This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNB.2017.2771350, IEEE
Transactions on NanoBioscience

A new adaptive gamma correction based algorithm using

DWT-SVD for non-contrast CT image enhancement
*, **Fathi KALLEL, *Ahmed BEN HAMIDA

*Laboratory of Advanced Technologies for Medicine and Signal, National Engineering School of Sfax, University of
Sfax, Route Soukra km 3, Sfax, B.P.W, 3038, Tunisia.
**Electrical Engineering Department, College of Engineering, Majmaah University, P.O. Box 66, Majmaah 11952,
Saudi Arabia
Corresponding author: [email protected]
Abstract— the performances of medical image processing illustrate valuable medical information. CT images were
techniques, in particular CT scans, are usually affected by poor widely used as an efficient medical imaging tool and replaced
contrast quality introduced by some medical imaging devices. many imaging techniques which have been inappropriate in
This suggests the use of contrast enhancement methods as a illustrating pathology, anatomy and more truculent diagnostic
solution to adjust the intensity distribution of the dark image. In
checks. Nevertheless, CT scans for some parts of the
this paper, an advanced adaptive and simple algorithm for dark
medical image enhancement is proposed. This approach is body, like the liver, have low contrast which eventually
principally based on adaptive gamma correction using discrete results in an inaccurate diagnosis. Some contrast agents may be
wavelet transform with singular value decomposition (DWT- introduced for the purpose of enhancing the CT scans’ contrast,
SVD). In a first step, the technique decomposes the input medical although they are hurtful or even deadly sometimes for some
image into four frequency sub-bands by using DWT and then patients due to the occurrence of anaphylaxis [1, 28].
estimates the singular value matrix of the LL sub-band image. In Image contrast and quality enhancement or image detail
a second step, an enhanced LL component is generated using an improvement algorithms are necessary to highlight the image
adequate correction factor and inverse SVD. In a third step, for details in different medical image processing applications,
an additional improvement of LL component, obtained LL sub-
particularly non-contrast CT scans. A common purpose in
band image from SVD enhancement stage is classified into two
main classes (low contrast and moderate contrast classes) based these applications is to enhance the contrast and the quality of
on their statistical information and therefore processed using an like image while preserving the edge information [14]. Indeed,
adaptive dynamic gamma correction function. In fact, an like degradations may have a crucial impression on the quality
adaptive gamma correction factor is calculated for each image of the image; and consequently, it affects the human
according to its class. Finally, the obtained LL sub-band image interpretation as well as the performances of computer assisted
undergoes inverse DWT (IDWT) together with the unprocessed methods in medical imaging.
low-high (LH), high-low (HL) and high-high (HH) sub-bands for Different techniques are proposed in literature to repair the
enhanced image generation. Different type of non-contrast CT damaged images and improve their quality, improve their
medical images is considered for performance evaluation of the
contrast and brightness ([13, 21, 22, ], Jha and Chouhan, 2014).
proposed contrast enhancement algorithm based on adaptive
gamma correction using DWT-SVD (DWT-SVD-AGC). Results In fact, contrast enhancement of dark medical images, like non-
show that our proposed algorithm performs better than other contrast CT images, is an important and crucial step in image
state-of-the-art techniques. processing, particularly when reimaging is not possible [35, 28,
45]. In general, we can distinguish two main categories of
Keywords— non-contrast CT image, contrast enhancement, contrast enhancement methods: spatial and frequency domain
DWT-SVD, adaptive Gamma correction methods. Spatial domain algorithms are considered for a direct
processing of image’s pixels. Histogram equalization (HE)
I. INTRODUCTION methods are one of the most image enhancement techniques
Medical images have an essential role in diagnosing a used in spatial domain [34]. Nevertheless, HE does not always
disease and monitoring the patient’s health condition and the give acceptable performances since it could produce a contrast
impact of the picked treatments. Despite the increasing lost for less frequent gray levels and an over-enhancement for
progress in the techniques of capturing these images, the frequent ones [26]. Singular Value Decomposition method is
generated images could not pose satisfactory quality for an also considered for image equalization [15]. This method
accurate diagnosis. Environmental noises, patients’ special preserves the general shape of the histogram and significantly
conditions in photography, emergency situations, lighting reduce the loss of information contained in the histogram.
conditions and technical constraints of imaging devices are Some other techniques, operating in the frequency domain,
amongst the reasons why images may have low quality [1]. are also presented and are based on changing the transform
Therefore, miscellaneous kinds of artefact and noise in imaging domain of an image [9]. In order to protect the edge
modalities degrade the quality of such image. CT image information from possible degradation, Bhandari et al proposed
obtains an important amount of information about the human applying SVD algorithm only on LL sub-band image, obtained
body to generate its corresponding cross-sectional images that using DWT [6]. Different contrast enhancement algorithms

