SPECIAL SECTION ON DEEP LEARNING FOR COMPUTER-AIDED MEDICAL DIAGNOSIS

Received April 15, 2019, accepted May 6, 2019, date of publication May 15, 2019, date of current version May 24, 2019.
Digital Object Identifier 10.1109/ACCESS.2019.2916934

A Fast Fractal Based Compression for MRI Images

SHUAI LIU 1, WEILING BAI2 , NIANYIN ZENG 3, AND SHUIHUA WANG 4
1 Hunan Provincial Key Laboratory of Intelligent Computing and Language Information Processing, College of Information Science and Engineering, Hunan
Normal University, Changsha 410000, China
2 College of Computer Science, Inner Mongolia University, Hohhot 010010, China
3 Department of Instrumental and Electrical Engineering, Xiamen University, Xiamen 361000, China
4 School of Architecture Building and Civil Engineering, Loughborough University, Loughborough LE11 3TU, U.K.

Corresponding authors: Nianyin Zeng ([email protected]) and Shuihua Wang ([email protected])

This work was supported in part by the National Natural Science Foundation of China under Grant 61502254, in part by the Program for
Yong Talents of Science and Technology in Universities of Inner Mongolia Autonomous Region under Grant NJYT-18-B10, in part by the
Open Project Program of the State Key Lab of CAD&CG under Grant A1926, Zhejiang University, and in part by the Open Project
Program of the State Key Laboratory of Complex Electromagnetic Environment Effects on Electronics and Information System under
Grant 2019K0104B.

ABSTRACT Magnetic resonance imaging (MRI), which assists doctors in determining clinical staging and
expected surgical range, has high medical value. A large number of MRI images require a large amount of
storage space and the transmission bandwidth of the PACS system in offline storage and remote diagnosis.
Therefore, high-quality compression of MRI images is very research-oriented. Current compression methods
for MRI images with high compression ratio cause loss of information on lesions, leading to misdiagnosis;
compression methods for MRI images with low compression ratio does not achieve the desired effect.
Therefore, a fast fractal-based compression algorithm for MRI images is proposed in this paper. First,
three-dimensional (3D) MRI images are converted into a two-dimensional (2D) image sequence, which
facilitates the image sequence based on the fractal compression method. Then, range and domain blocks
are classified according to the inherent spatiotemporal similarity of 3D objects. By using self-similarity,
the number of blocks in the matching pool is reduced to improve the matching speed of the proposed method.
Finally, a residual compensation mechanism is introduced to achieve compression of MRI images with high
decompression quality. The experimental results show that compression speed is improved by 2–3 times,
and the PSNR is improved by nearly 10. It indicates the proposed algorithm is effective and solves the
contradiction between high compression ratio and high quality of MRI medical images.

INDEX TERMS MRI, image compression, fractal compression, spatiotemporal similarity, lossy compres-
sion.

I. INTRODUCTION conditions are difficult to meet the real-time transmission

Medical imaging has become one of the most active and requirements of large data volumes. To meet effective storage
rapidly evolving fields in medical research and clinical and transmission of medical images, it is not only necessary
diagnosis. Medical images display the internal structure of to expand storage space and transmission bandwidth, but also
the human body in an intuitive form, providing clinicians to study how to efficiently compress medical data. Therefore,
with intuitive and accurate basic information on anatomy, it is necessary to implement effective compression of various
pathology and function. Typical imaging modalities include medical images using an image compression algorithm.
magnetic resonance imaging (MRI), computer-assisted Medical image compression methods are generally classi-
tomography (CT), ultrasound (US), computer-assisted X-ray fied into lossless compression and lossy compression. Loss-
(CR), digital subtraction angiography (DSA), et al. With less compression provides medical diagnostics with image
continuous advancement of medical imaging technology, information of the same quality as the original image. How-
especially the resolution of imaging devices, the amount of ever, the compression ratio of lossless compression is usu-
medical image data will continue to grow. Existing bandwidth ally low, which is difficult to meet the actual transmission
requirements of medical images. Lossy compression pro-
The associate editor coordinating the review of this manuscript and vides a higher compression ratio by losing some information.
approving it for publication was Yudong Zhang. Increase in compression inevitably brings a certain degree of
2169-3536 2019 IEEE. Translations and content mining are permitted for academic research only.
62412 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 7, 2019
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
S. Liu et al.: Fast Fractal Based Compression for MRI Images

approximate lossless compression of MRI images. The exper-

imental results show that the proposed method improves
compression efficiency under the premise of ensuring image
quality of MRI.
The rest of this paper is organized as follows. In Section 2,
we review the progress of medical image compression
research from both ROI coding and non-ROI coding.
Section 3 elaborates on a fast fractal-based compression
method for MRI image. In Section 4, we verify feasibil-
ity of the proposed method through experiments. Finally,
Section 5 summarizes our work and describes future research
directions.

