Content uploaded by Nacéra Benamrane
Author content
All content in this area was uploaded by Nacéra Benamrane on Dec 25, 2017
Content may be subject to copyright.
www.kips.or.kr
Copyright© 2017 KIPS
Adaptive Medical Image Compression Based on
Lossy and Lossless Embedded Zerotree Methods
Sid Ahmed Elhannachi*, Nacéra Benamrane*, and Taleb-Ahmed Abdelmalik**
Abstract
Since the progress of digital medical imaging techniques, it has been needed to compress the variety of
medical images. In medical imaging, reversible compression of image's region of interest (ROI) which is
diagnostically relevant is considered essential. Then, improving the global compression rate of the image can
also be obtained by separately coding the ROI part and the remaining image (called background). For this
purpose, the present work proposes an efficient reversible discrete cosine transform (RDCT) based embedded
image coder designed for lossless ROI coding in very high compression ratio. Motivated by the wavelet
structure of DCT, the proposed rearranged structure is well coupled with a lossless embedded zerotree
wavelet coder (LEZW), while the background is highly compressed using the set partitioning in hierarchical
trees (SPIHT) technique. Results coding shows that the performance of the proposed new coder is much
superior to that of various state-of-art still image compression methods.
Keywords
LEZW, Medical Images, ROI, RDCT, SPIHT
1. Introduction
Currently, medical image compression is done with lossless or near-lossless of information, to ensure
data integrity and avoid misdiagnosis due to degradation of image quality especially in parts called
regions of interest (ROI). However, this type of compression (lossless) doesn’t offer significant
reduction in the volume of such data. In this context, the lossy compression also has many advantages
for applications other than the diagnosis.
To find a good compromise between compression ratio and validation of diagnostic quality of
compressed images, it seems an advantage of combining the two techniques on a medical image. So
after an extraction step of the region of interest ROI in the image (e.g., tumor area in a brain magnetic
resonance image [MRI]), a lossless compression can be applied to the region of interest while the
remaining part called background will be compressed with loss.
In a medical image, a region of interest can be selected manually or else detected in a semi-automatic
or automatic manner. Detecting of the ROI part has been studied for several years and the existing
※ This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/lic enses/by-nc/3.0/) which
permits unrestricted non-commercial use, d istribution, and reproduction in any medium, provided the original work is properly cited.
Manuscript received July 6, 2015; first revision June 15, 2016; accepted August 31, 2016.
Corresponding Author:
Sid Ahmed Elhannachi (sidahmed.elhannachi@univ-usto.dz)
* Dept. of Computer Science, Faculty of Mathematics and Computer Science, University of Science and Technology of Oran
‘Mohamed Boudiaf,’ USTO-MB, BP El M’naouer, Oran, Algeria ({sidahmed.elhannachi, nacera.benamranei}@univ-usto.dz)
** Dept. of Automatic, University of Valenciennes and Hainaut-Cambrésis (UVHC), Valenciennes, France (Abdelmalik.Taleb-Ahmed
@univ-valenciennes.fr)
J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017
ISSN 1976-913X
(Print)
https://doi.org/10.3745/JIPS.02.0052
ISSN 2092-805X
(Electronic)
Sid Ahmed Elhannachi, Nacéra Benamrane, and Taleb-Ahmed Abdelmalik
J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017 | 41
algorithms include ROI extraction are based on low-level features or geometric properties [1-4].
JPEG2000 standard is a compression technique that can cover the needs of compression of medical
images. This method can perform a lossless compression in regions of interest and lossy compression
elsewhere [5].
In the literature, several existing works discussed the use of the ROI to improve medical image
compression. In [6], a multi-ROIs algorithm for medical image compression is proposed. Based on the
segmentation of Canny operator, the image is divided into ROIs and non-ROIs. In the ROIs, lossless
compression is applied, based on JPEG2000 algorithm. The set partitioning in hierarchical trees
(SPIHT) encoder was used to encode non-ROIs.
In [7], an approach based context-regions of interest to compress vascular images for peripheral
arteries, has been proposed. The vascular image is divided into primary regions (PROI), secondary
regions (SROI) and the background. Based on the JPEG2000 algorithm, these multiple ROIs was
compressed at various degrees of interest.
A study [8] presented a ROI based method with SPIHT algorithm. The scheme used a flexible
method to determine the ROI into the DWT and accepts multiple regions by defining the ROI mask on
the low frequency sub-band which reduces the amount of bits needed to the mask. In [9], a
compression method is proposed to compress the liver cancer computed tomography (CT) images. The
ROIs are compressed using the lossless Huffman encoder, while the remaining irrelevant region of
image is compressed using the near-lossless image compression techniques: SPIHT, embedded zerotree
wavelet (EZW), and DCT-RLE.
A study [10] has investigated an approach for medical image compression using ROI-EZW. Lossless
integer wavelet transform was included to partial EZW to encode the ROI part. The partial EZW uses
conventional EZW when the frequency of one’s is less than 0.2 and two new options are used otherwise,
making it more superior to EZW and SPIHT algorithm.
Another ROI coding scheme has been presented in [11] where the ROI part is selected by the user.
Using the fractal encoding, the ROI part is encoded by employing wavelet transform based on listless
SPECK (LSK), giving high compression ratio and better peak signal-to-noise ratio (PSNR) values.