1536-1241 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNB.2017.2771350, IEEE
Transactions on NanoBioscience

based on gamma correction are also presented [28, 31]. problems, different HE algorithms are proposed, such as
Bhandari et al. proposed a simple approach for enhancement of Brightness preserving bi-histogram equalization (BBHE) [27],
dark image based on knee function and gamma correction dualistic sub-image histogram equalization (DSIHE) [41],
using DWT-SVD [7]. Gamma correction factors are set brightness preserving dynamic histogram equalization
manually, which could not give a suitable enhancement for (BPDHE) [22], minimum mean brightness error bi-histogram
different types of images. equalization [12,24] and recursively separated and weighted
Tsai et al. classified the input image into six parts and HE algorithm, which is a combination of BBHE and DSIHE
therefore processing each part with the adequate enhancement [26]. Like improved algorithms split a histogram in different
technique [39]. In fact, this classification is accomplished sub-blocks and perform HE on each individual partition. Tan et
considering manual defined thresholds which could not always al. proposed a nonlinear local HE method to enhance the
fit an adequate contrast improvement of different types of contrast of CT brain images [38]. The main idea of this
images. To mitigate this problem, Rahman et al. proposed an algorithm is to decompose the input image into different sub-
adaptive gamma correction (AGC) method suitable for the blocks. The distribution of the maximum levels in the
enhancement of a huge type of images [30]. Gamma correction histogram are excluded for each sub-block and therefore
factor is calculated dynamically for each image according to its redistributed to the other gray levels with a threshold
statistical information. However, this method based contrast limitation. A gray level reallocation function is finally defined.
enhancement cannot conserve the boundary or edge details of The major drawback of this algorithm is the manual setting of
original image. the threshold used for contrast limitation. Huang and Yeh
In this paper, we propose a new approach for low contrast proposed a histogram equalization based image enhancement
CT scans enhancement based on adaptive gamma correction method including two main modules. The first histogram
and using DWT-SVD (DWT-SVD-AGC) technique. The main separation module is used to separate the histogram in small-
advantage of such a dynamic method is that it could be applied scale detail. The second intensity transformation module is
to large types of images for contrast enhancement with considered for a further contrast enhancement of the image
simultaneously preserving the edge information of the original with complete brightness preservation for each generated sub-
image. In fact, the SVD technique is applied in a first step for histogram [19]. Sundaram et al. described a technique to
contrast enhancement of LL sub-band image obtained using improve the contrast of mammogram images. This technique
DWT. In a second step, for a further contrast improvement, the included both histogram modifications and local contrast
obtained LL sub-band image is processed using a modified enhancement techniques [37]. Al-Juboori presented a method
transfer function based on adaptive gamma correction to enhance the contrast of mammographic image. This method
transformation. Parameters of gamma transformation are combined both contrast limited adaptive histogram equalization
dynamically and automatically calculated depending on the and retinex techniques [3]. Ganesan et al. presented a seed
statistical information of the processed LL sub-band image. dependent adaptive region growing approach for contrast
This paper mainly includes different sections distributed as enhancement of CT images [18].
following. Section 2 illustrates a short discussion of previous SVD techniques are also considered to improve the contrast
works. Section 3 provides a global presentation of DWT and of dark image and also to overcome the limits of HE methods
SVD algorithms. Section 4 details the idea of intensity [15, 47]. In fact, the singular value matrix contains the
transformation and gamma correction techniques. Section 5 illumination information in the image, so that the conversion of
explains the proposed AGC algorithm using DWT-SBD singular values will directly change the illumination of the
technique. Section 6 presents and discusses the obtained results image, and other information present in the image will be as
considering our proposed method compared to some others same as before. In order to preserve the edge information from
published algorithms. Section 7 concludes the paper. possible degradation, Bhandari et al. proposed the DWT-SVD
algorithm for satellite image enhancement [10]. In this
II. LITERATURE REVIEW algorithm, authors applied SVD method only on LL sub-band
Several algorithms are proposed in the literature to enhance the image computed using DWT. A new LL sub-band is obtained
contrast of different types of images. Global histogram by multiplying the initial singular value matrix by a correction
equalization (GHE) algorithm is one of the most used factor. This factor is equal to the ratio of highest singular value
techniques for contrast enhancement. The principle of GHE is of the created normalized matrix, having mean zero and
based on distributing the pixel values of an input image over its variance of one, over a normalized input image. Modified LL
dynamic intensity range according to its linear cumulative sub-band image is combined with unprocessed LH, HL, and
histogram [34]. In fact, the whole image is divided into HH sub-bands to generate the enhanced image. The
equidistant levels defining different boundaries. In order to performances of this technique are compared with GHE,
obtain a flat histogram, the boundaries limiting the initial levels BPDHE, and SVD techniques, and results showed the higher
are modified by intercalating a number of new boundaries performances of DWT-SVD over the others. Bhandari et al.
between the old ones. The new boundaries are calculated using proposed an improved algorithm for satellite images
a recursive relation [40]. enhancement based on artificial bee colony using DWT-SVD
Nevertheless, GHE algorithm conducts in some cases to an [6]. The method employed the artificial bee Colony method to
excessive contrast enhancement [26]. To overcome these learn the parameters of the adaptive thresholding function

1536-1241 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNB.2017.2771350, IEEE
Transactions on NanoBioscience

needed for optimum enhancement. Atta and Abdel-Kader III. DWT–SVD DECOMPOSITION
presented an improved method using SVD and DWT for
contrast enhancement [2]. Indeed, authors considered the A. Discrete wavelet transform
singular matrices of both input image and processed image Discrete wavelet transform coefficients of a given level could
using GHE to compute the enhanced singular value matrix. The be computed from the DWT coefficients of the preceding
technique provides enhanced images without the addition of level, which is expressed as following [17].
artefacts and modification in mean brightness of the original
image.
WL  p, q   W m, q  1h(m  2 p)
m
L (1)

 p, q  W m, q  1g (m  2 p)
Yang et al. reported a technique using both wavelet and
Haar transforms for the contrast enhancement of medical WH L (2)
images [45]. Wan and Shi proposed a contrast enhancement m
method based on Daubechies wavelets and exact histogram where, WL (p, q) denotes the p-th scaling coefficient at q-
equalization [44]. Kaur and Singh proposed an algorithm for th level and WH(p, q) expresses the p-th wavelet
contrast improvement of cephalometric images. This algorithm coefficient at q-th level. Furthermore, g(n) and h(n)
is based on adaptive HE with gray level information histogram expresses the dilation coefficients corresponding
using wavelet [25]. Cheng and Huang presented a method for respectively to wavelet and scaling functions.
image and video contrast enhancement using Bezier curve. The For initial signal reconstruction, the DWT coefficients
parametric Bezier curve is used to illustrate the mapping curve are upsampled and processed by a second pack of low-
for the intensity transformation [11]. Lee et al. proposed a pass and high-pass filters using equation 3.
W k, q  1h ( p  2k )  W
dominant brightness level analysis and adaptive intensity
WL  p, q   L
'
 