II. RELATED WORKS

FIGURE 1. Spatiotemporal correlation of MRI medical images.
Under application conditions such as telemedicine, the com-
pression ratio of lossless compression is difficult to meet
degradation to medical images. In the future of telemedicine transmission requirements for real-time images. It is neces-
applications, lossless compression will be difficult to provide sary to resort to a lossy compression method. At the expense
a low bit rate required for image transmission, while relying of image portion information, a target bit rate required for
on lossy compression to achieve real-time transmission of real-time transmission is exchanged. Selective image com-
images. At present, lossy compression technology for med- pression technology is gaining an increasingly important
ical images has become a research hotspot at domestic and role in telemedicine or medical imaging applications with
international, and its research goal is to improve the recon- large storage needs. Its production effectively alleviates the
struction quality of images as much as possible under a given contradiction between high compression ratio and lossless
code rate. compression of medical images. The compression method
MRI medical images contain rich temporal and spa- based on region of interest (ROI) has become an important
tial correlations. As shown in Fig. 1, three similar image selective image compression technology, which accounts for
blocks 1, 2, and 3 are respectively derived from three adjacent a large proportion in published literatures. This paper will
MRI medical image slices. The spatiotemporal correlation review the research progress of medical image compression
means that in a MRI image slice, its adjacent fields of a from both ROI coding and non-ROI coding.
certain pixel (region) in a layer of image (several layers
before and after) have a strong similar relationship with A. ROI CODING
the pixel (region). This spatiotemporal correlation makes a Medical image compression algorithm based on ROI cod-
large amount of (local) self-similar information contained in ing first divides an image into ROI and background (BG),
MRI images. Therefore, it is considered to compress this wherein the ROI region usually adopts lossless or near loss-
type of medical image using the fractal compression idea. less compression, and the BG region usually performs a
Fractal image compression is based on Iterated Function large degree of lossy compression. Bruckmann and Uhl [1]
System (IFS), omitting the image content, and only retain- compared the performance of lossy compression based on
ing self-similarity parameters of the local image content to wavelet transform and JPEG in selective image compres-
complete data compression. It has advantages of high com- sion techniques. Experimental results show that the selec-
pression ratio, reconstruction at any scale, and fast decoding. tive compression performance of wavelet transform was
However, traditional fractal compression algorithm fails to better than JPEG. Based on the ROI coding provided by
fully consider the spatiotemporal relationship of MRI images, JPEG2000, Tahoces et al. [2] proposed a selective coefficient
resulting in inefficient and poor results. Therefore, a fast mask displacement coding algorithm. The wavelet coeffi-
fractal-based compression algorithm for MRI images is pro- cients belonging to different sub-bands were shifted to imple-
posed in this paper. ment ROI coding. Zhang et al. [3] proposed a classification
In this study, our contribution is to apply sequence method based on neural network. Hosseini and Naghsh-
image-based fractal compression method to compress Nilchi [4] proposed context-based vector quantization (CVQ)
three-dimensional MRI images. Besides, to get better per- algorithm to achieve high-fidelity compression of medical
formance of compression, we proposed two improvements. ultrasound images. The algorithm used the region growing
First, range blocks and domain blocks are classified accord- method to separate ROI and BG in an image, and then used the
ing to the spatiotemporal similarity feature. A range block proposed CVQ scheme to compress the two parts separately.
only needs to search for the optimal matching block in a In the algorithm proposed by Sophia and Anitha [5], the ROI
certain type of domain block. Compression process is acceler- part of an image is losslessly compressed by run-length cod-
ated by reducing the capacity of matching pool. Second, the ing, Huffman coding or arithmetic coding, and the BG part is
residual compensation mechanism is introduced to achieve subjected to lossy compression based on vector quantization.