The authors in [12], proposed a ROI coding approach for abdominal CT image. Divided manually,
the region of no interest was coded by a hybrid technique based on fractal and improved SPIHT. The
Huffman technique was used to compress the region of interest, for lossless requirements.
In [11,12], measure performance showed that the PSNR values obtained with ROI compression are
not so high compared with those obtained in compression without ROI, this is probably due to the use
of fractal method which in some cases, cannot get enough information from the detail images of the
wavelet transform. Also, the Huffman coding does not always permit to obtaining high PSNR values.
A solution consists to use a robust lossless compression technique on the ROI part. This is an
important requirement for medical imaging domains, where, mainly high quality is demanded. The
algorithms for lossless compression can be further categorized into predictive coding and transform
coding algorithms.
Predictive coding requires intensive computation and relies on extensive side information to
implement the sophisticated predictors such as clustered differential pulse-code modulation (C-DPCM)
[13] and context-based adaptive lossless image coding (CALIC) [14,15].
Transform coding algorithms are more computationally efficient, but typically do not equal the
Adaptive Medical Image Compression Based on Lossy and Lossless Embedded Zerotree Methods
42 | J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017
lossless compression performance of predictive algorithms. Since the wavelet transform holds the
properties of the multi-resolution analysis and energy compactness, most transform based compression
algorithms are based on these. Therefore, transform based coding has become the popular approach
because of its high efficiency, as well as other advantages including progressive transmission, random
access to the bit stream, ROI coding.
Particularly, several works have shown that EZW wavelet-based embedded image coders help to
provide an excellent rate-distortion performance since dependencies of wavelet coefficients in subbands
are well exploited for high resolution images [16-18]. However, in very low bit rate image coding using
low resolution images (e.g., ROI size of 100×145) the embedded zerotree coding performance is
prominently reduced due to the insufficient wavelet decomposition.
For this purpose, beside wavelets, DCT based zerotree coding applications were investigated and used
by several researchers [19-22]. These works showed that DCT-based embedded coders can provide
competitive compression rates with a good image quality compared to the wavelet based embedded
coders.
In this paper, the aim of the proposed medical image compression approach is to compress the
important region strictly without loss, and to compress the remaining regions of the image with loss.
The pixels appertaining to the ROI part are encoded efficiently by using a reversible DCT-based lossless
embedded zerotree coder, thus yielding an overall high compression ratio and conserving the relevant
data diagnostic. To use the zerotree encoder with DCT coefficients, these coefficients are rearranged
into a hierarchical structure similar to the wavelet subbands. We obtain better compression results,
comparing to the DWT embedded zerotree and conventional DCT that use regular quantizer.
The remainder of this paper is organized as follows: Section 2 exhibits the proposed approach for
adaptive medical image compression. Section 3 illustrates and discusses the experiments results, while
last section concludes the paper.
2. Proposed Approach
We remind that the objective of the proposed approach is to make medical image coding more
efficient by preserving the content of the ROI part used in all diagnosis treatment. The proposed
method is based on the EZW algorithm, which is modified and combined to a reversible DCT (RDCT),
to improve its performance for lossless compression. The background is highly compressed with the
SPIHIT method. In the following paragraphs we will briefly present the basic tools having been used
and then we will detail our proposed compression system.
The RDCT has a good energy compaction property and gives better results for image compression by
providing an excellent rate-distortion [23-25]. Therefore we propose to use RDCT as a substitute for the
DCT with a lossless embedded compression to encode the ROI part of image without loss of decoding
accuracy.
Integer DCT performs as well as the DCT by employing lifting structures and rounding operations,
which convert signals from real values to integer values [26,27].
The lifting-structured method maps an integer input vector to also an integer output vector as done
for integer wavelet transform [28]. The integer approximation is achieved by applying a rounding
function after each addition. Various integer DCTs have been greatly innovated for lossy-to-lossless
Sid Ahmed Elhannachi, Nacéra Benamrane, and Taleb-Ahmed Abdelmalik
J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017 | 43
image coding [29-32]. The authors in [33] present a comparative study on many kinds of Int-DCT, they
referred to five algorithms of integer DCT type II (namely, the BinDCT-IIC, -IIL, -IIS, IntDCT-II,
LDCT-II). A comparative performance in aspect of lossless coding showed that the first order entropy
rate of LDCT-II is the best and confirmed that the LDCT-II is the best for lossless coding performance
Therefore, in our approach we opted for this type of lossless DCT (LDCT-II) [23].
For embedded zerotree methods, a zerotree data structure is defined by using the self-similarity of
wavelet transform sub-image at different scales. There are two main techniques, EZW and SPIHT.
In the EZW algorithm, two passes are used to code the image: a dominant pass and a subordinate
pass. The dominant pass finds pixel values above a certain threshold, and the subordinate pass refines
all significant coefficients found in the dominant pass [34].
In the dominant pass, a wavelet coefficient x is said to be insignificant with respect to a given
threshold T if: |x| <T. In this case, once all wavelet coefficients of the same orientation in the same
spatial location are insignificant, a zerotree can be formed, and it is called zerotree root (ZTR). In order
to encode a zerotree, three additional symbols are defined: if the coefficient is insignificant but has some
significant descendant, it is called isolated zero (IZ). In addition, the coefficients can be either positive
or negative significant. As a result, the wavelet coefficients can be efficiently represented as a string of
symbols from four symbol alphabet: ZTR, IZ, positive significant (POS), and negative significant
(NEG).
