A Novel Technique for Medical Image Fusion using

improved Contourlet Transformation with modified

DFBs Jaspreet kaur

Computer science and Engineering

Lovely Professional University

Phagwara, Punjab

Shefaly Sharma

Computers Science and Engineering

Lovely Professional University

Phagwara, Punjab

Shipra Gupta

Computer Science and Engineering

Lovely Professional University

Phagwara, Punjab

jaspreetkaurhunjal@gmail.com shefalysharma20@ymail.com er.shipra256@yahoo.com

Abstract— Medical Image fusion plays an critical and vital role

in today’s medical field to diagnose particular disease and this

concept is still under research. Image Fusion is a process in which

we merge two or more images captured by different instruments

or modalities to produce single image that provide accurate and

clear information without any degradation in quality of image.

Fused image will be analyses by doctors to diagnose the disease in

particular patient. In this paper, we proposed novel technique

which is based on contourlet transformation. We proposed

improved contourlet in which pyramid decomposition is replaced

by multiscale decomposition with improved DBFs: replace low

pass filter and high pass filter with log Gabor filters. This novel

technique not only sharply localizes image also eliminates DC

component to provide accurate information which is main

feature of log Gabor filter. In this paper, we are fusing CT and

MRI modalities images into single image without any

degradation in quality. We are considering registered images

only (same size and same alignment). Performance of new

technique is evaluated by quality measures like EN (entropy),

STD (standard deviation), MI (mutual information), Quality

index and PSNR.

Keywords- Contourlet transformation; Multimodality; Log

Gabor Filters; MIF; Multiscale decomposition; DFBs

I. INTRODUCTION

Image fusion plays an important role in day-to-day

applications. Image fusion is used to enhance the quality of a

degraded image.It is defined as process of combining two or

more images into single image that contain information of

both images without any loss of information. This information

can be used for different purposes like human visual and

machine perceptions, diagnosis, analysis. There are basically

two type of image fusion system. Single sensor image fusion

system in which sensor image fusion system in which we use

one sensor that capture multiple images and then fusion them

into one image. Imaging sensor has limited capability and

does not work in bad environment like clicking an image in

fog and rain. Second, Multisensor image fusion system in

which we use multiple sensor devices mainly we use infrared

cameras which has capability to work in bad environmental

conditions [10].In this paper, single sensor image fusion

system is used. Images are captured by single sensor device

and then they get fused. There are different modalities in

medical field like CT, MRI, X-ray. These modality images are

captured by single sensor device which is located within single

room and it does not get affected by weather conditions

because they are established inside room. So far, various

researchers proposed many IF methods for better results.

Generally there are three levels of fusion: pixel level; feature

level; decision level.

Figure. 1 Basic levels of image fusion

The main goal of pixel level fusion is to enhance the raw input

images and provide an output image with more useful

information than either input image. Pixel level fusion is

effective for high quality raw images. Feature level fusion is

done in case of feature extraction. In feature level fusion, the

information is extracted from each input image separately and

then fused based on features from input images. Decision level

fusion is effective for complicated systems with multiple true

or false decisions but not suitable for general applications

because in this the information is extracted from each input

image separately and then decisions are made for each input

channel. There are some issues that have to be dealt with

before doing fusion. Most of the time images that are captured

by different sensor are misaligned. For fusion we need to

establish correspondence between the sensor images, for this

process we use Registration technique. Registration techniques

align the images by exploiting the similarities between sensor

images. There are various algorithms that are available to

register images [10].

Image

Registr

-ation

Feature

Extract-

ion

Attribute

Descript-

ion

Image

Explan

-ation

Pixel

level

Feature

level

Decision

level

IEEE - 31661

4th ICCCNT 2013 July 4-6, 2013, Tiruchengode, India

A. Image Fusion Techniques

Image fusion techniques can be categories into two

domains: spatial domain and frequency domain. In spatial

domain we covert an image in the form of matrix and then

process it.

G (x, y) = T [ f ( x , y ) ]

where f(x, y) is an input image.

G(x, y) is the processed image.

