818 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 13, NO.

5, SEPTEMBER 2009

A Lossless Compression Method for Medical Image

Sequences Using JPEG-LS and Interframe Coding
Shaou-Gang Miaou, Member, IEEE, Fu-Sheng Ke, and Shu-Ching Chen

Abstract—Hospitals and medical centers produce an enormous
amount of digital medical images every day, especially in the form
of image sequences, which requires considerable storage space. One
solution could be the application of lossless compression. Among
available methods, JPEG-LS has excellent coding performance.
However, it only compresses a single picture with intracoding and
does not utilize the interframe correlation among pictures. There-
fore, this paper proposes a method that combines the JPEG-LS and
an interframe coding with motion vectors to enhance the compres-
sion performance of using JPEG-LS alone. Since the interframe
correlation between two adjacent images in a medical image se-
quence is usually not as high as that in a general video image
sequence, the interframe coding is activated only when the in-
terframe correlation is high enough. With six capsule endoscope
image sequences under test, the proposed method achieves aver-
age compression gains of 13.3% and 26.3% over the methods of
using JPEG-LS and JPEG2000 alone, respectively. Similarly, for
an MRI image sequence, coding gains of 77.5% and 86.5% are
correspondingly obtained.
Index Terms—Interframe coding, JPEG-LS, lossless compres-
sion, medical image sequences.
I. INTRODUCTION
OSPITALS and medical centers produce large volume
H of digital medical image sequences, including computed
tomography, MRI, ultrasound, and capsule endoscope (CE) im-
ages [1], etc. Theses sequences require considerable storage
space. One solution to this problem could be the application
of lossless compression. In fact, the need to compress medical
images is a driving force to develop lossless image compression
standards, such as JPEG-LS and JPEG2000. In addition, these
standards are considered in digital imaging and communica-
tion in medicine (DICOM), which is an industrial international
standard for medical image representation and coding [2].
JPEG-LS has excellent coding and computational efficiency,
and it outperforms JPEG2000 and many other image compres-
sion methods [3], [4]. JPEG-LS mainly consists of context mod-
eling, pixel prediction, and prediction error encoding [4], [5].
Manuscript received August 17, 2007. First published May 15, 2009; cur-
rent version published September 2, 2009. This work was supported in part by
the Chung-Shan Institute of Science and Technology, Lung-Tan, Taiwan, under
Grant BV95H02P.
S.-G. Miaou is with the Electronic Engineering Department, Chung-Yuan
Christian University, Chung-Li 32023, Taiwan (e-mail: miaou@wavelet.el.
cycu.edu.tw).
F.-S. Ke was with the Electronic Engineering Department, Chung-Yuan
Christian University, Chung-Li 32023, Taiwan. He is now with Top Victory
Electronics, Taipei County 23553, Taiwan (e-mail: alanke8@msn.com).
S.-C. Chen was with the Electronic Engineering Department, Chung-Yuan
Christian University, Chung-Li 32023, Taiwan. She is now with the CyberLink,
Taipei 231, Taiwan (e-mail: meet0512angel@yahoo.com.tw).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TITB.2009.2022971
1089-7771/$26.00 © 2009 IEEE
In the regular mode of JPEG-LS, a nonzero prediction error is
treated as an integer random variable that is entropy coded by
the efficient Golomb–Rice codes [6] using a context-based prob-
abilistic model. The Golomb–Rice codes can be implemented
at low computational cost. In the run mode of JPEG-LS, a run-
length coding scheme with low computational cost is used to
encode constant regions when a zero prediction error occurs.