VOLUME 7, 2019 62413

 S. Liu et al.: Fast Fractal Based Compression for MRI Images

The algorithm proposed by Kaura and Wasson [6] used the wavelet encoder (LEZW). The background was compressed
fractal algorithm to perform lossy compression on BG, and using a set partitioning algorithm in hierarchical tree (SPIHT)
performs lossless compression based on context tree weight- technology. Yee et al. [18] applied a lossless BPG compres-
ing (CTW) on ROI. sion algorithm to the ROI region and a lossy BPG for the
Lesions in medical images and other infected sites have non-ROI region.
a more important diagnostic significance than the overall
image. Schelkens et al. [7] combined the embedded zerotree B. NON-ROI CODING
wavelet (EZW) coding algorithm with the multi-ROI gen- Sridhar and Prasad [19] combined the integer DCT-based
eral protocol to improve scalability of the embedded code SPIHT algorithm with context adaptive variable length cod-
stream. Sridhar [8] proposed a dual ROI coding algorithm for ing (CAVLC) to encode important coefficients of medical
medical images in which the lesion area was considered to images. Then, only these important coefficients were trans-
be the ROI, and the surrounding lesion (usually susceptible mitted instead of transmitting the entire image data, thereby
area) was the secondary ROI, giving the two ROIs different achieving compression effect. Based on the curled DCT
levels of priority. Hu et al. [9] proposed a multi-ROI medical (Warped DCT, WDCT), Prabhu et al. [20] proposed a 3D
image compression algorithm based on edge feature protec- warped DCT (3D WDCT), by means of which a complete
tion. The algorithm used the Canny operator to extract use- medical image compression scheme was presented. Bhavani
ful image edge information, combined JPEG2000 to reduce and Thanushkodi [21] compared the performance of several
the ROI losslessly, and used a multi-level tree set split in fractal coding algorithms on MRI compression, including
hierarchical tree (SPIHT) to compress the BG. For vascu- fractal coding standards, quasi-lossless fractal coding, and
lar images, Firoozbakht et al. [10] proposed a compression improved quasi-lossless fractal coding. Based on this, a novel
algorithm based on context and multi-ROI coding. The algo- quasi-lossless fractal compression algorithm was proposed,
rithm divided a vascular image into a primary ROI (vas- which effectively retained important features in the image.
cular stenosis region), a secondary ROI (other important A machine learning method was used to reduce encoding
areas of blood vessels), and BG. The main ROI usually time of the algorithm and improve compression performance.
needed to be manually selected, while the secondary ROI was Juliet et al. [22] proposed a medical image compression
automatically detected by the regional growth method. The algorithm based on Ripplet transform, which introduced an
component priority-based ROI coding method proposed by anisotropic Ripplet transform to represent singular points
Bartrina-Rapesta et al. [11] utilized the optimal rate distortion on an arbitrary shape curve, and encoded its important
technique and combined a simple and effective ROI code coefficients by SPIHT algorithm. Automatic and accurate
rate allocation strategy to achieve coding of multiple ROIs classification of MRI images is important for the analysis and
at different priorities. interpretation of these images. Zhang et al. [23] combined
For stereoscopic medical images, multidimensional two successful techniques: pseudo Zernike moment and ker-
wavelet transforms are typically used to remove cor- nel support vector machine to pathological brain detection.
relations in various directions in the image. Agrafio- This approach performed better than eleven state-of-the-art
tis et al. [12] extended 3D SPIHT algorithm to enable smart pathological brain detection methods in three open
three-dimensional ROI coding. Wang and Cuhadar [13] datasets.
proposed the use of unbalanced three-dimensional tree In addition, Juliet et al. [24] proposed using sparse repre-
structure to achieve three-dimensional ROI coding of sentation to explore the geometrical rules of image structure,
medical images, and achieved multi-ROI and multi-quality and based on this, proposed a medical image compression
control. Victor et al. [14] proposed an improved three- algorithm. Geometric flow indicated the direction in which
dimensional scalable compression algorithm for medical the gray level of the image changed regularly. The image
images based on the optimal volume of interest (VOI) coding. was further refined along the direction of geometric flow
Nguyen et al. [15] proposed an efficient compression algo- by Bandelet transform. Bandelet coefficients were encoded
rithm using hierarchical vector quantization and motion by SPIHT algorithm, and then the global thresholding pro-
compensation. The algorithm used three-dimensional motion cess combined with fixed coding was performed. Selvi and
estimation to create uniform pre-processed data, and used a Nadarajan [25] proposed a fast compression method for
three-dimensional compression algorithm based on hierarchi- four-dimensional fMRI images. The method utilized data
cal vector quantization to compress the pre-processed data. recombination, Contourlet transform, and improved binary
Sanchez [16] proposed a lossy compression algorithm for array technology. The test results for fMRI showed that per-
medical images based on multiple 3D ROI. The use of joint formance of the algorithm was better than SPIHT or SPECK,
source channel coding enabled multiple three-dimensional and had lower complexity. Zhang et al. [26] proposed an MRI
ROIs to achieve higher transmission priorities in the context image classifier based on Particle Swarm Optimization and
of wireless transmission. Sid Ahmed et al. [17] proposed an kernel support vector machine. Juliet et al. [27] proposed a
embedded image encoder based on efficient reversible dis- projection-based medical image compression algorithm. Dis-
crete cosine transform (RDCT). The proposed rearrangement crete radon transform (DRT) was used to effectively represent
structure was well coupled to a lossless embedded zerotree direction information of the image, and RANHT was used

62414 VOLUME 7, 2019

S. Liu et al.: Fast Fractal Based Compression for MRI Images

to encode Randon transform coefficients. The test results for

MRI and CT showed that compression performance of this
algorithm was better than other types of SPIHT algorithms.
Patbhaje et al. [28] proposed a medical image compression
technique based on adaptive scanning wavelet differential
reduction (ASWDR). ASWDR technology utilized different
wavelet filters and achieved different compression efficien-
cies based on unique sparse characteristics. Selvi and Nadara-
jan [29] proposed a fast two-dimensional lossy compression
technique using wavelet-based contourlet transform (WBCT)
and binary array technique (BAT) for computed tomogra-
phy (CT) and magnetic resonance imaging (MRI) images.
Pathak et al. [30] analyzed various prediction methods using
spatial correlation properties of neighborhood pixels to min-
imize prediction errors. Experimental results showed that the
support prediction method improved compression ratio and
maintained marginal PSNR and minimum absolute error. The
new method proposed by Amri et al. [31] combined image
reduction and extension techniques, digital watermarking and
lossless compression standards such as JPEG-LS (JLS) and
TIFF formats. Zhang et al. [32] developed a novel machine
learning system that automatically diagnoses Alzheimer’s
disease from MRI images. Parikh et al. [33] established an
FIGURE 2. Flow chart of fractal MRI image compression based on
acceptable HEVC compression range for medical imaging sequence image.
applications based on established medically acceptable JPEG
2000 compression ranges. The complexity of diagnostically again divided into domain blocks (D blocks) having a size of
acceptable lossy compression and high depth medical image 2N × 2N , and D blocks may overlap. Before the encoding,
compression was investigated. D block is averaged by the four neighboring pixels, and
From the above review to research progress of medical its size is reduced to be the same as the size of R block.
image compression, it can be seen that the ROI-based com- The averaged sampled D block is subjected to eight kinds
pression method has received extensive attention and has a of equidistant transformations, and the transformed whole
large proportion in published literatures. Purpose of introduc- constitutes codebook . For each R block, it is necessary to
ing ROI coding is to protect some of the information useful find its best matching D block in the codebook . Each R
for subsequent medical diagnosis under limited transmission block is then approximated by the luminance transform of
bandwidth conditions. Transform-based method has been its best matching block D ∈ , that is R = s • D + o • 1.
widely used in medical image compression. The currently Where 1 is a unit matrix of N × N, and s, o are the contrast
used transformation method is still DWT. The adoption or and brightness adjustment factors of D block, respectively.
design of more advanced transformation methods is expected Fractal MRI image compression process based on sequence
to further improve compression performance. The proposed image is shown in Fig. 2.
method improves compression speed of MRI images while The following are specific steps of fractal MRI image
ensuring image quality. Experimental results show the effec- compression algorithm based on sequence image:
tiveness of this method. Input: MRI image F of size M × M .
Output: Fractal encoded file dx, dy, t, s, o.
III. FAST FRACTAL-BASED COMPRESSION METHOD FOR Step 1: Perform fixed block partitioning on MRI image F,
MRI IMAGES and divide it into a range block (R block) whose size is N ×N
A. FRACTAL MRI IMAGE COMPRESSION BASED ON but does not overlap each other.
SEQUENCE IMAGE Step 2: The 2N × 2N intercepting window is moved in
Three-dimensional MRI image is essentially a data cube horizontal and vertical directions of the image F by a step size
that adds third-dimensional information to a common two- δ, and the intercepted block after each movement constitutes
dimensional image. For the form in which three-dimensional a domain block (D block).
image has a data cube, we convert three-dimensional MRI Step 3: Perform average sampling and eight equidistant
image into two-dimensional sequence images for compres- transformations on all D blocks to form a codebook .
sion. First, the image to be compressed is divided into a series Step 4: For an arbitrary range block Ri , find the best
0
of fixed size N × N pixel sub-blocks. They do not overlap matching block Dj that satisfies Eq.1 in codebook .
each other and cover the entire image, which is called a range   0     0   2
block (R block). Subsequently, the image to be encoded is d Ri , Wi Dj = min Ri − si · tk Dj + oi (1)