T is operator on image pixel (x , y )

In frequency domain images are represented in form of signal

or sub bands. There are many methods which are based on

spatial domain like HIS (hue intensity saturation) method,

average method, PCA (principle component analysis), Brovey

method. These methods give better results but suffer from

spectral degradation. Here in this paper, we are working in

transform domain or frequency domain. Further there are three

main types which are discussed below

Wavelet Transformation

Wavelet transform has been greatly used successfully in

many areas, such as texture analysis, data compression, feature

detection, and image fusion [5].Wavelet transformation is

basically used to decompose an input image into sub bands.

We can apply this transformation in both spatial and frequency

domain. But it is always preferred to apply wavelet in

frequency domain in which images are decomposed into

frequency sub bands. Multi resolution analysis is done in

wavelet transform.

Figure. 2 Wavelet Decomposition

1, 2 are decomposition levels, H stands for high frequency

bands , L stands for low frequency bands.

Laplacian Method

Laplacian pyramid method is based on MRA

(multiresolution analysis). In this method, input image is

decomposed in form of pyramids. Decomposed image is

represented into two pyramids: First, Smoothing pyramid

contains the average pixel values. Second, Difference pyramid

contains the pixel differences that are edges. This pyramid is

also known as multi resolution edge representation of the input

image [6]. The basic idea behind this is to perform a pyramid

decomposition on each source image, then integrate all these

decompositions to form a composite representation, and

finally reconstruct the fused image [7].

Contourlet Transformation

Contourlet transformation is based on pyramid method. It is

recently developed transformations used to decompose an

image. This method implemented in two stages, first stage is

transformation and second stage sub band Decomposition [4].

Image is decomposed into different scale by laplacian

pyramids and then DFBs are applied to each decomposed

images. An image gets divided into low pass and high passes

sub bands. Filtering of these bands is done by directional filter

banks (DFBs).This is a part of a multi-dimensional signal

processing theory called as Multi scale Geometric Analysis

(MGA) Tools [4]. Major drawback of DWT in two

dimensions is their limited ability in capturing directional

information. to overcome this drawback contourlet

transformation was introduced. Contourlet not only possess

the main feature of DWT but also offers the high degree of

directionality.

PYRAMID DFB

METHOD

Figure. 3 Flow of Contourlet transformation

In this flow, the image is divided into different scale and

different frequency through Laplacian Pyramid. Then, as the

output of the Laplacian Pyramid, the sub-bands of image in

different scale and frequency are fed into directional filter

bank. In this way, directional information can be captured

effectively [1]

Figure. 4 Contourlet transformation with two levels of multi

scale Decomposition[9]

B. Modified DFBs

Original DFBs consist of low pass filters and high pass

filters but we are modifying DFBs by replacing low pass high

pass filters with Log Gabor Filters. The “Gabor Filter Bank” is

a popular technique used to determine a feature domain for

IMAGE

LL2 LH

2

LH 1

HL2 HH

2

HL 1 HH

1

(2,2)

(2,2)

IEEE - 31661

4th ICCCNT 2013 July 4-6, 2013, Tiruchengode, India

representing an image [3]. The main limitation of Gabor filter

is that, it was designed for a bandwidth maximum of 1 octave.

So, Field introduced Log Gabor Filter to overcome this

limitation. This filter is used for localizing the spatial and

frequency information of an image. A Log Gabor Filter has no

DC component and can be constructed with any arbitrary

bandwidth. Two important characteristics of Log-Gabor filter,

First, the Log-Gabor filter function always has zero DC

components which not only improve the contrast ridges but

also edges of images. Second, the Log Gabor filter function

has an extended tail at the high frequency end which allows it

to encode images [3].

Noisy Image

Gabor filter bank

Gabor Coeffcient

…..…... ……..

Gaussian Mask

……….. ………..

Gabor features

Figure. 5 Flow of Log Gabor Filtering Process

Gabor filter consist of Gaussian function. It is based on

logarithmic transformation which eliminates the DC

component allocated in medium and high pass filters. We use

Fourier transformation in log Gabor filter. This transformation

decomposes an image into sin and cosine component. Major

applications of FT (Fourier transformation) are image

compression, image analysis and image filtering.

G(ρ,θ,p,k) = exp(-1/2(ρ-ρk/σρ)2)exp(-1/2(θ-θk,p/σθ)2)

Log Gabor filters consist of complex-filtering arrangement in

p oritentations and k scales, whose expression in the Log-polar

Fourier domain is given in above equation. In this, (ρ, θ) are

the log-polar coordinates and (σρ,σθ) are the angular and

radial bandwidths(common for all the filters).p and k represent

the orientation and scale selection respectively.

