coefficients
and the other containing an index to the position of the coefficients in the original transformed matrix. It is within this step that the wavelet compression is
x 256
array
used
The second
are usually
magnitude
of
sparse
ments
mode,
during
in
can
coefficients
which
a nearest-integer
be
on
are
function,
use
these
with
array into row-indexed
during
mapping
is desired
ranging
the IDL SPRSIN
is achieved scale
this
the 16-bit
scale,
using
not
by a factor
filter
transformed
reversing do
this
an integer
been
using
wavelet
to
[5]. Rather,
format.
resulting
data
reduced
used in other wavelet-coding we attempt to maintain one to points of accuracy by multiplying the
has
256
The
data,
into
We
The images are compressed using the built-in wavelet transform function (WTN) within IDL. This function performs a discreet wavelet transform coefficients. thresholded compression.
is only
This com-
Compression
filter
file size
coefficient data
decompression.
storage
wavelet
wavelet-transformed
the largest factor that will maintain
selected.
a Daubechies
an
compression
the size of the original
have
used
images.
16
or CT
the largest 10 coefficients in 32-bit form. thereby decreasing the dynamic range of the data to be mapped. The mapping factor is stored with the
compression)
pertinent
imaging
the 90%
of the
image,
the
compression
schemes
of run-
was
are 32 bit, introducing
times
wavelet
by mapping
rule
has the option
space
image
50-60 kB with 90% compression, achieving actual file compression in the range of 40%. We achieve higher compression ratios by conveiling the 32-bit storage mode of the sparse matrix arrays to 16 bit, removing this factor-of-four increase in size from the original data. For the floating point
and
two coeffi-
of storage
for MR
Thus,
size
is four
in each
wavelet
if the original the case
to the
which
approximately
if more detail
are
However,
and
the amount
in size by two.
three decimal
also
of the
SPRSIN
image.
(position
the new arrays
increase
and series fold-
takes
by
of elements
original
Furthermore,
introduced
then
opens
The
been transferred to the SPARC2O. the user to change the percentage
over the modem connection. Fetch has been configured to automatically reverse the lossless porof
input.
and
and a
1 sec per 256 x 256 image.
scan-
pression
Telnet
to sorting
sion).
the
disk
number
examination
and
ning
closes
to
transferred
the images into the appropriate series and performs a wavelet transform-based compression. The wavelet-compressed data is then compressed further using a lossless technique (Gzip compresuser
the exami-
system
and automatically
scanner
sorted into appropriate
program
sorts
The
(General
on the SPARC2O. The image files are from the image database on the scanner
directory
ers.
displays
scanner
the user for an examination
host
Once
from GINX This
on the
login using Telnet at the
adapted
Systems).
on-site
institution.
ination
C program
compression
will be equal to the number
generated
set. For a 137-kB
external
coefficients,
the number
bit (as is frequently
data,
gram for performing lossless image compression (Gzip) using Lempel-Ziv coding. The compressed
although
wavelet
returned
arrays
in the
arrays required.
is networked
compression,
of actual
arrays
of the resulting
shows that SPARC2O workstation
90%
floating point (32 bit) for wavelet coefficients and long integer (32 bit) for matrix position. If no corn-
(unzip
Fig. 1.-Diagram
of
in terms of surviving
18-25
is applied.
kB
to the appropriate
image compression
wavelet-compressed
files
using
program
Gzip, which uses Lempel-Ziv
achieves
another
twofold
when
These
to threefold
90%
compressed directories.
is appliedlo the
shareware
coding. decrease
This in the
AJR:169, July 1997
Image size of the files, reducing range
(94%
of information
for
Transmission
of MR Data
them to the 9- to 12-kB
compressed
files).
This
compression
performed on the SPARC2O workstation. ing can be reversed on different platforms loss
Compression
using
is
The codwithout
the appropriate
shareware
program for that platform (i.e., MacGzip for the Macintosh). The process of wavelet compression. conversion to sparse mode. and lossless Gzip compression takes approximately 4 sec per 256 x 256 image
on the SPARC2O
Image
Transmission
The
wavelet-compressed
transmitted
Fetch.
This
to automatically
rates
via
Internet
is easily
Theoretically,
files
as they
high
and
rates
vented
by establishing
modem through
can
modest
volumes
the transmission
to the SPARC2O
a 28.8-kBps range
speeds,
transmitted
reduce
Image
be configured
connection
At the more
file
can Gzipped
using
a TCP/IP
are 2-3 kBps,
as 6 kBps.
are
the shareware
program the
files
using
using MacGzip (GNU, Software Foundation. Champaign, IL). Typi-
transmission
connected
site
unzip
are transferred Development
pressed
Gzipped
to the remote
program
cal
workstation.
the
as high
each
com-
in less than
5 sec.
of Internet
traffic
Fig. 2-67-year-old no compression, 94% compression,
man with hemorrhagic infarct of right basal ganglia. T2-weighted MR images displayed with 90% compression, and 94% compression show that details of image are retained at 90% and including hemorrhage with hemosiderin ring (arrowon 90% compressed image).
can
rate. This can be circuma direct
modem
connection
workstation.
Decompression
All image decompression and viewing is performed in IDL on the Macintosh PowerPC. A program was written with a graphical user interface for performing these functions. The decompression process
simply
restores
the
sparse
storage
form
of
the compressed images to full storage mode. remapping the wavelet coefficients to floating point (truncated at .01-.000l) and the matrix position vector to the appropriate matrix elements. The inverse
wavelet
transform
array to obtain ing
program
and
is then
the decompressed incorporates
magnification
simple
controls
images.