VOLUME 7, 2019 62415

 S. Liu et al.: Fast Fractal Based Compression for MRI Images

0
In Eq.1, Dj is the domain block after average sampling. tk ∈
{t1 , · · · , t8 } are eight isometric transforms. si and oi are con-
trast factor and luminance factor of gradation transformation,
respectively, and are calculated according to Eqs.2-3. n is the
total number of pixels, and rp , dp are the p-th pixel values of
range block and domain block, respectively.
n np=1 dp rp − np=1 dp np=1 rp
P P P
s= P 2 (2)
Pn 2 n
n p=1 dp − p=1 d
 
n n
1 X X
o=  rp − s dp  (3)
n
p=1 p=1

B. CLASSIFICATION OF MRI IMAGE SUB-BLOCKS

In traditional fractal MRI image compression algorithm, each
range block needs to perform a global search on the entire
codebook  to find the best matching block, which is the
main reason for long compression time [34]. In order to effec-
tively solve this problem, we classify image blocks according
to the spatiotemporal similarity feature. It makes matching
range of range block greatly reduced. Moreover, continuous
FIGURE 3. Flow chart of classified MRI image compression.
m-layer MRI images are combine them into one large image
matrix for overall compression. The ’codebook’ is expanded
from a single frame image to a number of consecutive frames. Suppose the class CDi contains a number of domain
2
When matching a range block, it is no longer limited to a blocks of ωi . If ωi  ( M −2N
δ + 1) /m, the threshold is
single frame. Make full use of the inter-frame correlation adjusted to σDi = σ p−1 , where σp−1 is an order of magnitude
to expand the matching pool. Matching error is reduced and smaller than σp . And so on until ωi − ( M −2N + 1) /m ≤
2
δ
compression quality is improved.
10( 10 i ) or ωi − ωi ≥ 10( 10 i ) .
ω ω
log −1 0 log −1
First, set the number m of classes, and then set a set of
threshold sequences {σi }, i = 1,
2, · · · , σ1 = 0. Find the Similarly, if ωi
2
 δ +1) /m, the threshold
( M −2N
mean DA = hA1 , hA2 , · · · , hA2b of gray-scale feature vec- 2
tors for all remaining domain blocks (excluding the former σp is adjusted to σp+1 . Until ωi − ( M −2N δ +1) /m ≤
i-1 class) as the initial cluster center of the i-th domain block 10 (log10 ωi −1) 0
or ωi − ωi ≥ 10 (log10 ωi −1)
. Adjustment of the
in Eq.4, where b is the number of bits per pixel, n is the total threshold is stopped, and the i-th class domain block CDi is
number of pixels. determined.

1 Finally, m classes of domain blocks CD1 , CD2 · · · CDm are
hA1 = (h11 + h21 + · · · + hn1 )




 n obtained.
 1 After getting domain block class, we also classify range
hA2 = (h12 + h22 + · · · + hn2 )


n (4) blocks.