C. Medical Image Fusion

Medical images are used by doctors to get information

about the patient body and to detect any disease. There are

several types of diagnostic tools available to capture an image

of patient’s body. Different types of medical images like X

rays which is used to detect broken bones. CT (computer

tomography) provides us structure of bones in patient body.

These images are used to provide us detail information

regarding body parts. MRI is used to get information about

soft tissues, organs in human body. It does not use X rays and

radiations that harm to our body. Ultrasound is a safe and less

harmful image processing, in which high frequency sound

waves are used.PET that stands for positron emission

tomography. Medical imaging is an important term in field of

research, diagnosis and treatment. Multimodality medical

image fusion is a process to merge images that are captured by

different modalities such as CT, MRI and PET and which

results in more useful information for diagnosis or doctors

regarding patient’s body. Various fusion techniques has

attracted more researchers in this domain to assist the

physicians in fusing and retrieving information from different

modalities such as CT, MRI, PET, SPECT and so on[4].

II. PROPOSED SCHEME

In this paper, we proposed an efficient fusion technique based

on contourlet transformation, in which we are using multi

scale decomposition and improving the DFBs by replacing

low pass and high pass filter with Log Gabor filter, which

reduces the DC component in fused image. Most important

consideration: medical images which we are fusing must be

registered images that are images should be same in size and

of same alignment. Algorithm is given below:

1. Select two registered images

2. Apply Multi scale decomposition on both images and

DFBs on decomposed images.

3. Apply multi scale decomposition with modified DFBs (Log

Gabor filters) for both images

4. Merge the frequencies.

5. Resultant single fused image. 6. Calculate PSNR and Standard Deviation for fused image.

III. EXPERIMENTAL RESULTS

Figure. 4 GUI for Image fusion

I

RT Ri R2 R1

F1 F2 Fi FT

IEEE - 31661

4th ICCCNT 2013 July 4-6, 2013, Tiruchengode, India

Table 1: Fusion of medical images using contourlet

transformation.

Original image1 Original image2 Fused image

Standard Deviation

56.6812

PSNR

23.8897

Table 2: Fusion of medical images using improved contourlet

transformation.

Original image1 Original image2 Fused image

Standard Deviation

83.2753

PSNR

41.2502

To evaluate the performance of our proposed scheme we

calculate following factors :

A. Peak Signal to Noise Ratio :

PSNR computes Peak Signal to Noise Ratio, in decibels,

between two images. This ratio is used to provide the quality

measurement between two resultant fused images. Efficiency

of particular technique is calculated by calculating

PSNRvalue. Higher the PSNR more will be the quality.

PSNR = 2

1010log 255 / MSE (1)

Where MSE is calculated as:

MSE =

1

0

,

, 0

12([ ( , ) ( , )] / *

M N

i j

OI i j DI i j M N

(2)

Where OI is fused image with contourlet transformation,DI is

fused image with improved contourlet transformation, M*N is

number of pixels.More the value of MSE lower the value of

PSNR.

B. Standard deviation:

STD is used to measure the contrast in fused image. An

image with high contrast would have high Standard deviation

value.

STD = =1 =1

21/ ( ( , ) )M N

m n

MN F m n MEAN (3)

Where M*N denotes the size of an image and f(m,n)indicates

the gray-value of the pixel of image f at position(m,n) and

MEAN =

=1 =1

1/ ( , )M N

m n

MN F m n [8] (4)

C. Entropy:

EN of an image is a measure of information content. It is

average number of bits needed to quantize the intensities in

the image. It is defined as

EN =

1

2

0

( ) log ( )L

g

p g p g

(5)

Where p(g) is probability of g gray level g and the range of g

is [0… L-1]. An image with high information content would

have high entropy. If the entropy of fused image is high then

parent images then it indicates that the fused image contain

more information[8]

D. Mutual Information:

MI measures the degree of dependence of the two images.