Most lossless image compression techniques deal with sin-
gle image independently without utilizing the relation among
images in an image sequence. However, some literature does
exploit the interframe correlation for lossless compression of
CE images [7] and video [5], [8], [9]. The compression ratio
(CR) obtained in [7] is about 2.5, which is unsatisfactory be-
cause using JPEG-LS alone can obtain a CR value higher than
2.7 according to our experiment. Memon and Sayood [8] elab-
orated a promising entropy analysis for exploiting interframe
correlation, but no real encoding/decoding development was de-
scribed. The compression technique proposed by Brunello et al.
is based on motion compensation, optimal 3-D linear prediction,
and context-based entropy coding [5]. A small compression gain
with respect to the static JPEG-LS coding, applied on a frame-
by-frame basis, is obtained at the price of significant increase of
computational complexity, delay time, and memory cost. Zhang
et al. used the combination of temporal and spatial prediction
and context-based entropy coding [9], resulting in a 10% coding
gain over the state-of-the-art techniques, such as JPEG-LS and
context-based, adaptive, lossless image coding (CALIC). How-
ever, a 5% loss (not a gain) was obtained for one of the five video
sequences under test. Thus, exploiting temporal correlation does
not always work very well in lossless compression.
In this paper, a new lossless compression method based on
JPEG-LS and motion estimation is proposed. To avoid the pos-
sible coding loss and significant increase of computational cost
like those in [5] and [9], respectively, we only exploit high
enough correlation between adjacent image frames by using a
correlation estimation approach.
II. SEQUENCE COMPRESSION WITH JPEG-LS
AND MOTION ESTIMATION
A. Overview of the Proposed Approach
A processing flow chart for the encoder side of the proposed
method is shown in Fig. 1. Here, an 8-bit grey-level image or
an 8-bit color component of a color image is considered. Given
an image sequence, the first input image is always compressed
by 8-bit JPEG-LS coding alone due to no reference frame, then
the second frame will be the current frame, and the first frame
becomes the reference frame, and so on. First, estimate the

Authorized licensed use limited to: R.V.S. College of Engineering and Tech. Downloaded on April 29,2010 at 10:24:27 UTC from IEEE Xplore. Restrictions apply.
MIAOU et al.: LOSSLESS COMPRESSION METHOD FOR MEDICAL IMAGE SEQUENCES USING JPEG-LS AND INTERFRAME CODING
Fig. 1. Processing flow chart for the encoder side of the proposed method.
correlation (see Section II-B) between the current frame to be
encoded and the reference frame in the buffer. If it is lower than a
correlation threshold, we mark bit “1” as a flag and compress the
current frame with 8-bit JPEG-LS coding. Otherwise, we mark
bit “0” and perform motion estimation and motion compensation
(MEMC) as in traditional video coding, followed by subtract-
ing the reconstructed frame (which has gone through the MEMC
process) from the current frame to obtain a difference image.
The difference image is then preprocessed and encoded by using
9-bit JPEG-LS compression (see Section II-B). The motion vec-
tor (MV) derived from motion estimation will be encoded (see
Section II-B). Finally, all the flag bits and the encoded bits for
one image frame are concatenated into single bit stream. This
processing cycle repeats for next image frame until the end of the
image sequence. The decoder simply implements a reverse oper-
ation of encoding, including standard 8-bit JPEG-LS decoding
or MV-based decoding plus 9-bit standard JPEG-LS decoding.
B. Further Details
1) Determining Correlation Between Adjacent Image
Frames: The correlation is determined by estimating the vari-
ance of the difference image. To find a proper threshold σT2
for separating high and low correlation, we examine the com-
pression results obtained when the two CE image sequences
Authorized licensed use limited to: R.V.S. College of Engineering and Tech. Downloaded on April 29,2010 at 10:24:27 UTC from IEEE Xplore. Restrictions apply.
819
Fig. 2. Example images of the two CE image sequences used in the preliminary
tests: (a) #4, (b) #6, and (c) MRI.
(#4 and #6), as shown in Fig. 2, are tested. One sequence is
smooth in the sense that images are similar and change gradu-
ally, while the other contains complex pathological images that
change abruptly. The best CR occurs when σT2 = 180. Thus, we
set σT2 = 180 as our desired threshold.
2) Macroblock Size and Searching Range: Again, the CE
image sequences of #4 and #6 are tested. Various search ranges
and MB sizes are considered in the test. The macroblock (MB)
size and the searching range of our choice are 16 × 16 and
[−7, 7], respectively, because they provide the largest average
CR.
3) Encoding of MVs: First, the x and y components of a
MV are treated separately. For the x component, a bit 0 is
sent or stored if it is zero; a bit 1 is sent or stored, otherwise,
followed by the 4-bit binary representation of x component.