 ··· According to above steps of finding the initial cluster cen-
ter, class centers c1 , c2 · · · cm for m classes of domain blocks


 1

hA2b =
 h12b + h22b + · · · + hn2b
n are obtained. Calculate the Euclidean distance between
0 range block Rn and each class center c1 , c2 · · · cm , select
Secondly, domain block Dj = hj1 , hj2 , · · · , hj2b is taken


n 0 0 o the class with the smallest distance to perform a matching
0
out one by one from domain block set D1 , D2 , · · · ,Dn . The search, and obtain corresponding m class of range blocks
distance to cluster center DA is calculated according to Eq.5. CR1 , CR2 · · · CRm . In this way, each class of range blocks
b
only needs to perform matching search in the same class
2
 0
 X
2 of domain blocks, for example, the m-th range block class
dis DA , Dj =[ hAk − hjk ]/2b (5) CRm and the m-th domain block class CDm is matched. The
k=1
matching process between different classes of range blocks
Again, let the initial threshold be σp (σp is the median of is performed independently. Reduction of matching range
threshold sequence {σi }, and the i-th class domain block CDi effectively reduces encoding time. The flow chart of MRI
corresponds to the threshold σDi = σp . Then, the domain image compression in this paper is shown in Fig. 3.
block whose distance is less than or equal to σDi is divided The following are specific steps of classified MRI image
into the i-th class CDi . compression algorithm in this paper:

62416 VOLUME 7, 2019

S. Liu et al.: Fast Fractal Based Compression for MRI Images

Input: continuous m-layer MRI image {F1 , F2 , · · · , Fm }.

Output: Fractal encoded file dx, dy, t, s, o}.
Step 1: Take out continuous m-layer images
{F1 , F2 , · · · , Fm } in MRI image slices, and combine them
into one large image matrix F for overall compression. Split
F into N × N range blocks that do not overlap each other.
Step 2: In a window of 2N × 2N , a step size δ intercepts
domain blocks along F. Shrink domain blocks and perform
eight equidistant transformations.
Step 3: According to above method, domain blocks are
classified and the corresponding class of range blocks is FIGURE 4. Flow chart of residual compensation mechanism.
obtained.
Step 4: Take a range block Ri from the p-th class of range
blocks, and take a domain block Dj from the p-th class of TABLE 1. Information of ADNI data set.

domain blocks. Calculate s and o according to (2) and (3),

until all domain blocks in the p-th class are all searched, find
the best matching block of range block Ri .
Step 5: All range domain classes are matched with their
corresponding domain block classes, and fractal codes of
continuous m layer image is obtained.

C. RESIDUAL COMPENSATION MECHANISM

Since the reconstructed image after lossy compression retains
most of information for the original image, it is considered IV. EXPERIMENTAL RESULTS AND ANALYSIS
that there is a strong correlation between the reconstructed Experiments are deployed on a computer with an Intel-
image and the original image. Residual image of the original Core i5-4590 CPU and a 12GB memory. The operat-
image and the reconstructed image is not random noise. This ing environment is Matlab2016a. MRI image sequences
conclusion has direct guiding significance for the following in the ADNI data set are taken as examples [35], and
coding. Eq.6 gives the basic formula for calculating residual image size is 160 × 192 × 192, 8 bits/pixel. Descrip-
image in this paper. tion of data set is shown in Table 1. We compare the
proposed algorithm with traditional fractal MRI image com-
r (x, y) = f (x, y) − g (x, y) (6) pression algorithm from two aspects. They are a com-
where r(x, y) is the pixel value of the residual image, f (x, y) parison of single-layer image compression and continuous
is the pixel value of the original image, and g(x, y) is the pixel layer image compression, respectively. In addition, the algo-
value of the reconstructed image. rithm of this paper is compared with the BWT–MTF
Residual image of the reconstructed image is obtained after algorithm [36]–[37]. In these experiments, size of range
the original image is greatly compressed, and correspond- block is selected as 4 × 4, and size of domain block is
ing ROI region is found on the residual image. A Huffman 8 × 8, and the horizontal step and the vertical step are
coding based on an integer squared quantization threshold is both 8.
performed on the region with a finer threshold. The obtained We evaluate quality of the decoded image by calculating
code stream data is transmitted along with the original coded PSNR by Eq.7. The peak signal-to-noise ratio (PSNR) is the
data, thereby completing lossless coding of the ROI region logarithmic value of mean squared error (MSE) between the
under Huffman coding. The algorithm framework is shown original image and the decoded image relative to (2n −1)2 ,
in Fig. 4. where 2n − 1 represents upper limit of gray level. n is the
The following are specific steps of the residual compensa- number of bits per pixel. The larger the PSNR value, the less
tion mechanism algorithm: distortion is represented. In addition, comparison indicators
Input: Original image and reconstructed image. include compression time (T), speedup ratio (SR), and com-
Output: Residual code file. pression ratio (CR) by Eqs.7-9.
Step 1: Calculate the residual image.
Step 2: Retain the important coefficient of the threshold 1) PEAK SIGNAL TO NOISE RATIO (PSNR) INDICATOR
under a certain quantization threshold and generate a basic
8−1 2
 "
#
image to find the ROI region of the remaining coefficient.  2
PSNR = 10 × log10

Step 3: Huffman coding encodes a finer integer squared


MSE
quantization threshold for the ROI region. n (7)
1 X
Step 4: The obtained code stream data is transmitted fol- (Xi − Yi )2

MSE = n



lowing the original coded data. i=1

VOLUME 7, 2019 62417

 S. Liu et al.: Fast Fractal Based Compression for MRI Images

where n is the number of pixels for the image. Xi and Yi TABLE 2. Comparison for decoding quality of traditional algorithm and
proposed algorithm.
represent gradation values of the i-th pixel for images X and
Y , respectively.

2) SPEEDUP RATIO (SR) INDICATOR

TS
SR = (8)
TC
where TS is the time required to adopt traditional fractal MRI
image compression method. TC is the time required for the
method in this paper.