In the subordinate pass, a refinement bit is coded for each significant coefficient. A coefficient is
significant if it has been coded POS or NEG in a previous bit-plane. Its current refinement bit is simply
its corresponding bit in the current bit-plane. The interested reader is suggested to refer to [34,35] for
further details on the EZW coding scheme.
The SPIHT algorithm is a refined version of EZW algorithm. It can perform better at higher
compression ratios for a wide variety of images [36,37].
The main idea is based on partitioning of wavelet coefficients in three sets: when a threshold value is
given, it partitions the coefficients or the sets of coefficients into significant or insignificant coefficients.
Significant coefficients are added to the list of significant pixels (LSP), and insignificant coefficients to
the list of insignificant pixels (LIP) or the list of insignificant sets (LIS).
During the compression procedure, the sets of coefficients in the LIS are refined and, if coefficients
become significant, they are moved from the LIP to the LSP. The LSP is set as an empty list and all the
nodes in the highest subband are put into the LIP. The root nodes with descendants are put into the
LIS.
Two encoding passes which are the sorting pass and the refinement pass are then performed in the
SPIHT coder. During the sorting pass, a significance test is performed on the coefficients according to
the order in which they are stored in LIP. Elements in LIP that are found to be significant with respect
to the threshold are moved to the LSP. A significance test is then performed on the sets in the LIS. Here,
if a set in LIS is found to be significant, the set is removed from the list and is partitioned into four
single elements and a new subset. This new subset is added back to LIS and the four elements are then
tested and moved to LSP or LIP, depending on whether they are significant or insignificant with respect
to the threshold. In the refinement pass, another bit of precision is added to the magnitudes of
coefficients in the LSP. Finally, the threshold is halved and SPIHT coding is repeated until the bit
budget is exhausted [37].
Adaptive Medical Image Compression Based on Lossy and Lossless Embedded Zerotree Methods
44 | J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017
Fig. 1. Proposed compression method.
2.1 The Proposed Compression System
Fig. 1 shows the overall system flow diagram. The proposed system starts with a manual partition of
image into two parts namely: ROI and background for selective compression. Once the ROI is
extracted, a reversible DCT with lossless embedded technique are used to compress the ROI with higher
compression ratio. Also, we have chosen not to discard the background (non-ROI), but rather to highly
compress it by SPIHT technique. Finally, an entropy coding based on adaptive arithmetic coder is
applied to the final bit stream.
2.1.1 Lossless ROI coding
Rearrangement of RDCT coefficients
In our approach, the ROI is selected manually by a rectangular shape. It is decomposed into non-
overlapping 8×8 blocks, which are then transformed with LDCT-II. The LDCT-II coefficients of each
block are rearranged into 3-level wavelet pyramid structure. These rearranged coefficients are coded by
using embedded zerotree coding based on a modified EZW coder for lossless compression, followed by
entropy coding to further compress the result. The detailed descriptions are given in the following
subsections.
Consider the ROI part which is composed of K×L blocks with sizes of M×M, where each block is 2D
LDCT-II transformed. Each LDCT-II block of size M×M, including M2 coefficients is treated as a
hierarchical subband structure. Rearranging all blocks of the ROI part in this way, we obtain a 3-scale
hierarchical subband structure.
Arithmetic coding
Lossless EZW
Reorganized
coefficients
Reversible DCT
SPIHT
ROI
ROI selection
Original image
Compressed image
Yes No
Lossless coding
Lossy coding
Sid Ahmed Elhannachi, Nacéra Benamrane, and Taleb-Ahmed Abdelmalik
J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017 | 45
In our approach, for all the DCT blocks and according to their order, we adopt a method to collect
the coefficients of the same frequency in order to take full advantage of this relationship between DCT
and subband decompositions. For each block of size of 8×8, we apply the relation illustrated in the
following equation:
]][[]88)8%][(88)8%[( jiDjNjiMiW
(2)
Where:
D[M][N]: the matrix of DCT coefficients
W[M][N] : the matrix of rearranged coefficients
I=1,2,...,M and J=1,2,...,N
% denotes modulus operator
(a) (b)
Fig. 2. (a) 4×4 DCT block, (b) reordered DCT coefficients into subband structure.
From Fig. 2, observing the energy distribution among the DCT blocks, the energy compaction is
presented in the multi-resolution dimension in so that the largest energy is located at the first subband
LL3that includes DC coefficients of all 8×8 DCT blocks and the energy declines monotonically in the
subbands in lower scale that contains the alternative current (AC) coefficients. Since most of the energy
is concentrated in the direct current (DC) coefficients, the quality of the decoded image depends mostly
upon DCcoefficients then on the AC coefficients. Therefore the rearranged integer DCT having the
same structure of multi-resolution wavelet subbands is suitable for the zerotree encoding that allow
more efficient coding.
The proposed lossless EZW (LEZW)
We propose a simple coder that yields perfect quality reconstruction at good compression rates. It
performs an embedded coding, based on reversible DCT. The lossless embedded zerotree coding of
LDCT-II coefficients is based on three steps: 1) determining the significant coefficients across scales by
exploiting the self-similarity inherent in reorganized LDCT-II coefficients, 2) successive differences
quantizing of LDCT-II coefficients, 3) lossless compressing of the data from the output of the
embedded zerotree coder by using an adaptive arithmetic coder [38].
Therefore, we propose some modifications to simplify the conventional EZW algorithm for a lossless
coding and improved compression ratio. The modification will change the passes used in the coding
algorithm.