This coding scheme is applied to y component as well. This
coding design is based on the observation that a zero MV in
x- or y-direction occurs much more frequently than other MVs.
This coding approach is more efficient than the direct 4-bit
natural code for each component but could be less efficient than
a more powerful entropy code. However, the proposed approach
is essentially a fixed length code, and it is thus simple and easier
to encode and decode than the more powerful variable length
entropy code.
4) Preprocessing of Difference Images: Since the pixel value
of a difference image is in the range from −255 to 255 and a
P-bit JPEG-LS coding can only handle the input with sample
value from 0 to 2P –1 for a positive integer P, each pixel value
of a difference image is added by 28 –1 = 255, resulting in the
new value range of [0, 510]. Then, we apply a 9-bit JPEG-LS
coding to the resulting image.
III. EXPERIMENTS AND RESULTS
The personal computer we used for the experiment is
equipped with 2.4-GHz CPU, 266-MHz double data rate II
820 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 13, NO. 5, SEPTEMBER 2009
TABLE I
CRS OBTAINED BY COMPRESSING R, G, AND B COMPONENTS INDEPENDENTLY
USING JPEG-LS
Fig. 3. Comparison of compression results for CE image sequence #1.
(DDRII), and 512-MB RAM, and the main software used is
Microsoft Visual Studio .Net 2003. The image data for testing
include six color CE image sequences #1–#6 with different im-
age characteristics and one set of gray-level MRI sequence (as
shown in Fig. 2). Each CE image sequence consists of 100 im-
age frames and the MRI sequence contains 75 frames. All the
images are 256 × 256 and 8 bit for each color component.
The R, G, and B components of the CE image sequences
are compressed independently with JPEG-LS coding and the
resulting CRs are given in Table I. Obviously R component
is somewhat more difficult to compress than the other two.
For simplicity, we will present the compression results for R
component of the CE images from now on and expect to have
slightly better results for G and B components.
A. Correlation Estimation Between Adjacent Frames
The compression performance for CE Sequence #1 is shown
in Fig. 3, where the dashed line denoted as JPEG-LS shows the
result obtained by applying JPEG-LS directly and separately
to each image frame; the solid line denoted as JPEG-LS+MV
gives the result obtained by applying MEMC and JPEG-LS
for every image frame in the sequence; the solid line denoted
as JPEG-LS+VAR+MV shows the result obtained by either
applying JPEG-LS directly or MEMC plus JPEG-LS according
to the result of correlation estimation; and the solid line denoted
as JPEG2000 gives the results obtained by applying JPEG2000
to each image. The results for other CE image sequences are
similar and are not shown here. As expected, MEMC does not
work very well in some image frames and the successful result
of correlation estimation can be verified easily from Fig. 3.
Note that not every decision based on correlation estimation is
satisfactory. See the results for the 41th and the 81th frames, and
for the frames around the 70th frame in Fig. 3 as an example.
Authorized licensed use limited to: R.V.S. College of Engineering and Tech. Downloaded on April 29,2010 at 10:24:27 UTC from IEEE Xplore. Restrictions apply.
TABLE II
PERFORMANCE IN AVERAGE CR FOR THE CE IMAGE SEQUENCES
TABLE III
PERFORMANCE IN AVERAGE CR FOR THE MRI SEQUENCE
Fig. 4. Compression results for the MRI sequence.
However, in general, most decisions based on our correlation
estimation result are correct. In fact, all wrong judgments can
be avoided if both approaches are applied to the same frame and
choose the one with better performance. However, this would
add a little coding gain at the expense of too much computational
effort.
B. Performance Comparison and Discussion
Fig. 3 has demonstrated that JPEG-LS performs better than
JPEG2000 on a frame-by-frame basis for sequence #1. Next,
we examine the performance of average CR and the results
are shown in Table II, where JPEG-LS and JPEG2000 mean
using JPEG-LS and JPEG2000 alone, respectively, and JPEG-
LS+VAR+MV represents the proposed method. The results
show that the proposed method can provide a coding gain rang-
ing from 4% to 31.6% over JPEG-LS and 15% to 51.9% over
JPEG2000.