3) COMPRESSION (SR) INDICATOR

160 × 192 × 10
CR = (9)
H × (8 + 8 + 3 + 8 + 3)
where H is the number of range blocks. {8,8,3,8,3} represents
the quantization level of fractal parameters {xi , yi , ti , si , oi },
respectively, which is the memory required to save these TABLE 3. Comparison for performance of traditional algorithm and
proposed algorithm.
parameters.

A. COMPARISON FOR CONTINUOUS LAYER OF MRI

IMAGES
The first 4 layers of each MRI image are selected for com-
pression. Calculate average compression time and PSNR for
the 4 layers. The size of the original image slice is 160 × 192,
size of range block is 4 × 4, and segmentation yields 40 × 48
range blocks. Size of domain block is defined as 8×8, the hor- The ‘‘codebook’’ is expanded from a single layer image to
izontal step and the vertical step are both 8, and segmentation a 4-layer image, effectively improving matching accuracy.
yields 20 × 24 domain blocks. According to the spatiotem- We first compress the 4-layer image without residual com-
poral similarity of image blocks, domain blocks are divided pensation mechanism, then compress it using the improved
into 3 classes. Accordingly, range blocks are also divided fractal coding method with residual compensation mecha-
into 3 classes. Table 2 is a restored image after compression nism. After combining, domain blocks and range block are
and decompression of the first layer image for the three MRI divided into corresponding 10 classes. Table 4 is a perfor-
images using the traditional fractal encoding algorithm and mance comparison of the three methods, where PSNR is the
the proposed algorithm. Table 3 is experimental data [average average of the 4-layer image.
T(s), average PSNR(dB)]. It can be seen from Table 4 and Table 5 that the image
It can be seen from Table 2 and Table 3 that quality of quality obtained by the proposed algorithms with residual
the reconstructed image obtained by traditional algorithm compensation mechanism is optimal. Although its compres-
and proposed is not much different observed by the human sion time is higher than the the proposed algorithms without
eye, which indicates that the algorithm is feasible. Secondly, residual compensation mechanism, it is still much smaller
compared with the traditional fractal coding algorithm, PSNR than the time required by traditional method, and its PSNR
value is slightly reduced, but coding time is significantly is also the highest of the three. When four layers of MRI
shortened. This is because the traditional method uses global images are combined and compressed together, the time is
search, and the method first classifies all image blocks. As four times that of a single layer image. But at this time all
mentioned above, all range and domain blocks are divided range and domain blocks are divided into 10 classes, so there
into 3 classes. In theory, the speed should be up to three times. is still speed-up effect. The compression time is 2.5 times
However, due to the non-uniformity of the classification, the that of traditional method. In addition, the introduction of
actual acceleration is less than three times. RCM significantly improves quality of reconstructed images.
It should effectively compensate for the error loss.
B. COMPARISON FOR MRI IMAGES WITH RESIDUAL
COMPENSATION MECHANISM C. COMPARISON BETWEEN THE PROPOSED ALGORITHM
We compress the first 4 layers of the three MRI images. AND BWT–MTF ALGORITHM
A combination strategy is used that combines the four lay- The BWT-MTF algorithm is a medical image compression
ers of images together and compresses them. Classification model that implements efficient transmission of medical
of domain blocks and range blocks is also done globally. images using Huffman coding and hybrid fractal coding block

62418 VOLUME 7, 2019

S. Liu et al.: Fast Fractal Based Compression for MRI Images

TABLE 4. Comparison for decoding quality of three algorithms. TABLE 7. Performance comparison for the proposed algorithm and
BWT–MTF algorithm.

image. Since each MRI image contains 192 layers, the time
is the total compression time.
As can be seen from Tables 6 and 7, compression ratio
of Screening and Complete 2Yr are reduced compared to
TABLE 5. Performanc comparison for different strategies of the algorithm.
the BWT-MTF algorithm, but their PSNR are significantly
improved. Compression ratios of Complete 1Yr is higher
than that of BWT–MTF algorithm. PSNRs of the three MRI
images are higher than the BWT–MTF algorithm. Unfor-
tunately, compression time of proposed method is greater
than the BWT–MTF algorithm. This is a flaw of fractal
compression algorithm. Compared to traditional algorithm,
compression speed is increased by about 2 to 3 times. Since
classification of range blocks and domain blocks is not uni-
form, it causes a difference between actual acceleration and
TABLE 6. Comparison of decoding quality for the proposed algorithm and theoretical acceleration.
BWT–MTF algorithm.

V. CONCLUSION
With the rapid development of medical imaging technology,
a large number of three-dimensional medical data, such as
MRI, CT and three-dimensional ultrasound, have been pro-
duced. The volume of three-dimensional medical image data
is large, resulting in high storage and transmission costs
of network traffic during diagnostics and treatment. Proper
compression of medical images reduces the amount of data
transferred. Compressed data is then stored and transmitted.
It not only saves storage space, but also improves transmis-
sion efficiency and shortens transmission time during remote
diagnosis. Thereby promoting the development of medical
care. In addition to retaining important information in med-
ical images, increasing the compression ratio and decoding
ability of compressed images is a major problem in medical
image compression.
BWT-MTF. In this method, a block-based Burrows-Wheeler In response to the above problems, we propose a fast MRI
compression algorithm is used to separate the regions con- image compression method based on fractal. First, three-
taining the most needed diagnostic features and then encoded dimensional MRI image is converted into two-dimensional
without significant loss in diagnostic quality. The remaining sequence image. The sequence image-based fractal com-
regions are encoded using a hybrid fractal coding algorithm. pression method is used to compress it. Secondly, range
Finally, the two coding regions are combined to reconstruct blocks and domain blocks are classified according to the
the output image. Table 6 and Table 7 are comparison data spatiotemporal similarity feature. A range block only needs
of decoding quality and performance for the algorithm in the to search for the optimal matching block in a certain type
paper and the BWT-MTF method, respectively. PSNR is the of domain block. Accelerate by reducing the capacity of
average value of all the images included in each MRI medical matching pool. Finally, a residual compensation mechanism