Reconstruction
image
on the Macintosh
applied image.
time
view-
window,
for
level,
displaying
for a single
7100
to this The
256
is less than
the x 256
4 sec.
Results The image
quality
the compression Figure
that can be achieved
scheme
2. These
outlined
images
with
is shown
are displayed
in
at origi-
nal resolution and at 90% and 94% compression. The original image file size was 137 kB. After wavelet and Gzip compression, file sizes were
reduced
The
90%
to 15 kB and 8 kB, respectively. compressed
image
loss of image
detail.
image
the key diagnostic
retains
original
image.
Even
Some
shows
the 94%
image
minimal
compressed
features blurring
of the
however,
particularly
at sulcal
boundaries
and
at gray
matter-white
matter
interfaces.
This
can be noted
AJR:169, July 1997
in cortex. (arrow) with 90% and 94% compression.
is evi-
dent,
in Figure
Fig. 3.-62-year-old man with small infarct in left motor cortex. A, T2-weighted MR images show subtle area of high signal intensity B, Balanced MR images also show subtly increased signal intensity
2, with progressive
decrease tion
in sulcal
of the internal
tralateral
to the
defmition
and loss of resolu-
and external area
capsules
of hemorrhage.
conDespite
this blurring,
subtle
at 94%
compression
pressed
image
can
detail (Fig.
can be retained 3). The
be transferred
90% over
even coma con-
25
Maldjian
ventional
modem
decompression compressed formed
line
within
see,
5
with
eral
in less than 4 sec. For the 94% image,
within
both
processes
Electric
easily
can be per-
More image
We
describe
an
networking
gram,
IDL, license.
data.
The
infrastructure
costs remote
7100,
obtained
for less than
software
used
about
is widely
imaging
require
are not capable data.
windows,
soft-tissue
within
the
dural
windows.
These
for
a
a Macinpow-
networking and can
be
via the Internet at no cost. The component of the system is the
the
transfer
as a separate
($20,000-$30,000), a similar
most workstation
to the MR scanner.
tions usually perform volume reconstructions Windows
These
pur-
university
diagnostic
the
associated
lower
substantially transfer
larger
Workstation;
times.
commercial
size
methods
(one
of
the would for use
images. involve require We
More more
easily
is
resulting for
these
is being
Tl -weighted,
complicated than
200
axial more cases
images,
which
120 easily would
transmission times of more than 1 hr. describe a teleradiology system that
can effectively compress, compress a 256 x 256 MR
and
approximate
original
raw
need
for
desired
data.
separate
images.
flexibility
trans-
The
system
to allow
compression
the user
ratio.
Fur-
the remote
login,
specific
key images
to compress
ratio
if better
method
can
resolution
be easily
institutions
components
the user at
is needed.
implemented
that already
have
at
the essen-
in place.
transfer, and image within
References 1. Aberle DR. Gleeson F, Sayre JW, et al. The effect of irreversible image compression on diagnostic accuracy in thoracic imaging. invest Radio! 1993:28:
398-403 2. Goldberg
MA,
Pivovarov
de10-
M. Mayo-Smith
WW,
et al. Application of wavelet compression tized radiographs. AiR 1994:163:463-468 3. Angelidis PA. MR image compression
wavelet transform
enhanced than can
This
quality
through
tial networking
are on the order of 30 this may be barely
T2-weighted, with
a lower
the
images
for
ratios
(sagittal
the
can select
many
adequate for a CT examination of 30 images, this would be unacceptable for an MR study sequences
to enter
with
with
resulting
of the
windowed
thermore,
sub-
which
the
was designed
and
times
of
compression The
range
eliminates
mission
images
and
Transmission
used by our department) sec per image. Although
already
Gem-
this
compression file
This
grab-
frequently
stage,
the dynamic
remote
the penalty
worksta-
CPU-intensive image and manipulations
methods
at the compression
dual-echo
26
maintain
of
window,
90-94%
transpar-
frame
separate
using
transform.
of transmitting
of
15 sec wavelet
faster
axial
investment
Advantage
be
compression. may perform
Tl-weighted)
(e.g.,
can
On most
some form of lossless image Although lossless methods
a substantial have
run
For a CF scan,
bone
computer,
available
which
raw
three
networked
can
methods
the
$1500
The
bers,
with
hospitals
but
rely on image
sys-
is not a particularly
$2000.
conventional transmission
SPARC2O workstation networked to the hospital MR scanner. Although this can represent chase
process
The
or expensive platform. More powerful, generation PowerPC computers can be
downloaded essential
this
using
setting. It requires the use available software pro-
which
tosh PowerPC
capa-
cost-effective,
university hospital of a commercially
erful later
of MR
is extremely
existing
easy-to-use
teleradiology
for transmission
tern
user
efficient,
for implementing
bility
Systems),
to run IDL as well.
ently in the background via remote login without affecting the normal daily activity.
Discussion
method
Medical
configured
workstations,
8 sec.
et al.
Imaging 4. Goldberg Making
coding algorithm.
sion
using
Magn
a
Reson
1994:12:1111-1120 MA, global
Sharif
HS,
telemedicine
Rosenthal
DI,
practical
architecture
in a PACS
for medical
environment.
Med
et al.
and afford-
able:demonstrations from the Middle 1994:163:1495-1500 5. Azpiroz Leehan J. Lerallut iF, Magana processor
to digi-
image
East. AiR I. A multicompres-
Prog
Technol
1994:20:101-110
AJR:169, July 1997