We also test the proposed method with the MRI sequence. The
result of average CR is shown in Table III, and the result for each
image frame is shown in Fig. 4. The coding gain provided by the
proposed method is very significant. Although the parameters
of the proposed approach are tailored for the CE images, they
also work well for the MRI images.
Next, we consider the computing time. In order to have
the most accurate searching result and the best compression
MIAOU et al.: LOSSLESS COMPRESSION METHOD FOR MEDICAL IMAGE SEQUENCES USING JPEG-LS AND INTERFRAME CODING
efficiency, this paper uses the full search and mean absolute er-
ror (MAE) computation in MEMC. However, it is well known
that the full search and MAE operation is quite time consuming.
The time for compressing a 100-image sequence is 29.83 ± 0.47,
24.75 ± 2.61, and 4.93 ± 0.33 s when we use JPEG-LS+MV,
JPEG-LS+VAR+MV, and JPEG-LS only, respectively. This
result demonstrates the expensive time demanding of the full
search and MAE computation in the MEMC process. The com-
pression time difference between the use of JPEG-LS and the
proposed method for one image frame may be ignored, but if
an 8-h long sequence of CE images is considered, this time dif-
ference is too significant to be neglected. Fortunately, there are
many fast searching methods for block matching.
The authors in [10]–[12] study how accurate can a fast search-
ing method find the best matched MB (obtained by the full
searching method, i.e., full search and MAE) and the searching
time saved by using the fast searching methods. It is concluded
that the diamond search (DS) method [10] is the best among
the methods under study in terms of both accuracy and speed.
With the DS method, the accuracy is 90% and the time saved
is 93% on the average. Therefore, if we took DS as the search
method, the compression time of the proposed method might
become comparable to that of using JPEG-LS only and the CR
performance might degrade slightly.
IV. CONCLUSION
The proposed method performs better than using JPEG-LS
and JPEG2000 alone in CR by about 13.3% and 26.3% on the
average for CE images, respectively. Similarly, the coding gains
for MRI images are about 77.5% and 86.5%, respectively. Since
not every correlation in adjacent frames can be exploited effec-
tively in lossless compression, the proposed method achieves
such coding gains by judiciously exploiting the correlation be-
tween adjacent frames with a low cost correlation estimation
method.
The MEMC process could be implemented by a faster version
for some applications. Better correlation estimation methods
could be developed for more image types. To further improve the
coding performance of the proposed method, we may consider
sharing MVs among multiple color components, using color
transformation, reordering image frames, and designing entropy
codes for the MVs.
REFERENCES
[1] D. G. Adler and C. J. Goustout, “Wireless capsule endoscopy,” Hosp.
Phys., vol. 39, no. 5, pp. 14–22, 2003.
[2] (2008, Feb. 6). [Online]. Available: http://dicom.nema.org/
[3] G. Schaefer, R. Starosolski, and S. Y. Zhu, “An evaluation of lossless
compression algorithms for medical infrared images,” in Proc. IEEE Eng.
Med. Biol. Conf., Sep. 2005, pp. 1673–1676.
[4] M. J. Weinberger, G. Seroussi, and G. Sapiro, “The LOCO-I lossless
image compression algorithm: Principles and standardization into JPEG-
LS,” IEEE Trans. Image Process., vol. 9, no. 8, pp. 1309–1324, Aug.
2000.
[5] D. Brunello, G. Calvagno, G. A. Mian, and R. Rinaldo, “Lossless compres-
sion of video using temporal information,” IEEE Trans. Image Process.,
vol. 12, no. 2, pp. 132–139, Feb. 2003.
[6] M. Weinberger, G. Seroussi, and G. Sapiro, “LOCO-I: A low complex-
ity, context-based, lossless image compression algorithm,” in Proc. IEEE
Data Compression Conf., Snowbird, UT, Mar./Apr. 1996, pp. 140–149.
Authorized licensed use limited to: R.V.S. College of Engineering and Tech. Downloaded on April 29,2010 at 10:24:27 UTC from IEEE Xplore. Restrictions apply.