VOLUME 7, 2019 62419

 S. Liu et al.: Fast Fractal Based Compression for MRI Images

is introduced to achieve approximate lossless compression of [17] E. S. Ahmed, N. Benamrane, and A. Taleb-Ahmed, ‘‘Adaptive medical
MRI images. The experimental results show that compared image compression based on lossy and lossless embedded zerotree meth-
ods,’’ J. Inf. Process. Syst., vol. 13, no. 1, pp. 40–56, 2017.
with the traditional fractal MRI image compression method, [18] D. Yee, S. Soltaninejad, D. Hazarika, G. Mbuyi, R. Barnwal, and
the proposed method significantly improves compression A. Basu, ‘‘Medical image compression based on region of interest using
speed when image quality is similar. Compared with the better portable graphics (BPG),’’ in Proc. IEEE Int. Conf. Syst., Oct. 2017,
pp. 216–221.
BWT–MTF algorithm, although compression ratio is some- [19] K. V. Sridhar and K. S. R. K. Prasad, ‘‘Medical image compression
times slightly worse, the PSNR is higher than BWT–MTF. using advanced coding technique,’’ in Proc. 9th Int. Conf. Signal Process.,
Validity of the proposed method is verified. Oct. 2008, pp. 2142–2145.
[20] K. M. M. Prabhu, K. Sridhar, M. Mischi, and H. N. Bharath, ‘‘3-D warped
In the future, we need to improve works as follows. discrete cosine transform for MRI image compression,’’ Biomed. Signal
1) consider to recruit more subjects in our future studies Process. Control, vol. 8, no. 1, pp. 50–58, 2013.
and perform statistical test; 2) further construct new features [21] S. Bhavani and K. G. Thanushkodi, ‘‘Comparison of fractal coding meth-
ods for medical image compression,’’ IET Image Process., vol. 7, no. 7,
to enhance the definition of domain partition classification; pp. 686–693, Oct. 2013.
3) based on cloud computing, deep learning and other con- [22] S. Juliet, E. B. Rajsingh, and K. Ezra, ‘‘A novel medical image compression
cepts, propose a more effective medical image compression using Ripplet transform,’’ J. Real-Time Image Process., vol. 11, no. 2,
pp. 401–412, 2016.
algorithm; 4) start from the compression algorithm itself for [23] Y.-D. Zhang, Y. Jiang, W. Zhu, S. Lu, and G. Zhao, ‘‘Exploring a smart
improving its resistance to error. pathological brain detection method on pseudo Zernike moment,’’ Multi-
media Tools Appl., vol. 77, pp. 22589–22604, Sep. 2018.
REFERENCES [24] S. Juliet, E. B. Rajsingh, and K. Ezra, ‘‘A novel image com-
pression method for medical images using geometrical regularity of
[1] A. Bruckmann and A. Uhl, ‘‘Selective medical image compression tech- image structure,’’ Signal Image Video Process., vol. 9, pp. 1691–1703,
niques for telemedical and archiving applications,’’ Comput. Biol. Med., Oct. 2015.
vol. 30, pp. 153–169, 2000. [25] U. V. S. G and R. Nadarajan, ‘‘A rapid compression technique for
[2] P. G. Tahoces, J. R. Varela, M. J. Lado, and M. Souto, ‘‘Image compres- 4-D functional MRI images using data rearrangement and modified
sion: Maxshift ROI encoding options in JPEG2000,’’ Comput. Vis. Image binary array techniques,’’ Australas. Phys. Eng. Sci. Med., vol. 38, no. 4,
Understand., vol. 109, pp. 139–145, Feb. 2008. pp. 731–742, 2015.
[3] Y.-D. Zhang, Z. Dong, L. Wu, and S. Wang, ‘‘A hybrid method for MRI [26] Y.-D. Zhang, K. Muhammad, and C. Tang, ‘‘Twelve-layer deep convo-
brain image classification,’’ Expert Syst. Appl., vol. 38, pp. 10049–10053, lutional neural network with stochastic pooling for tea category clas-
Aug. 2011. sification on GPU platform,’’ Multimedia Tools Appl., vol. 77, no. 17,
[4] S. M. Hosseini and A.-R. Naghsh-Nilchi, ‘‘Medical ultrasound image pp. 22821–22839, Sep. 2018.
compression using contextual vector quantization,’’ Comput. Biol. Med., [27] S. Juliet, E. B. Rajsingh, and K. Ezra, ‘‘Projection-based medical image
vol. 42, pp. 743–750, Jul. 2012. compression for telemedicine applications,’’ J. Digit. Imag., vol. 28, no. 2,
[5] P. E. Sophia and J. Anitha, ‘‘Implementation of region based medical pp. 146–159, 2015.
image compression for telemedicine application,’’ in Proc. IEEE Int. Conf. [28] U. Patbhaje, R. Kumar, A. Kumar, and H.-N. Lee, ‘‘Compression
Comput. Intell. Comput. Res., Dec. 2014, pp. 1–4. of medical image using wavelet based sparsification and coding,’’ in