'
H l , q  1 g ( p  2l ) (3)
transformation based contrast enhancement method. This
technique performs brightness-adaptive intensity transfer k l
functions using the LL sub-band image and therefore where, g’(p) and h’(p) indicates respectively the high-
transforms intensity values according to the transfer function pass and low-pass synthesis filter of considered mother
[30]. wavelet.
Different others methods stretching from traditional gamma B. Singular value decomposition
correction to more complex algorithms using adaptive gamma
correction are also proposed. In traditional gamma correction, The singular value matrix is calculated by applying SVD on
the correction factors are set manually, which could not give a original image matrix. In fact, a given real matrix I may be
suitable enhancement for different types of images in some decomposed into a product of three matrix, commonly
cases [17]. Huang and Chen presented a hardware-oriented calculated according to equation (4) (Bhandari et al, 2014).
contrast enhancement method to analyze the context of the I  U I  I VIT (4)
image and to pipeline the different processing stages for where, ΣI is the singular value matrix which is a diagonal
process speed up [4]. Bhandari et al. proposed a simple matrix, UI and VI are used as orthogonal matrices and T
approach for enhancement of dark image based on knee is transpose operator. The singular value matrix ΣI
function and gamma correction using DWT-SVD [7]. Frosio
includes intensity information of the image, that’s why
presented a contrast enhancement algorithm of cephalometric
radiographs based on soft tissue filter using histogram SVD technique is considered for image equalization.
clustering and gamma correction [16]. Histogram based The ratio of highest singular value of generated
clustering method is firstly considered to divide the image normalized matrix, with mean zero and variance of one,
pixels into bony and not bony pixels. Local gamma correction for a particular input image may be calculated according
is therefore applied for a better visualization of both bony and to equation (5).
soft tissue features. Somasundaram and Kalavathi proposed an max  N  0,VAR 1
 (5)
max I 
automatic method for medical image contrast enhancement
based on gamma correction. The gamma value is calculated
using cumulative histogram. Rahman et al. proposed an where, ΣN(μ=0, var=1) is the singular value matrix of the
adaptive gamma correction (AGC) for image enhancement, synthetic intensity matrix. These coefficients could be
which is computationally low complex and simultaneously considered for equalized image ‘Iequalized’ reconstruction
suitable for a huge type of images. AGC’s parameters are according to equation (6).
computed dynamically according to the statistical information I equalized  U I   I VIT (6)
of the respective images. In fact, input dark image is firstly
classified into either low or moderate contrast class according
IV. INTENSITY TRANSFORMATION AND GAMMA
to the available contrast of the image. The value of the
CORRECTION
correction factor γ is calculated differently for each class [30].
The gamma correction transformation is given by the
following equation (Gonzalez and Woods, 2008).
I out  c.Iin (7)

1536-1241 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNB.2017.2771350, IEEE
Transactions on NanoBioscience

where Iin and Iout represent respectively the input and obtained curves are more steepness for the lower
output image intensities. Both c and γ parameters are contrast, as desired.
used to adjust the shape of the transformation function.
In fact, a set of parameters (γ, c) could produce high
performances for some considered images but not for
some others [20]. In order to overcome this problem,
Rahman et al. proposed a new method where γ and c
parameters are determined dynamically and
automatically for each input image according to its
statistical characteristics [30]. Indeed, authors proposed
to classify an original image I into either low-contrast
class C1 or moderate contrast class C2 according to the
contrast of considered image using equation (8).
(a)

C if D 1
g (I )   1  (8)
 C
 2 otherwise
where D = diff((μ + 2σ), (μ − 2σ)) and τ is a constant
considered to define the contrast of an input image.
Experiments showed that τ = 3 is an optimal choice to
characterize the contrasts of different images. The standard
deviation and the mean of the considered image intensity are (b) (c)
respectively denoted by σ and μ [30]. According to equation
(8), we classify an image as a low-contrast (class C1) if 4σ ≤
1/τ, meaning that the major pixel intensities of considered
image are grouped within a small range. Otherwise, the image
is classified as moderate contrast (class C2).
A. Dark images with low contrast (in C1)
(d)
The majority intensities of a dark input image from the (e)
first class are grouped in a small range of dark gray Fig.1. (a) Transformation curves for low-contrast dark images for different
levels around the mean intensity of the considered values of µ and σ (b) Example of an original low-contrast dark image (c)
image. In order to improve the contrast of like input histogram of original (d) Enhanced images (e) histogram of enhanced image.
image, the transformation curve requires to flaunt the
dark intensities to the higher intensities. Therefore, the Figures 1(b) and 1(c) present respectively an example of
main constraint in this case is to generate a a low-contrast dark image (µ= 0.069 and σ= 0.06) and
transformation function that lies above the line Iout = Iin. its corresponding histograms. Figures 1(d) and 1(e) show
As a response to this constraint, Rahman et al. show that respectively the enhanced image (γ= 4.0383) and its
the value of γ could be calculated using equation (9) corresponding histograms.
[30]. B. Dark images with moderate contrast (in C2)
   log2 ( ) (9) The intensities of a dark input image from the second
The parameter c is also calculated dynamically for class are scattered over the available dynamic range. In
different images according to the nature of the respective this case, Iout and c are calculated similarly using
image using equation (10). equations (7) and (10). However, the correction factor γ
1 is expressed differently using equation (13), not to make
c (10)
1  Heaviside (0.5   ).(k  1) much stretching of the contrast [30].
where k is defined by equation (11). 
1    
  exp    (13)
k  I in  (1  I in ).  (11)  2 