821
[7] Y. D. Wang, “The implementation of undistorted dynamic compression
technique for biomedical image,” Master’s thesis, Dept. Electr. Eng., Nat.
Cheng Kung Univ., Taiwan, 2005.
[8] N. D. Memon and K. Sayood, “Lossless compression of video sequences,”
IEEE Trans. Commun., vol. 44, no. 10, pp. 1340–1345, Oct. 1996.
[9] M. F. Zhang, J. Hu, and L. M. Zhang, “Lossless video compression using
combination of temporal and spatial prediction,” in Proc. IEEE Int. Conf.
Neural Netw. Signal Process., Dec. 2003, vol. 2, pp. 1193–1196.
[10] L. M. Po and W. C. Ma, “A novel four-step search algorithm for fast block
motion estimation,” IEEE Trans. Circuits Syst. Video Technol., vol. 6,
no. 3, pp. 313–317, Jun. 1996.
[11] S. Zhu and K. K. Ma, “A new diamond search algorithm for fast block-
matching motion estimation,” IEEE Trans. Image Process., vol. 9, no. 2,
pp. 287–290, Feb. 2000.
[12] J. L. Guo, “Thresholded two-path search method for video block motion
estimation,” Master’s thesis, Dept. Comput. Commun. Eng., Kaohsing
First Univ. Sci. Technol., Taiwan, 2001.
Shaou-Gang Miaou (M’94) was born in Taipei,
Taiwan, in 1961. He graduated from the Cheing-
Kuo Senior High School, Taipei, in 1979, and re-
ceived the B.S. degree in electronic engineering from
the Chung-Yuan Christian University, Chung-Li,
Taiwan, in 1983, and the M.S. and Ph.D. degrees in
electrical engineering from the University of Florida,
Gainesville, FL, in 1990 and 1993, respectively.
From 1983 to 1988, he was a Research Assistant at
the Chung-Shan Institute of Science and Technology,
Lung-Tan, Taiwan. From 1989 to 1993, he was either
a Teaching Assistant in the Department of Electrical Engineering, University of
Florida, or a Research Assistant in the Center for Information Research, Uni-
versity of Florida. In 1993, he joined the Department of Electronic Engineering,
Chung-Yuan Christian University, Chung-Li, Taiwan, as an Associate Professor.
He is currently a Professor and the Chairman of the Department of Electronic
Engineering, Chung-Yuan Christian University. His current research interests
include data compression, pattern recognition, and wireless multimedia commu-
nications. He is an Editor for the International Journal of Electrical Engineering
and the Journal of Advanced Engineering. He has authored or coauthored over
100 technical papers in referred journals and conference proceedings. He holds
two invention patents with nine others pending. He is the author or coauthor of
several books on digital image processing, communication systems, probability
and statistics, and engineering mathematics.
Prof. Miaou is a permanent member of the Image Processing and Pattern
Recognition Society of R.O.C., and a member of the Biomedical Engineering
Society of R.O.C. He has been a Reviewer for many technical projects, con-
ferences, and prestigious journals, including IEEE TRANSACTIONS ON BIOMED-
ICAL ENGINEERING, IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, IEEE
TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, IEEE TRANS-
ACTIONS ON MEDICAL IMAGING, IEEE TRANSACTIONS ON MULTIMEDIA, and
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS.
Fu-Sheng Ke was born in Taipei County, Taiwan,
in 1982. He received the B.S. and M.S. degrees
in electronic engineering from the Chung-Yuan
Christian University, Chung-Li, Taiwan, in 2004 and
2006, respectively.
He was involved in the research of data com-
pression and its application to medical images. He
is currently an R&D Engineer with the Top Victory
Electronics, Taipei County, Taiwan.
Shu-Ching Chen was born in Taipei, Taiwan, in
1983. She received the B.S. and M.S. degrees in elec-
tronic engineering from the Chung-Yuan Christian
University, Chung-Li, Taiwan, in 2005 and 2007,
respectively.
She was involved in the research of digital wa-
termarking for medical images. She is currently an
R&D Engineer with the CyberLink, Taipei, Taiwan.