International Journal of Electronics Signals and Systems

Volume 4 Issue 2 Article 1

October 2014

A NOVEL FEATURE EXPANSION ALGORITHM FOR NONRIGID CT/

MRI BRAIN IMAGE REGISTRATION
N. HARSHA VARDHAN
Department of ECE, Annamacharya Institute of Technologies and Sciences, Rajampet, A.P, India,
[email protected]

S. ASIF HUSSAIN
Department of ECE, Annamacharya Institute of Technologies and Sciences, Rajampet, A.P, India,
[email protected]

Follow this and additional works at: https://www.interscience.in/ijess

Part of the Electrical and Electronics Commons

Recommended Citation
VARDHAN, N. HARSHA and HUSSAIN, S. ASIF (2014) "A NOVEL FEATURE EXPANSION ALGORITHM FOR
NONRIGID CT/MRI BRAIN IMAGE REGISTRATION," International Journal of Electronics Signals and
Systems: Vol. 4 : Iss. 2 , Article 1.
Available at: https://www.interscience.in/ijess/vol4/iss2/1

This Article is brought to you for free and open access by Interscience Research Network. It has been accepted for
inclusion in International Journal of Electronics Signals and Systems by an authorized editor of Interscience
Research Network. For more information, please contact [email protected].
 A NOVEL FEATURE EXPANSION ALGORITHM FOR NONRIGID
CT/MRI BRAIN IMAGE REGISTRATION.

HARSHA VARDHAN.N.1, S.ASIF HUSSAIN2

1
Research Scholar, 2Assistant professor, Department of ECE, Annamacharya Institute of Technologies and
Sciences, Rajampet, A.P, India

Abstract- Image registration is one of the challenging tasks in medical image analysis. While coming to non rigid image
registration there are mainly two issues to consider. They are i) intensity similarity and ii) gray level transformation. The
issue with intensity similarity is it is not necessarily equivalent to anatomical similarity when the anatomical
correspondences between source and target images are established. Another issue is choosing an appropriate registration
algorithm. It should be robust against monotonic gray-level transformation when aligning anatomical structures in the
presence of bias fields. Here new feature- intensity based registration method developed for nonrigid brain image registration
to overcome the above stated issues named as Anatomical Region Descriptor (ARD). This method is developed on image
feature, it encodes geometric properties of anatomical structures and pixel wise interaction details. It is efficient and
theoretically monotonic gray level transformation invariant. This method is integrated with intensity based registration
algorithm named as residual complexity for Registration purpose. This proposed method is compared with three other non
rigid image registration algorithms. Experimental results of the proposed method show that it achieves the highest accuracy
rate among the compared methods.

Keywords- Component; formatting; style; styling; insert (key words)

I. INTRODUCTION ii) intensity based and iii) feature based registration

methods. In landmark based registration method first
Image registration is one of the challenging tasks in select a set of reliable landmark from input images.
medical image analysis. It has many applications in This landmark selection can be manually semi
navigation, clinical diagnosis etc..,. Registration is a automatically or even fully automatically performed.
process of aligning two images into a common Features are extracted from those landmark points.
coordinate system thus aligning them in order to The optimal transformations are then estimated based
monitor the changes between them. There are number on these extracted features.
of registration algorithms for compute transformations
to set the correspondences between the images. Image Intensity-based registration methods operate directly
registration is about determining geometrical on the image gray values, without reducing the gray-
transformation that aligns points in one view of an level image to relatively sparse extracted information.
object, with corresponding points in another view of The basic principle of intensity-based techniques is to
that object or another object. In other words, image search, in a certain space of transformations, the one
registration can be seen as a process where we try to that maximizes (or minimizes) a criterion measuring
find a mapping (transformation) between two images, the intensity similarity of corresponding voxels. Some
both spatially and with respect to intensity. measures of similarity are sum of squared differences
in pixel intensities, regional correction or mutual
Both rigid and nonrigid image registration plays a information. Mutual information has proved to be an
vital role in clinical diagnosis. Nonrigid image excellent similarity measure for cross-modality
registration plays a crucial role in medical image registrations, since it assumes only that the statistical
analysis and processing. Imaging modalities like dependence of the voxel intensities is maximal when
computed tomography (CT), magnetic resonance the images are geometrically aligned. The intensity
imaging (MRI) or positron emission tomography similarity measure, combined with a measure of the
(PET) are widely used and literally provide physicians structural integrity of the deforming scan, is optimized
with invaluable insight. The amount of data generated by adjusting parameters of the deformation field. Such
by medical imaging devices exceeds the available an approach is typically more computationally
time and concentration potential of practitioners. In a demanding, but avoids the difficulties of a feature
way, there is a need of computerized technology extraction stage.
therefore requires other new digital aids that assist
humans in evaluating acquired data. There are many Feature-based approaches attempt to find the
applications of nonrigid image registration including correspondence and transformation using distinct
anatomical analysis, image data mining, brain disease anatomical features that are extracted from images.
diagnosis. Nonrigid image registration is classified These features include points, curves or a surface
mainly into three types. They are i) landmark based, model of anatomical structures. Feature-based

International Journal of Electronics Signals and Systems (IJESS), ISSN: 2231- 5969, Vol-4, Iss-2