[6] M. Kaur and V. Wasson, ‘‘ROI based medical image compression for Proc. 4th Int. Conf. Signal Process. Integr. Netw. (SPIN), Feb. 2017,
telemedicine application,’’ Procedia Comput. Sci., vol. 70, pp. 579–585, pp. 394–398.
Dec. 2015. [29] G. U. V. Selvi and R. Nadarajan, ‘‘CT and MRI image compres-
[7] P. Schelkens, A. Munteanu, and J. Cornelis, ‘‘Wavelet-based compression sion using wavelet-based contourlet transform and binary array tech-
of medical images: Protocols to improve resolution and quality scalabil- nique,’’ J. Real-Time Image Process., vol. 13, no. 2, pp. 261–272,
ity and region-of-interest coding,’’ Future Gener. Comput. Syst., vol. 15, 2017.
no. 2, pp. 171–184, 1999. [30] K. C. Pathak and J. N. Sarvaiya, ‘‘Lossless medical image compression
[8] K. V. Sridhar, ‘‘Implementation of prioritised ROI coding for medical using transform domain adaptive prediction for telemedicine,’’ in Proc. Int.
image archiving using JPEG2000,’’ in Proc. Int. Conf. Signals Electron. Conf. Wireless Commun., Signal Process. Netw. (WiSPNET), Mar. 2017,
Syst., Sep. 2008, pp. 239–242. pp. 1026–1031.
[9] M. Hu, C. Zhang, J. Lu, and B. Zhou, ‘‘A multi-ROIs medical image [31] H. Amri, A. Khalfallah, M. Gargouri, N. Nebhani, J.-C. Lapayre, and
compression algorithm with edge feature preserving,’’ in Proc. 3rd Int. M.-S. Bouhlel, ‘‘Medical image compression approach based on image
Conf. Intell. Syst. Knowl. Eng., Nov. 2008, pp. 1075–1080. resizing, digital watermarking and lossless compression,’’ J. Signal Pro-
[10] M. Firoozbakht et al., ‘‘Compression of digital medical images based on cess. Syst., vol. 87, no. 2, pp. 203–214, 2017.
multiple regions of interest,’’ in Proc. 4th Int. Conf. Digit. Soc., Feb. 2010, [32] Y.-D. Zhang, C. Pan, X. Chen, and F. Wang, ‘‘Abnormal breast identifica-
pp. 260–263. tion by nine-layer convolutional neural network with parametric rectified
[11] J. Bartrina-Rapesta, J. Serra-Sagristà, and F. Aulí-Llinàs, ‘‘JPEG2000 ROI linear unit and rank-based stochastic pooling,’’ J. Comput. Sci., vol. 27,
coding through component priority for digital mammography,’’ Comput. pp. 57–68, Jul. 2018.
Vis. Image Understand., vol. 115, no. 1, pp. 59–68, Jan. 2011. [33] S. Parikh, D. Ruiz, H. Kalva, G. Fernández-Escribano, and V. Adzic, ‘‘High
[12] D. Agraftotis, D. R. Bull, and N. Canagarajah, ‘‘Region of interest cod- bit-depth medical image compression with HEVC,’’ IEEE J. Biomed.
ing of volumetric medical images,’’ in Proc. Int. Conf. Image Process., Health Inform., vol. 22, no. 2, pp. 552–560, Mar. 2018.
Sep. 2003, p. III-217. [34] X. W. Jiang, ‘‘Application of networking technology in special education,’’
[13] K. Wang and A. Cuhadar, ‘‘Unbalanced 3-D tree structure for region-based J. Nanjing Tech. College Special Educ., vol. 4, pp. 5–7, Dec. 2012.
coding of volumetric medical images,’’ in Proc. IEEE Int. Conf. Eng. Med. [35] Alzheimer’s Disease Neuroimaging Initiative. Accessed: Mar. 20, 2019.
Biol. Soc., Aug./Sep. 2006, pp. 3286–3289. [Online]. Available: http://adni.loni.usc.edu
[14] V. Sanchez, R. Abugharbieh, and P. Nasiopoulos, ‘‘3-D scalable medical [36] C. P. Devadoss and B. Sankaragomathi, ‘‘Near lossless medical image
image compression with optimized volume of interest coding,’’ IEEE compression using block BWT–MTF and hybrid fractal compression tech-
Trans. Med. Imag., vol. 29, no. 10, pp. 1808–1820, Oct. 2010. niques,’’ Cluster Computing, 2018. doi: 10.1007/s10586-018-1801-3.
[15] B. P. Nguyen, C.-K. Chui, S.-H. Ong, and S. Chang, ‘‘An efficient compres- [37] X. Jiang, ‘‘Isolated Chinese sign language recognition using gray-level
sion scheme for 4-D medical images using hierarchical vector quantization co-occurrence matrix and parameter-optimized medium Gaussian support
and motion compensation,’’ Comput. Biol. Med., vol. 41, pp. 843–856, vector machine,’’ in Proc. 7th Int. Conf. Frontiers Intell. Comput., Theory
Sep. 2011. Appl. (FICTA), 2019, pp. 1–10.
[16] V. Sanchez, ‘‘Joint source/channel coding for prioritized wireless trans-
mission of multiple 3-D regions of interest in 3-D medical imaging data,’’
IEEE Trans. Biomed. Eng., vol. 60, no. 2, pp. 397–405, Feb. 2013.

62420 VOLUME 7, 2019