In preceding paragraphs, it mentioned that the EZW algorithm used two passes: dominant pass and
subordinate pass. The dominant pass looks for the significant coefficients for each bitplane, and the
Adaptive Medical Image Compression Based on Lossy and Lossless Embedded Zerotree Methods
46 | J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017
subordinate then quantizes these found significant coefficients. For lossless compression, the results
from two passes must be stored for decoding. Therefore, the resulting compression ratio is a sum of the
bit rates from the dominant and subordinate passes. We note that the subordinate pass contributes
averagely one-third of the total bit rate.
So, in order to save more of the bit rate (improving compression ratio), we consider removing the
subordinate pass. Unlike Shapiro’s EZW, the subordinate pass is eliminated by coding residual values
obtained per successive differences (i.e., subtracting the subordinate threshold from the coefficients in
the subordinate list) to provide an exact representation of the coefficients for each subordinate pass.
The coefficient Cij whose absolute value is larger than the current threshold, is recognized as a
significant pixel and replaced by a residual value in the image. This is an alternative way to replace the
subordinate pass, but still complete the job of quantizing coefficients. The dominant pass can thus be
implemented as (LIP; Sn(Des(i,j)), significance of descendants):
1) Set initial threshold T=2 [log2(MAX(|Cij|))]
2) For each current threshold Tk≥ 2 do
For each coefficient Cij
LIPdo
If |Cij| <Tk Then
Output ‘0’;
Remove Cij from LIP;
Else
Output ‘1’ and Set Dij=|Cij| - Tk;
Remove Cij from LIP;
End If;
If (Sn(Des(i,j))==1) Then
Add Children of Cij to LIP;
Output ‘1’;
Else Output ‘0’;
End If;
3) Set Tk+1=Tk/2 and go to step (2)
The LEZW lossless algorithm continues until the last threshold is equal to 2, at which the residual
image only contains binary values. We note that, in this lossless coding method, the algorithm forces
previously scanned significant coefficients being coded and appended to the dominant list only once
during the repeated scanning processes. The final bit stream is stored as a byte stream to shorten its
length for adaptive arithmetic coding [38].
2.1.2 Lossy background coding
In order to improve the global compression rates, the background which contains no useful
information to establish a medical diagnosis, was highly compressed by using the wavelet zerotree
image coding technique (SPIHT) developed by Said and Pearlman [36]. This technique provides a high
performance in image compression with low complexity.
Sid Ahmed Elhannachi, Nacéra Benamrane, and Taleb-Ahmed Abdelmalik
J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017 | 47
3. Experimental Results
We conducted a number of experiments to evaluate the performance of the proposed scheme, on
various kinds of medical images including: MRI, X-rays, CT, and ultrasounds images givenin Figs. 3–8.
(a) (b) (c)
Fig. 3. Original test images with size of 256×256 and 8 bit per pixel: (a): a brain MRI, (b) a lung X-ray,
and (c) a pancreas CT image.
We use PSNR (in dB) and correlation coefficient (CoC) as metrics for whole image quality evaluation
and bit rate (bit per pixel, bpp) for lossless compression assessment. The mean square error (MSE) is
used to calculate the PSNR that is given in the equation 4.
2
1
0
1
0
)),(),((
1yxîyxi
NM
MSE
M
x
N
y
(3)
MSE
LogPSNR 255255
10 10
(4)
With M×N, image size; i, original image; î, reconstructed image.
The CoC suggests how closely the reconstructed image is correlated with an original image, on a scale
of 0–1:
1
0
1
0
1
0
1
0
22
1
0
1
0
),(),(
),(),(
M
x
N
y
M
x
N
y
M
x
N
y
yxîyxi
yxîyxi
CoC (5)
The coding results were compared with those obtained applying the original SPIHT algorithm and
the standards JPEG [39] and JPEG2000 [40] on the whole image. The SPIHT algorithm and JPEG2000
had been utilized with the irreversible 9/7 biorthogonal filters [41].
Adaptive Medical Image Compression Based on Lossy and Lossless Embedded Zerotree Methods
48 | J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017
(a) (b)
(c) (d)
(e) (f)
Fig. 4. Visual compression results for brain MRI image at bit rate=0.5 bpp. (a) Original image, (b) ROI
selection, (c) ROI encoding—lossless encoding of image background and ROI losslessly coded (PSNR=33.60
dB), (d) image difference between (a) and (c), (e) entire image encoding by SPIHT (PSNR=29.30 dB), and
(f) image difference between (a) and (e).
In our scheme, the ROI region is selected by a rectangular shape for each test image, as shown in Figs.
4(b) and 5(b). Thus, information to be transferred to the decoders is its corner coordinates or its origin
and radius. The ROI region is coded losslessly by using a reversible integer DCT transform with the
modified algorithm LEZW and the background is coded with loss at low bit-rate by SPIHT using 9/7
Daubechies wavelet filter, with 4 levels of decomposition in order to increase the compression ratio.
Perceptually the gains are very significant in term of diagnostic information conservation and
compression ratio, as illustrated in the following.