A large measure implies better quality. Given two images xF

and xR; MI defined as:

MI = I (xA; xF) + I (xB;xF) (6)

Where I (xR;xF) =1

, 2 ,

1

( , ) log ( , ) / ( ) ( )L L

R F R F R R

u v

h u v h u v h u h v

(7)

Where hR,hF are the normalized gray level histograms of xR

and xF respectively. hR,F is the joint gray level histogram of xR

and xF and L is the number of bins. xR and xF correspond to the

reference and fused images respectively. I(xR ;xF ) indicates

how much information the fused image xF conveys about the

reference xR. Higher the mutual information between xF and

xR [8].

After the experiments performed on registered Images in

MATLAB, we have realized that above factors determine the

quality of fused image. Our technique is better than those of

other techniques of image fusion, because this technique

provides high quality of information. The quality of image is

measured by three factors which are given in following table:

IEEE - 31661

4th ICCCNT 2013 July 4-6, 2013, Tiruchengode, India

Table 3: Calculation of PSNR, Standard deviation and Quality

index, entropy values by applying contourlet and improved

contourlet transformation

IV. CONCLUSION

Our proposed scheme is applied on medical images. There

are many techniques have been proposed for image fusion;

like wavelet transformation, laplacian pyramid method,

contourlet transformation, curvelet transformation. The

proposed fusion technique is evaluated on the multimodality

medical images to fuse them with better quality so that there is

no degradation of quality of image. And can be send to

doctors without any loss of information with better quality.

From series of experiments we have concluded that our

proposed technique is yielding higher PSNR value then other

techniques mentioned above.

V. FUTURE SCOPE

A lot of work has to be done to fuse the image in order to

improve the quality of the image fusion. In future, the

implementation can be done on the hybrid scheme that can

handle the 2D and 3D image fusion so to achieve this, more

work can be done with our proposed scheme to have higher

PSNR so that reconstructed fused image will be of higher

quality and useful results can be produced. We can also

implement registration techniques so that unregistered images

can also be fused by applying these techniques.

REFERENCES.

[1] Mo Tingting, Shi Yuehua, Wei Yipeng, ZhaoFeng,

ZhuYongxin,(2009),”Implementing Contourlet Transform for

Medical Image Fusion on Heterogenous Platform”, IEEE, Pp

115-120.

[2] Thanushkodi K., Umaamaheshvari A. ,(2010),“Image fusion

techniques”, IJRRAS 4 (1).

[3] R. Redondo, F. Šroubek, S. Fischer, G. Cristóbal,

(2008),“Multifocus image fusion using the log-Gabor transform

and a Multisize Windows technique”

[4] Kavitha S, Rajkumar S, (2009),“Redundany Discrete Wavelet

Transform and Contourlet Transform for Multimodality Medical

Image Fusion with Quantitative Analysis”,IEEE Pp 134-139

[5] Fang Zhijun, Huang Shuying, Park Dong Sun, Yang Yong,

Wang Zhengyou,(2009),”Wavelet based Approach for Fusing

Computed Tomography and Magnetic Resonance Images”,

Chinese Control and Decision Conference (CCDC 2009) Pp

5770-5774.

[6] Singh Nupur , Tanwar Pinky,(2012),”Image fusion using

improved Contourlet Transform Technique”,International

Journal of Recent Technology and Engineering (IJRTE) ISSN:

2277-3878, Volume-1, Issue-2,Pp 131-139

[7] Kaur Jaspreet , Sharma Chirag (2012)”An Efficient technique of

multimodality medical image fusion using contourlet

transformation”, (IJITEE), Vol 2 ,issue 1,Pg 28-30.

[8] Das Sudeb, Kundu Kumar Malay (2011)”Fusion of

Multimodality Medical images using Combined Activity level

measurement and Contourlet Transformation.”(ICIIP),

[9] Do Minh N. and Lu Yue, (2010),“A new contourlet transform

with sharp frequency localization”,National Science Foundation

under Grant CCR-0237633.

[10] Canga Eduardo Fernández, Signal and Image Processing

Group,(June 2002), “image fusion”.

Quantitative

Measures

Technique

Contourlet

transformation

Improved Contourlet

transformation

MI

4.1131

4.2937

EN

0.95352

1.0371

STD

56.6812

83.2753

PSNR

23.8897

41.2502

Quality Index

38.8031

76.2091

IEEE - 31661

4th ICCCNT 2013 July 4-6, 2013, Tiruchengode, India