 
and the Heaviside function is given by equation (12) [33].
Figure 2(a) shows that the transformation curves for
0, if x  0 low-contrast dark images generated for different values
Heaviside ( x)   (12)
 1, if  0 of µ and σ. Figures 2(b) and 2(c) show respectively an
Figure 1(a) shows that the transformation curves for example of the input dark image with a moderate
low-contrast dark images generated using equations (7), contrast (µ=0.4 and σ=0.11) and its corresponding
(10) and (11) for different values of µ and σ fall histogram. Figures 2(d) and 2(e) illustrate the enhanced
effectively above the line Iout=Iin. In addition, the image (γ=1.3492) and its corresponding histograms.

1536-1241 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNB.2017.2771350, IEEE
Transactions on NanoBioscience

Hence, separating the high-frequency sub-bands and

applying a contrast enhancement on only LL sub-band
will protect the edge information from possible
degradation. In a first enhancement step, SVD method is
applied over both low frequency components LL and
to generate respectively UL, ΣL, VL, and , , . The
maximum element in UL and VL, from LL and the
maximum element in and from are
respectively calculated to determine the correction factor
ξ. The correction factor ξ, the enhanced singular value
matrix and the enhanced LL sub-band SVD are
respectively calculated using equations (14), (15) and
(a)
(16) [7].

 
 
maxUˆ L  maxVˆL   (14)
maxU L   maxVL 

L   x ˆ L (15)

(b) (c)
LL SVD  Uˆ L L Vˆ T (16)
L

After that, the enhanced LL sub-band using SVD

approach, SVD, is classified according to equation (8)
into either low-contrast class C1 or moderate contrast
class C2 depending on the available contrast of this sub-
(d) (e)
Fig.2. (a) Transformation curves for moderate-contrast dark images for band image (µ and σ represent respectively the mean and
different µ and σ (b) Example of an original moderate-contrast dark image (c)
histogram of original image (d) Enhanced image (e) histogram of enhanced
the standard deviation of the enhanced LLSVD sub-band
image image in this case). An adaptive gamma factor correction
is calculated dynamically according to the obtained
V. PROPOSED MEDICAL IMAGE ENHANCEMENT class: equation (9) is considered for LL sub-band images
ALGORITHM with low contrast (class C1) and equation (13) for LL
The general method of the proposed algorithm for sub-band images with moderate contrast (class C2). In a
medical image enhancement is concerted in different second enhancement step, an adaptive intensity
parts described as follows. The dark input medical image transformation using gamma correction is applied on
‘Ii ’is firstly processed by GHE algorithm in order to SVD sub-band according to equation (17) to generate
compute ‘ ’. Both images are decomposed by DWT the final enhanced γ sub-band image.
into LL, LH, HL, and HH for ‘Ii’, and , , , and 
LL   c. LL SVD  (17)
for ‘ ’. Indeed, as shown in figure 3, enlightenment where c is computed using equation (10) and (18).
information is surrounded in LL sub-band but the edges
are concerted in other sub-bands (i.e., LH, HL, and HH). 
k  LL SVD   1  LL SVD  . (18)
 
The generated enhanced LL sub-band using gamma
correction, γ is recombined with others sub-band images of
the original image (LH, HL and HH) using IDWT to generate
the resultant equalized medical image ‘ ’.
Ii  IDWT LL  , LH , HL, HH  (19)
 
Figure 4 shows a block diagram describing the mechanism of
contrast enhancement of dark medical image according to our
proposed method.

(a) (b)
Fig. 3. (a) Original image, (b) LL, LH, HL and HH sub-band images

1536-1241 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNB.2017.2771350, IEEE
Transactions on NanoBioscience

B. Qualitative evaluation
The enhanced images using AGC, DWT-SVD, DWT-
SVD-KGC and our proposed DWT-SVD-AGC methods
are respectively illustrated in figures 5(b1-b4), 5(c1-c4),
5(d1-d4) and 5(e1-e4). As shown in figure 5, enhanced
non- contrast CT scans using Demirel’s algorithm
maintain average brightness level and better enhance
with overall image quality, however some edge details
are lost in low and high intensity ranges. Nevertheless,
the low-intensity range was not sufficiently enhanced
using DWT-SVD method and this could be explained by
singular-value constraint of the considered image. When
we considered gamma transformation as a second
processing step, a further improvement in image quality
is observed. A better overall contrast enhancement with
minimum degradation and higher edge preservation are
noted when proposed DWT-SVD-AGC is considered.
C. Quantitative evaluation
For further performance evaluation of considered image
enhancement methods, we present obtained quantitative
experimental results. Hence, different evaluation
parameters are considered to measure the fidelity of
enhanced images to the original image.
Fig. 4. Block diagram of the proposed enhancement algorithm The Mean (µ) indicates the average of all intensities of
the image. The standard deviation (σ) indicates the
VI. PERFORMANCE EVALUATION AND RESULTS deviation of intensity values about the mean. µ and σ are
In this paper, we evaluated and compared the respectively calculated using equations (20) and (21).
performances of our proposed DWT-SVD-AGC M 1N 1