Sid Ahmed Elhannachi, Nacéra Benamrane, and Taleb-Ahmed Abdelmalik
J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017 | 49
(a) (b)
(c) (d)
(e) (f)
Fig. 5. Visual compression results for lung X-ray image at bit rate=1.0 bpp. (a) Original image, (b) ROI
selection, (c) ROI encoding—lossless encoding of image background and ROI losslessly coded
(PSNR=36.97 dB), (d) image difference between (a) and (c), (e) entire image encoding by SPIHT
(PSNR=30.81 dB), and (f) image difference between (a) and (e).
Figs. 4 and 5 show a comparison results of a brain MRI image and lung X-Ray image, respectively,
compressed by proposed method (based ROI) and regular compression method by SPIHT (not based
ROI), at the same rates. Figs. 4(b) and 5(b), show the separated parts of the original image, namely the
ROI and the background.
In Figs. 4(c) and 5(c), both regions are compressed and merged and formed the compressed images.
Figs. 4(e) and 5(e) show the reconstructed images obtained by a regular compression.
Note that the value of PSNR in our method is higher than those obtained by ordinary SPIHT
Adaptive Medical Image Compression Based on Lossy and Lossless Embedded Zerotree Methods
50 | J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017
compression method, this is due to the lossless compression applied to the ROI part, whereas the
background is almost useless in diagnosis, therefore it is compressed with higher compression rate.
So we managed to fully preserve the ROI in similar bit rates to the regular compression where
distortions are visible in reconstructed image by the SPIHT method as seen in Figs. 4(f) and 5(f),
contrary to our proposed method that keeps all the critical information as we can distinguish in Figs.
4(d) and 5(d), where the difference is equal to zero at the ROI portion required by medical care
professionals.
Table 1. Comparison of coding results for a brain MRI image
Bit rate
(bpp)
PSNR CoC
JPEG SPIHT JPEG2000 Proposed JPEG SPIHT JPEG2000 Proposed
0.125 20.12 25.43 28.32 30.50 0.883 0.910 0.925 0.975
0.25 22.05 27.11 30.98 31.20 0.885 0.914 0.928 0.978
0.5 23.85 29.30 32.27 33.60 0.890 0.922 0.933 0.982
1 26.64 30.94 35.71 35.26 0.897 0.925 0.937 0.986
Table 2. Comparison of coding results for a pancreas CT image
Bit rate
(bpp)
PSNR CoC
JPEG SPIHT JPEG2000 Proposed JPEG SPIHT JPEG2000 Proposed
0.125 21.25 25.40 30,30
31.17 0.884 0.915 0.926 0.977
0.25 22.60 27.07 32.41 33.32 0.888 0.917 0.931 0.981
0.5 24.09 29.30 33.20
34.90 0.893 0.923 0.935 0.984
1 26.64 31.94 35.56 37.25 0.905 0.928 0.938 0.989
Table 3. Comparison of coding results for a lung X-ray
Bit rate
(bpp)
PSNR CoC
JPEG SPIHT JPEG2000 Proposed JPEG SPIHT JPEG2000 Proposed
0.125 20.20 26.03 29.09
29.60 0.883 0.913 0.923 0.974
0.25 21.50 28.10 31.05 32.68 0.887 0.916 0.927 0.979
0.5 23.07 29.70
34.21 33.02 0.890 0.923 0.933 0.984
1 26.33 30.81 36.08 36.97 0.900 0.930 0.940 0.990
In order to compare our method with other conventional compression methods such as: JPEG [36],
SPIHT [36], JPEG2000 [40], Tables 1–3 summarize the coding results in terms of PSNR and CoC,
obtained on the three test images presented in Fig. 3, at bit rates ranging from 0.125 to 1.0 bpp. On
average, notice the gain of close to 10 dB of our ROI-based coder over JPEG coder. Also, on average, for
the three test images, our technique improves, respectively, general SPIHT and JPEG2000 by 5 dB and 1
dB.
In fact, the distortion achieved at 1.0 bpp with SPIHT method, corresponds to a compression rate of
0.125 bpp with proposed approach. Compared to JPEG2000, for the brain MRI image and pancreas CT
image, the distortion achieved in 0.5 corresponds to a compression rate of 0.25 bpp with our approach.
The JPEG2000 method outperforms our proposed method only in the case of the X-ray image at bit
Sid Ahmed Elhannachi, Nacéra Benamrane, and Taleb-Ahmed Abdelmalik
J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017 | 51
rate=0.5. This proving clearly the improvement obtained with our ROI based compression technique
that guaranties an absolute conservation of the ROI part to make the diagnostic details significantly
legible without lowering the compression ratio.
So, in general, our region-based coding method performed substantially better than JPEG, SPIHT,
and JPEG2000 (applied to the whole image, with 9/7 irreversible Daubechies filters). The reported CoC
of these methods (JPEG, SPIHT and JPEG2000) compared with our method, in Tables 1–3, confirm this
conclusion. It is clear that the proposed method yields better CoC value, owing to lossless compression
on the ROI part where a perfect reconstruction was obtained.
Fig. 6. A comparativeevaluation of average PSNR for JPEG, SPIHT, JPEG2000, and our proposed approach.
Fig. 7. A comparative evaluation of average CoC for JPEG, SPIHT, JPEG2000, and our proposed approach.
Plots of average PSNR and CoC parameters versus different bit rates are respectively given in Figs. 6
and 7. The climb of PSNR versus rate for our proposed scheme is better compared to JPEG, SPIHT, and
JPEG2000 methods.
On the other hand, Fig. 7 plots the CoC measurements that show the high difference in image quality
reconstruction performance with the reversible compression on the ROI part, confirming the improvement
obtained with our scheme over conventional techniques.
So, we can deduce that our method’s advantage over all three reported methods is more evident as the
compression ratio increases since it allows a perfect reconstruction of ROI and slightly degraded
background for the entire image at relatively low bit rates.
Adaptive Medical Image Compression Based on Lossy and Lossless Embedded Zerotree Methods
52 | J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017
Fig. 8. Original ultrasound image with size of 512×512.
Table 4. Evaluation of lossless ROI coding on ultrasound image
Lossless method Bit rate (bpp)
JPEG-LS 3.25
JPEG2000 1.33
Reference [43] 1.44
Proposed 1.36
Moreover, to show the effectiveness of our proposed technique for lossless ROI coding (RDCT+
LEZW), we have considered the ultrasound image, sized 512×512 pixels, presented in Fig. 8. Table 4
enumerates bit rate results (bit per pixel, bpp) for lossless coding of ROI part of the ultrasound image.
Compared with JPEG-LS [42], and results presented in [43], we can see that the bit rate required by our
method for fully reconstruction of ROI is the minimum (1.36 bpp). Yet, lossless JPEG2000 technique
with reversible 5/3 wavelet filter [44], provides a very slight bit rate advantage over our approach.
Table 5. Comparison of PSNR (dB) for different ROI based methods
Bit rate (bpp) EBCOT MaxShift Reference [43] Proposed ROI coder
0.125 24.95 30.08 28.17 31.14
0.25 27.69 33.48 32.58 34.97
0.5 31.52 38.91 38.09 39.58
1 37.77 47.71 37.77 45.07
To more justify our results, we have evaluated the proposed method against the other ROI based
methods namely: EBCOT[45], MaxShift [46] and the work presented in [43]. Table 5 lists the PSNR
values at different bit rates for the ultrasound image. Compared to the two methods EBCOT and that
presented in [43], PSNR values reveal that reconstructed image quality for the proposed method is
Sid Ahmed Elhannachi, Nacéra Benamrane, and Taleb-Ahmed Abdelmalik
J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017 | 53
much better. These obtained results are very competitive to those obtained by MaxShift method.
The obtained results by our approach show that with the proposed lossless compression technique
applied on ROI part, compression results are satisfactory compared to the stat-of-art compression
methods based ROI.
In addition, applied on MRI brain image, our approach compared to the both based ROI approaches
presented respectively in [10,11], is clearly superior in term of PSNR. At a bit rate equal to 1 bpp, with
our approach we obtain 45.61 dB of PSNR which is higher to 36.10 dB and 40.91 dB, respectively,
obtained by [10,11]. Thus, our approach brings a high compression rate maintaining faultless quality of
the ROI while the quality of the background is allowed to have degraded quality, since it is considered
to be less important.
4. Conclusion
In this work, to provide solutions for efficient region based image compression, we presented an
embedded lossless wavelet-based image coding algorithm based on successive differences, which used
only the subordinate pass. Applied on the ROI part, the proposed method is simple in concept and
implementation but achieving promising lossless compression efficiency as compared with some
reversible methods. Irreversible lifting wavelet transform with SPIHT coder was used for compressing
the background.
As we have seen in section of experimental results, the proposed reversible system has not limited
efficiency in terms of compression rates and provides a solution to increasingly important conservation
of diagnostically relevant information. So, our proposed compression scheme can provide efficient
compression for various medical images and offer potential advantages in yielding much higher
compression rates, while maintaining the integrity of data in ROI part, combining therefore the
advantages from both DCT and wavelet based compression algorithms. The choice of integer
transformations is an important aspect of the lossless compression.
Medical images, by their inherent features, are more sensitive to distortion and nonlinear
transformation and the definition of ROI varies greatly between different categories. Identifying and
extracting the region of interest accurately is very important before coding and compressing the image
data for efficient transmission or storage thus further investigation is required to automatically detect
the ROI.
References
[1] M. Haindl and S. Mikes, “Texture segmentation benchmark,” in Proceedings of the 19th International Conference
on Pattern Recognition, Tampa, FL, 2008, pp. 1-4.
[2] D. Martin, C. Fowlkes, D. Tal, and J. Malik, “A database of human segmented natural images and its
application to evaluating segmentation algorithms and measuring ecological statistics,” in Proceedings of the 8th
International Conference on Computer Vision, Sydney, Australia, 2001, pp. 416-423.
[3] H. Lin, J. Si, and G. P. Abousleman, “Knowledge-based hierarchical region-of-interest detection,” in Proceedings
of the International Conference on Acoustics, Speech and Signal Processing, Orlando, FL, 2002, pp. 13-17.
Adaptive Medical Image Compression Based on Lossy and Lossless Embedded Zerotree Methods
54 | J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017
[4] G. P. Nguyen and M. Worring, “An user based framework for salient detail extraction,” in Proceedings of the
IEEE International Conference on Multimedia and Expo, Taipei, Taiwan, 2004, pp. 1831-1834.
[5] JPEG2000 Part I Final Committee draft version 1.0 (ISO/IEC JTC 1/SC 29/WG 1 (ITU-T SG8) [Online].
Available: https://www.ics.uci.edu/~dan/class/267/papers/jpeg2000.pdf.
[6] H. Min, C. Zhang, J. Lu, and B. Zhou, “A multi ROIs medical image compression with edge feature preserving,”
in Proceedings of the 3rd International Conference on Intelligent System and Knowledge Engineering, Beijing,
China, 2008, pp. 1075-1080.