enhancement technique to three different state-of-the-art 

1
MN  I x, y
x 1 y 1
(20)
techniques in the case of CT medical images: Adaptive
gamma correction method (AGC) [30], DWT-SVD M 1N 1

technique [14] and knee and gamma correction using  

1
MN  I x, y   
x 1 y 1
2 (21)
DWT-SVD technique (DWT-SVD-KGC) [7].
DWT is applied on selected dark image using Meyer Where (M, N) indicate the size of considered image and I
(dmey) wavelet as a mother wavelet and with only one (x, y) corresponds to the intensity value of original pixel
level of sub-bands decomposition. Our main goal is to (x, y).
get an output image with higher contrat and maximum PSNR in dB (PSNRdB) is another quality measurement
edge preservation. Both qualitative and quantitative parameter and corresponds to the peak signal-to-noise
parameters are considered for performance assesment. ratio in decibels between original and enhanced images.
PSNRdB parameter is calculated using equation (22)
A. Material [42].
For qualitative evaluation, we selected a variety of four  L  12 
non-contrast CT images to illustrate the effectiveness of PSNRdB  10 log10   (22)
 MSE 
our proposed algorithm. These images are respectively  
illustrated in figures 5 (a1-a4) and corresponds to non- 2
    
1 M N
contrast CT scans of the lungs [36], the liver [1], the MSE    I x, y  I x, y (23)
MN x 1 y 1
brain [5] and the kidneys [43].
For quantitative evaluation, we considered a database Where L is the global number of gray-scale, MSE is
including 115 high-resolution CT images coming from calculated according to equation (23) and indicates the
39 subjects. CT scans was performed using general cumulative squared error between processed and original
electric equipment with four detector rows and using the images, and Ī (x, y) respresents the intensity level of
following parameters: in-plane resolution 0.78 x 0.78 enhanced pixel (x, y).
mm, slice thickness 1.25 mm, tube voltage 140 kV, and
tube current 200mA [36].

1536-1241 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNB.2017.2771350, IEEE
Transactions on NanoBioscience

(a1) (a2) (a3) (a4)

(b1) (b2) (b3) (b4)

(c1) (c2) (c3) (c4)

(d1) (d2) (d3) (d4)

(e1) (e2) (e3) (e4)

Fig. 5. Qualitative evaluation of non-contrast CT scans enhancement: (a1-a4) Original dark with non- contrat CT images, (b1-b4) enhanced image using AGC
method [13], (c1-c4) enhanced image using DWT-SVD method [7] (d1-d4), enhanced image based on DWT–SVD–KGC method [4] and (e1-e4) enhanced image
based on DWT–SVD– AGC proposed method.

1536-1241 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNB.2017.2771350, IEEE
Transactions on NanoBioscience

where, Ω indicates the entire image, SL(X) represents the

The other assessment parameter considered is this work
similarity between enhanced and original images and
for performance evaluation is the Weighted Peak Signal PCm(X) represents the phase congruency which is a
to Noise Ratio in dB (WPSNRdB) calculated according to dimensionless measure of the significance of a local
equation (24) [46]. structure.
 L 1 
2 All non-contrast CT images are processed using
WPSNRdB  10 log10   (24) considered enhancement algorithms. Figure 6 compares
 MSE  NVF 

the avereage Mean, Variance, PSNR, WPSNR and
where NVF is the Noise Visibility Function. It is QRCM values between AGC, DWT-SVD, DWT-SVD-
variable according to a texture masking function and KGC and proposed DWT-SVD-AGC methods.
ranges from zero (for extremely textured area) to one
(for smooth area).
We considered also the quality-aware relative contrast
measure (QRCM) for comparison. It uses pixel gradients
for image quality assessment as given by equation (25).
 RCM  Q RCM  0 
QRCM    (25)
1  RCM  Q  1 RCM  0
where Q  1 1 u v GMSu,v μ  ω2 u,v , 2 u, v  1 1  GO u, v ,
G
 2G u, vG p u, v 
GMSu, v  
O
. Subscripts (a) (b)

 GO u, v2  G p u, v2 ^ 
 

‘o’ and ‘p’ indicates the original and processed images,

G u, v 
  255 RCM  u v G p,o u, v 1u, v , 1u, v   O ,
k l Go k , l 
2 ,
G u, v  Go u, v
G p,o u, v   p
G p u, v  Go u, v   . For a given 3x3 patch,
both derivatives  v in vertical direction and  u in
horizontal direction are used to calculate the gradient by
G  2u  2v . and better image quality
is obtained with the QRCM value closest to unity [10]. (c) (d)
Fig.6. Performance evaluation of DWT-SVD method [14], DWT-SVD-KGC
Structure similarity index measurement (SSIM) is an method [7], AGC method [30] and proposed DWT-SVD-AGC method using
image quality evaluation based on the degradation of Comparison of the (a) Mean, (b) Variance, (c) PSNR, (d) EME, (e) QRCM.
structural information. In this comparative study, SSIM
is considered to compare the structures of original and Figure 7 compares the average SSIM and FSIM values
processed image [42]. The SSIM index is formulated between same above considered enhancement methods.
using equation (26)
SSIM x, y  
2 x  y  1 2 xy  2 
 x2   2y  1  x2   y2  2  (26)