[7] M. Firoozbakht, J. Dehmeshki, M. Martini, Y. Ebrahimdoost, H. Amin, M. Dehkordi, A. Youannic, and S. D.
Qanadli, “Compression of digital medical images based on multiple regions of interest,” in Proceedings of the
4th International Conference on Digital Society, St. Maarten, Netherlands Antilles, 2010, pp. 260-263.
[8] R. F. Larico,Y. Iano, R. S. Higa, R. Arthur, and O. Saotome, “Generalized region of interest coding applied to
SPIHT,” Cyber Journals: Multidisciplinary Journals in Science and Technology, Journal of Selected Areas in
Telecommunications (JSAT), vol. 2012, pp. 23-31, 2012.
[9] N. D. Londhe and S. Chawre, “Region based coding of liver cancer CT images,” Canadian Journal on Image
Processing and Computer Vision, vol. 4, no. 1, pp. 9-15, 2013.
[10] D. V. Babu and N. R. Alamelu, “Wavelet based medical image compression using ROI EZW,” In tern atio nal
Journal of Recent Trends in Engineering, vol. 1, no. 3, pp. 97-100, 2009.
[11] H. Jiang, Y. Zhang, S. Shi, Z. Ma, A. Dong, Q. Tong, B. Yang, and L. Zhang, “An image ROI compression
algorithm based on hybrid fractal model,” Journal of Information & Computational Science, vol. 11, no. 4, pp.
1201-1208, 2014.
[12] T. M. Rajkumar and M. V. Latte, “An efficient ROI encoding based on LSK and fractal image compression,” The
International Arab Journal of Information Technology, vol. 12, no. 3, pp. 220-228, 2015.
[13] J. Mielikainen and P. Toivanen, “Clustered DPCM for the lossless compression of hyperspectral images,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 41, no. 12, pp. 2943-2946, 2003.
[14] X. Wu and N. Memon, “Context-based lossless interband compression-extending CALIC,” IEEE Transactions
on Image Processing, vol. 9, no. 6, pp. 994-1001, 2000.
[15] E. Magli, G. Olmo, and E. Quacchio, “Optimized onboard lossless and near-lossless compression of
hyperspectral data using CALIC,” IEEE Geoscience and Remote Sensing Letters, vol. 1, no. 1, pp. 21-25, 2004.
[16] A. Ouafi, A. T. Ahmed, Z. E. Baarir, and A. Zitouni, “A modified embedded zerotree wavelet (MEZW)
algorithm for image compression,” Journal of Mathematical Imaging and Vision, vol. 30, no. 3, pp. 298-307, 2008.
[17] N. M. A. Ahmed and A. M. A. Brifcani, “A new modified embedded zerotree wavelet approach for image
coding (NMEZW),” International Journal of Scientific & Engineering Research, vol. 4, no. 3, pp. 1-11, 2013.
[18] A. Sattar, “The medical image compression with embedded zerotree wavelet,” International Journal of Science,
Spirituality, Business and Technology, vol. 1, no. 2, pp. 67-70, 2013.
[19] Z. Xiong, O. G. Guleryuz, and M. T. Orchard, “A DCT-based embedded image coder,” IEEE Signal Processing
Letters, vol. 3, no. 11, pp. 289-290, 1996.
[20] D. M. Monro and G. J. Dickson, “Zerotree Coding of DCT coefficients,” in Proceedings of the IEEE International
Conference on Image Processing, Santa Barbara, CA, 1997, pp. 625-628.
[21] Y. A. Jeong and C. K. Cheong, “A DCT-based embedded image coder using wavelet structure of DCT for very
low bit rate video codec,” IEEE Transactions on Consumer Electronics, vol. 44, no. 3, pp. 500-507, 1998.
[22] C. K. Cheong, K. S. Cho, and S. W. Lee, “Significance tree image sequence coding with DCT-based pyramid
structure,” in Proceedings of the IEEE International Conference on Image Processing, Vancouver, Canada, 2000,
pp. 859-862.
[23] V. Britanak, P. C. Yip, and K. R. Rao, Discrete Cosine and Sine Transforms: General Properties, Fast Algorithms
and Integer Approximations. New York: Academic Press, 2007.
[24] K. Komatsu and K. Sezaki, “Reversible discrete cosine transform,” in Proceedings of the International Conference
on Acoustic, Speech, Signal Processing, Seattle, WA, 1998, pp. 1769-1772.
Sid Ahmed Elhannachi, Nacéra Benamrane, and Taleb-Ahmed Abdelmalik
J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017 | 55
[25] Z. He and M. Bystrom, “Improved conversion from DCT blocks to integer cosine transform blocks in
H.264/AVC,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 18, no. 6, pp. 851-857, 2008.
[26] G. Plonka and M. Tasche, “Invertible integer DCT algorithms,” Applied and Computational Harmonic Analysis,
vol. 15, no. 1, pp. 70-88, 2003.
[27] J. Liang and T. D. Tran, “Fast multiplierless approximations of the DCT with the lifting scheme,” IEEE
Transactions on Signal Processing, vol. 49, no. 12, pp. 3032-3044, 2001.
[28] W. Sweldens, “The lifting scheme: a custom-design construction of biorthogonal wavelets,” Applied and
Computational Harmonic Analysis, vol. 3, no. 2, pp. 186-200, 1996.