where, μx, μy, σx and σy represent respectively the mean

intensity and the standard deviations of x and y, σxy is the
local sample correlation coefficient between x and y. λ1
and λ2 are two constants, and are considered to shun
instability when μ2x+μ2y are very close to zero. Here, λ1 =
λ2 = 0.065. SSIM value are ranged betwenn -1 and 1, and
better performances are obtained with heiger SSIM (a) (b)
values . Fig. 7. Performance evaluation of DWT-SVD method [14], DWT-SVD-KGC
The last considered assessment method is Feature method [7], AGC method [30] and proposed DWT-SVD-AGC method using
Comparison of the (a) SSIM and (b) FSIM parameters.
similarity index measurement (FSIM) which compute
also the feature similarity between input and enhanced In addition to visual comparison, quantitative
images. FSIM is determined using equation (27) [48]. comparisons approved the superiority of our proposed
FSIM 
 X 
S L  X PC m  X  method. Simulation results showed better performances
 X 
PC m  X  (27)
for the enhanced images using proposed DWT-SVD-
AGC method. In fact, proposed algorithm presented a

1536-1241 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNB.2017.2771350, IEEE
Transactions on NanoBioscience

better mean value equal to 101.94 (compared to 120.97, However, AGC algorithm is applied on the overall of the
114.14 and 102.21 when DWT-SVD, DWT-SVD-KGC image, and therefore some details could be lost during
and AGC contrast enhancement algorithms are enhancement process.
respectively considered). A good average standard We compared also the performances of the considered
deviation is also obtained when proposed method is enhancement algorithms in term of running time using
considered for non-contrast CT images enhancement (in an Intel Core i5-5200U-2.2GHz processor machine and
this case, the average standard deviation is equal to 4Go of RAM under windows operating system. Results
103.34 compared to only 64, 91.12 and 89 for are reported in Table 1.
respectively DWT-SVD, DWT-SVD-KGC and AGC
methods). PSNR evaluation parameter showed also the Table 1. Running time for DWT–SVD, DWT–SVD-KGC, AGC and Proposed
heigher performances of our proposed algorithm. Indeed, methods
the average value of PSNR was about 12.5 dB for Contrast enhancement Running time
proposed algorithm, and only 8.22 dB, 8.71 dB and 8.38 algorithm (second)
dB for DWT-SVD, DWT-SVD-KGC and AGC methods DWT–SVD 0.2
respectively. Average WPSNR values indicated also the DWT–SVD-KGC 0.21
heigher performances of our proposed algorithm. The AGC 0.017
mean values were approximately equal to 9.4 dB, Proposed 0.23
10.26dB, 9.15d B and 15,88dB for DWT-SVD, DWT-
SVD-KGC, AGC and DWT-SVD-AGC methods
We note that the AGC contrast enhancement algorithm
respectively.
has the shortest running time among the considered methods.
QRCM is also considered to compare the performances
Nevertheless, it cannot guarantee the quality of enhanced
of considered enhancement method. The better QRCM image. The computing time for the others three contrast
mean value are achieved from DWT-SVD-AGC method enhancement algorithms using DWT technique (DWT-
at 0.155. Mean QRCM values are limited to 0.082, 0.11 SVD, DWT-SVD-KGC and proposed algorithms) is
and 0.065 for DWT-SVD, DWT-SVD-KGC and AGC approximately equal with a little higher value for the proposed
methods respectively. method.
The above results indicated that both gamma correction Finally, we conclude that results showed an important
using DWT-SVD algorithms (i.e. DWT-SVD-KGC and complementarity between DWT-SVD and gamma
proposed DWT-SVD-AGC) performed better than correction techniques for non-contrast CT image
DWT-SVD algorithm (without gamma correction). This enhancement. Furthermore, compared with DWT-SVD,
observation emphasize that Gamma correction is an DWT-SVD-KGC and AGC algorithms, the proposed
important tool given further improvement on processed enhancement algorithm balance between image quality and
image using only DWT-SVD method. In addition, we running time, which is more feasible in practical image
note the importance of an ideal choose of the gamma enhancement.
correction parameters. Hence, gamma transformation
parameters, which are adaptively and dynamically VII. CONCLUSION
calculated according to the nature of the respective non- This paper proposed a new contrast enhancement
contrast CT scan image, may explain the superiority of technique based on adaptive gamma correction using
our proposed method compared to DWT-SVD-KGC DWT-SVD method. This algorithm is dedicated for
where parameters are set manually for each image. contrast enhancement of CT images. DWT algorithm is
According to SSIM and FSIM measurements, lower considered to decompose the dark original image into
degradation of structural information and therefore better different sub-band images. In order to obtain an
image quality are obtained when input images are improved image characterized by higher contrast with
processed using our proposed method. Performances are edges preservation, only LL sub-band images are
respectively decreased with DWT-SVD-KGC, DWT- processed using SVD method and adaptive intensity
SVD and AGC methods. Furthermore, it can be noted transformation using gamma adjustment function.
that DWT-SVD based algorithms performed Parameters of gamma correction are computed
significantly better in term of image structure dynamically and automatically for each image according
conservation with higher values of SSIM and FSIM to its statistical information.
parameters than AGC method. We can explain this The performances of proposed AGC-DWT-SVD
observation by the fact that when DWT-SVD based algorithm are assessed with three state-of the-art
methods are considered, transformations were applied enhancement algorithms. Experimental results revealed
only on low frequency image details (high frequency that the proposed AGC-DWT-SVD algorithm can
image details are approximately unchanged) which allow successfully enhance the non-contrast CT images and
a preserving of the edge information of original image.

1536-1241 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNB.2017.2771350, IEEE
Transactions on NanoBioscience

achieved that the overall quality and visibility of the [21] Hasikin K, Mat Isa, N.A. Adaptive fuzzy contrast factor enhancement
technique for low contrast and non-uniform illumination images, SIViP, 2014,
local details were better than others techniques. 8(8), pp. 1591-1603.
[22] Ibrahim H, Kong N.S.P. Brightness preserving dynamic histogram
ACKNOWLEDGMENTS equalization for image contrast enhancement, IEEE Trans. Consum. Electron,
2007, pp. 1752-1758.
The author would like to thank Deanship of Scientific [23] Jha R.K, Chouhan R. Noise-induced contrast enhancement using
Research at Majmaah University for funding this work stochastic resonance on singular values, SIViP, 2014, 8(2), pp. 339-347.
[24] Jagatheeswari P, Suresh Kumar S, Rajaram M. Contrast Enhancement for
under Project No. 37/55. Medical Images Based on Histogram Equalization Followed by Median Filter,
Proceedings of the International Conference on Man-Machine Systems
REFERENCES (ICoMMS), 11-13 October 2009, Malaysia.
[1] Al-Ameen Z, Al-Ameen S, Sulong G. Latest Methods of Image [25] Kaur N, Singh E. Enhancement of Medical Images using Histogram
Enhancement and Restoration for Computed Tomography: A Concise Based Hybrid Technique, International Journal of Advanced Engineering,
Review. Applied Medical Informatics; 2015; 36(1), pp. 1-12. Management and Science, 2016, 2(9), pp. 1425-1432.
[2] Atta R, Abdel-Kader R.F. Brightness preserving based on singular value [26] Kim M, Chung MG. Recursively separated and weighted histogram
decomposition for image contrast enhancement, Optik - Int. J. Lig. and Elec. equalization for brightness preservation and contrast enhancement. Consum.
Opt., 2015, 126(7-8), pp. 799-803. Electron. IEEE Trans; 2008, 54(3), pp. 1389-1397.
[3] Al-Juboori R.A. Contrast Enhancement of the Mammographic Image [27] Kim Y.T. Contrast enhancement using brightness preserving bi-
Using Retinex with CLAHE methods, Iraqi Journal of Science, 2017, 58(1B), histogram equalization. Consum. Electron. IEEE Trans; 1997, 43(1), pp. 1-8.
pp. 327-336. [28] Liang Y, Yanga L, Fanb H. Image enhancement for liver CT images.
[4] Huang S, Chen W. A new hardware-efficient algorithm and reconfigurable Proceedings of the SPIE International Conference on Optical Instruments and
architecture for image contrast enhancement, IEEE Trans. Image Process., Technology: Optoelectronic Imaging and Process Technology; 2009.
2014, 23(10), pp 4426-4437. [29] Lee S. An efficient contrast-based image enhancement in the compressed
[5] Arora AJ, Bolla H, Arora R, Kaul S, Yarlagadda J. Basal ganglia domain using retinex theory. IEEE Transactions on Circuit and Systems for
infarction due to mild trauma in a case of mineralizing angiopathy: A rare Video Technology; 2007, 17(2), 199-213.
entity. Neurol India 2015; 63:771-2 [30] Lee E, Kim S, Kang W, Seo D and J. Paik, Contrast Enhancement Using
[6] Bhandari A, Soni V, Kumar A, Singh G. Artificial Bee Colony-based Dominant Brightness Level Analysis and Adaptive Intensity Transformation
satellite image contrast and brightness enhancement technique using DWT- for Remote Sensing Images, IEEE Geoscience and Remote Sensing Letters,
SVD. International Journal of Remote Sensing; 2014, 35(5), pp. 1601-1624. 2013, 10(1), pp. 62-66.
[7] Bhandari AK, Kumar A, Singh GK, Soni V. Dark satellite image [31] Monobe Y, Yamashita H, Kurosawa T, Kotera H. Dynamic range
enhancement using knee transfer function and gamma correction based on compression preserving local image contrast for digital video camera. IEEE
DWT–SVD, Multidimensional Systems and Signal Processing; 2016, 27(2), Transactions on Consumer Electronics; 2005, 51(1), pp. 1-10.
pp. 453-476. [32] Rahman S, Rahman M. M, Abdullah Al Wadud M, Al Quaderi G. D,
[8] Bhandari, A. K., Kumar, A., & Padhy, P. K. (2011). Enhancement of low Shoyaib M. An adaptive gamma correction for image enhancement.
contrast satellite images using discrete cosine transform and singular value EURASIP Journal on IVP; 2016, 35, pp. 2-13.
decomposition. World Academy of Science, Engineering and Technology, 79, [33] Riley K.F, Hobson M.P, Bence S.J. Mathematical Methods for Physics
35–41. and Engineering: a Comprehensive Guide. Cambridge University Press,
Cambridge, 2006.
[9] Bhandari, A. K., Soni, V., Kumar, A.,&Singh, G. K. (2014b). Artificial
[34] Shanmugavadivu P, Balasubramanian K. Particle swarm optimized multi-
Bee Colony-based satellite image contrast and brightness enhancement objective histogram equalization for image enhancement. Optics & Laser
technique using DWT-SVD. International Journal of Remote Sensing, 35(5), Technology, 2014, 57, pp. 243-251.