[29] S. Fukuma, K. Ohyama, M. Iwahashi, and N. Kambayashi, “Lossless 8-point fast discrete cosine transform
using lossless Hadamard transform,” Technical Report No. IEICE-DSP-99-103, 1999.
[30] Y. J. Chen, S. Oranintara, and T. Nguyen, “Integer discrete cosine transform (IntDCT),” in Proceedings of the
2nd International Conference on Information, Communications and Signal Processing, Singapore, 1999.
[31] T. D. Tran, “The BinDCT: fast multiplierless approximation of the DCT,” IEEE Signal Processing Letters, vol. 7,
no. 6, pp. 141-144, 2000.
[32] J. Liang and T. D. Tran, “Fast multiplierless approximations of the DCT with the lifting scheme,” IEEE
Transactions on Signal Processing, vol. 49, no. 12, pp. 3032-3044, 2001.
[33] S. Udomsiri and M. Iwahashi, “Comparative study on recent integer DCTs,” World Academy of Science,
Engineering and Technology, International Scholarly and Scientific Research & Innovation, vol. 2, no. 12, pp.
2656-2662, 2008.
[34] J. M. Shapiro, “Embedded image coding using zerotrees of wavelet coefficients,” IEEE Transactions on Signal
Processing, vol. 41, no. 12, pp. 3445-3462, 1993.
[35] V. J. Rehna and M. K. Jeya Kumar, “Wavelet based image coding schemes: a recent survey,” International Journal
on Soft Computing, vol. 3, no. 3, pp. 101-118, 2012.
[36] A. Said and W. A. Pearlman, “A new fast efficient image codec based on set partitioning in hierarchical trees,”
IEEE Transactions on Circuits and System Video Technology, vol. 6, no. 3, pp. 243-250, 1996.
[37] M. D. Baylon and S. J. Lim, “Transform/subband analysis and synthesis of signals,” Research Laboratory of
Electronics, Massachusetts Institute of Technology, Cambridge, MA, Technical Report No. RLE-559, 1990.
[38] T. V. Ramabadran and K. Chen, “Efficient compression of medical images through arithmetic coding,” in
Proceedings of SPIE 1234: Medical Imaging. Bellingham, WA: International Society for Optics and Photonics,
1990, pp. 761-775.
[39] A. Kaluanpur, V. P. Neklesa, C. R. Taylor, A. R. Daftary, and J. A. Brink, “Evaluation of JPEG and wavelet
compression of body CT images for direct digital teleradiologic transmission,” Radiology, vol. 217, no. 3, pp.
772-779, 2000.
[40] J. Askelof, M. L. Carlander, and C. Christopoulos, “Regions of interest coding in JPEG2000,” Signal Processing:
Image Communication, vol. 17, no. 1, pp. 105-111, 2002.
[41] M. Antonini, M. Barlaud, P. Mathieu, and I. Daubechies, “Image coding using wavelet transform,” IEEE
Transactions on Image Processing, vol. 1, no. 2, pp. 205-220, 1992.
[42] M. J. Weinberger, G. Seroussi, and G. Sapiro, “From LOCO-I to the JPEG-LS standard,” in Proceedings of the
International Conference on Image Processing, Kobe, Japan, 1992, pp. 68-72.
[43] H. Yang, M. Long, and H. M. Tai, “Region-of-interest image coding based on EBCOT,” IEEE Proceedings-Visual
Image Signal Process, vol. 152, no. 5, pp. 590-596, 2005.
[44] A. Skodras, C. Christopoulos, and T. Ebrahimi, “The JPEG2000 still image compression standard,” IEEE Signal
Processing Magazine, vol. 18, no. 5, pp. 36-58, 2001.
[45] D. Taubman, “EBCOT: embedded block coding with optimized truncation,” Technical Report N1020R,
ISO/IEC JTC1/SC29/WG1, 1998.
[46] P. G. Tahoces, J. R. Varela, M. J. Lado, and M. Souto, “Image compression: maxshift ROI encoding options in
JPEG2000,” Computer Vision and Image Understanding, vol. 109, no. 2, pp. 139-145, 2008.
Adaptive Medical Image Compression Based on Lossy and Lossless Embedded Zerotree Methods
56 | J Inf Process Syst, Vol.13, No.1, pp.40~56, February 2017
Sid Ahmed El Hannachi
He received his engineering diploma in 2002 in computer science from USTO-MB
University. In 2005, he received his Magister, in department of computer science in
USTO-MB University. He is lecturer and PhD candidate atthe same department. His
current research interests include pattern recognition and medical image coding.
Nacéra Benamrane
She received her engineering degree in Computer Science from University of Oran,
the M.Sc. and Ph.D. degrees from University of Valenciennes, France, in 1988 and
1994. Since 2002, she is the head of vision and medical imaging group at SIMPA
Laboratory. She is currently a professor in informatics departement at University of
Science and Technology of Oran-Mohamed Boudiaf (USTO-MB). She has published
more than 80 papers in journals and conference proceedings. Her main research
interests include image processing, medical imaging, computer vision and pattern
recognition.
Taleb-Ahmed Abdelmalik
He received Ph.D. degree in 1992 from the University the Sciences and technology of
Lille 1 and the HDR degree in 2003 from the university côte d'opale at Calais. He is
professor in the IUT University of Valenciennes and member in LAMIH laboratory.
His researches include image processing and data learning.