1601–1624. [35] Somasundaram K. Kalavathi P. Medical image contrast enhancement
[10] Celik T. Spatial mutual information and page rank-based contrast based on gamma correction, International Journal of Knowledge Management
enhancement and quality-aware relative contrast measure, IEEE Transactions & e-Learning, 2011, 3(1), pp. 15-18.
on Image Processing, 2016, 25 (10), pp. 4719-4728. [36] Sørensen L, Shaker S. B, Bruijne M. Quantitative Analysis of Pulmonary
[11] Cheng F, Huang S. Efficient Histogram Modification Using Bilateral Emphysema using Local Binary Patterns, IEEE Transactions on Medical
Bezier Curve for the Contrast Enhancement. Journal of Display Technology, Imaging, 2010, 29(2), pp. 559-569.
2013, 9, pp. 44-50. [37] Sundaram M, Ramar K, Arumugam N, Prabin G. Histogram-Modified
[12] Chen S.D, Ramli A.R. Preserving brightness in histogram equalization Local Contrast Enhancement for mammogram images, 2012, Int. J.
based contrast enhancement techniques. Digital Signal Processing; 2004, Biomedical Engineering and Technology, 9(1), pp. 60-71.
14(5), pp. 413-428. [38] TAN T. L, SIM K. S, TSO C. P, CHONG A. K. CONTRAST ENHANCEMENT
[13] Chouhan R., Biswas P.K, Jha R.K. Enhancement of low-contrast images OF COMPUTED TOMOGRAPHY IMAGES BY ADAPTIVE HISTOGRAM
by internal noise-induced Fourier coefficient rooting, SIViP, 2015, 9(1), pp. EQUALIZATION-APPLICATION FOR IMPROVED ISCHEMIC STROKE DETECTION,
S255 - S263. 2012, 22, PP 153-160.
[14] Demirel H, Anbarjafari G. Image resolution enhancement by using [39] Tsai C.M, Yeh Z.M, Wang Y.F. Decision tree-based contrast
discrete and stationary wavelet decomposition. IEEE Transaction on Image enhancement for various color images. Mach. Vis. Appl; 2011, 22(1), pp. 21-
Processing; 2011, 20(5), pp. 1458-1460. 37.
[15] Demirel H, Anbarjafari G, and Jahromi M. Image equalization based on [40] Verdenet J, Cardot J.C, Baud M, Chervet H, Duvernoy J, Bidet R.
singular value decomposition; In Proc. IEEE Int. Conference on CIS, , 2008, Scintigraphic Image Contrast-Enhancement Techniques: Global and Local
vol. 1-5. Area Histogram Equalization, Eur J Nucl Med, 1981, 6, pp 261-264.
[16] Frosio I. Real time enhancement of cephalometric radiographies, Proc. [41] Wang Y, Chen Q, Zhang B. Image enhancement based on equal area
3rd IEEE Int. Sym. on Bio. Imag. Nano to Macro, Virginia, 2006, pp. 972- dualistic sub-image histogram equalization method. Consum. Electron. IEEE
975. Trans; 1999, 45(1), pp. 68-75.
[17] Gonzalez R. C, Woods R. E. Digital image processing, Pearson/Prentice [42] Wang Z, Bovik A. L, Sheikh H. R, Simoncelli R. E. Image Quality
Hall, Upper Saddle River, NJ, 3rd ed. (2008). Assessment: From Error Visibility to Structural Similarity, IEEE transactions
[18] Ganesan B, Yamuna G, Suman S. K. Hybrid Contrast Enhancement on image processing, 2004, 13(4), pp. 600-612.
Approach for Medical Image, International Journal of Computer Applications, [43] Wang Y, Wei C, Mou L, Zhang Q, Cui Z, Li X, Ye J, Lai Y. Sporadic
Proceedings on National Conference on Emerging Trends in Information & renal haemangioblastoma: Case report and review of the literature, 2013, 5(1).
Communication Technology, 2013. [44] Wan Y., Shi D., Joint exact histogram specification and image
[19] Huang and Yeh, Image Contrast Enhancement for Preserving Mean enhancement through the wavelet transform, IEEE Trans. on Img. Proc., 2007,
Brightness without Losing Image Features. ELSEVIER Engineering 16(9), pp. 2245-2250.
Applications of Artificial Intelligence, 2013, 26(5), pp. 1487-1492. [45] Yang Y, Su Z, Sun L. Medical image enhancement algorithm based on
[20] Huang S.C, Cheng F.C. Chiu Y.S. Efficient contrast enhancement using wavelet transform. IEEE Electron Lett, 2010, 46(2).
adaptive gamma correction with weighting distribution. Image Processing. [46] Voloshynovskiy S., S. Pereira, A. Herrigel, N. Baumgartner and T. Pun.
IEEE Transaction; 2013, 22(3), pp. 1032-1041. A Stochastic Approach to Content Adaptive Digital Image Watermarking.

1536-1241 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNB.2017.2771350, IEEE
Transactions on NanoBioscience

Proceedings of the Third International Workshop on Information Hiding,

pp.211-236, 1999.
[47] Zear A, Singh A K, Kumar P, A Proposed Secure Multiple Watermarking
Technique based on DWT, DCT and SVD for Application in Medicine,
Multimedia Tools and Applications, 2016, DOI: 10.1007/s11042-016-3862-8.
[48] Zhang L, Zhang L, Mou X, Zhang D. FSIM: A Feature Similarity Index
for Image Quality Assessment, IEEE Transactions on Image Processing, 2011,
20(8), pp. 2378-2386.

1536-1241 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.