64
 A Novel Feature Expansion Algorithm for Nonrigid CT/MRI Brain Image Registration.

methods are typically applied when the local structure II. ANATOMICAL REGION
information is more significant than the information DESCRIPTOR
carried by the image intensity. They can handle
complex between-image distortions and can be faster, This proposed Region descriptor is extension work
since they don't evaluate a matching criterion on every from [1] [2]. This method is a 2-D joint histogram
single voxel in the image, but rather rely on a consisting features of Uniform Spherical Structure
relatively small number of features. The simplest set Pattern (USSP) and Uniform Gradient Spherical
of anatomical features is a set of landmarks. However, Pattern(UGSP). Both ussp and ugsp encodes the
the selection of landmarks is recognized to be a voxel information of first order and second order
difficult problem, whether done automatically or voxels.
manually. For many images, this is a serious
drawback because registration accuracy can be no Methodology diagram to produce an ARD Signatured
better than what is achieved by the initial selection of Image:-
landmarks. For practical reasons, the number and
precision of landmark locations is usually limited.
Hence, spatial coordinates and geometric primitives
often oversimplify the data by being too sparse and
imprecise.

Landmark based registration methods are usually

computationally efficient and includes priory known
points extracted from the selected landmarks.
However, to obtain maximum and accurate
registration results, a number of reliable landmarks are
needed, and this can bring a lot of computational
burdens additionally. Intensity based registration
methods are usually fully automatic. However,
intensity similarities may not be necessarily equivalent
to anatomical similarities. Feature-based approaches
drive the registration process based on the adopted
features. After adopting the features of images then
calculate the features in the edgeless domain. The
images registered in the edgeless domain are then
forwarded to further alignment to be performed in
intensity domain. Then intensity standardization
process is to be done to remove the effect of intensity
gray-level differences.

In non rigid image registration there are two major

challenges are there one is intensity similarity.
Intensity distributions are not fully characterize
lonely. It leads to anatomical differences between
different tissues, so there is a need of designing an
anatomical region descriptor for defining signature for
each voxel. Another challenge is volume rotation and
monotonic gray level bias fields. It generally exists
during the imaging acquisition process. So, there is
need of a registration algorithm that should be robust
against such invariants. For feature based registration Fig.1
methods rotation invariance one of its property and Fig.1 shows process of generating an ARD signature image.
this algorithm may prefer to align the bias fields
A. USSP
between the source and target images, instead of
aligning their anatomical structures if the registration USSP is developed from local binary points
algorithm is not robust against bias fields. There is a introduced in [3]. The USSP is a 2-D region
chance of degradation of the performance of descriptor that encodes voxel interaction of the first-
order image voxel interaction information. USSP
registration algorithms due to noise in the input
images. The effect of noise is overcome by using computes and preserves the rotation and monotonic
different filtering techniques during the registration gray-level transformation invariant properties of
procedure to remove the noise information from the Local Binary Points(LBP)s. This USSP has some key
physical meanings with respect to the fundamental
input images.
image structures.
International Journal of Electronics Signals and Systems (IJESS), ISSN: 2231- 5969, Vol-4, Iss-2

65
 A Novel Feature Expansion Algorithm for Nonrigid CT/MRI Brain Image Registration.

Let consider input image as Im, for each voxel v transformation invariant. Then check that acquired
belongs to Im, spherical neighborhood Sv centered at USSP is either monotonic gray-level transformation
v with radius R is defined. Nv samples are taken on invariant or not. After that an ID is determined from
the surface of Sv . Here the deterministic incremental USSP based on number of ‘0’s on surface of USSP. It
grid sequence sampling method proposed in [4] is makes USSP as rotation invariant. Then checks this
adopted as the sampling method, which is uniform USSP is feature rotation invariant. To know the
and computationally efficient. extract rotation and monotonic gray-level
transformation invariant USSP features as signatures
For each sample vi on the surface of Sv , it is for each voxel using following algorithm.
converted to binary number Bi by comparing its
intensities to the intensities of voxel v, which is Algorithm 1
expressed by Input: Input image(im) , local cubic square window
for each voxel(w), radius(r), and the number of
neighbouring samples(N).
Output: Vector image, each voxel is represented by a
…(1)
USSP signature.
where Ii denotes the intensity of sample vi , and Iv
1. w=8;
denotes the intensity of the centred voxel v of Sv .
2. r=4;
Based on equation (1), BSSP defined as follows:
3. n=3;
4. k=zeros(r,c);
Definition 1: The BSSPs are the binary threshold
5. for v1=0:size(im1,1)/w
spherical surfaces obtained by (1) for each voxel v
6. for v2=0:size(im1,2)/w
belongs to Sv.
7. sub=im1(v1+1:v1+8,v2+1:v2+8;
8.
The BSSP is monotonic gray-level transformation
9. H=zeros(1,w);
invariant as monotonic gray-level transformation
10. N=numel(H);
does not alter the relative intensity differences
11. bi1=im2bw(sub);
between the central pixel and the neighbouring pixels.
12. [q1,num]=bwlabel(bi1,r);
Therefore, the binary label assigned to each
13. For pr=1:numel(bi1)
neighbouring pixel remains the same before and after
14. pattid=0;
the monotonic gray-level transformation; thus, the
15. if bi1(pr)==0
BSSP remains the same. Therefore, we have the
16. pattid=pattid+1;
following property of the BSSP.
17. H(pattid)=H(pattid)+1;
18. else
Property 1: BSSPs are monotonic gray-level
19. H(N)=H(N)+1
transformation invariant.
20. end
Therefore, the BSSP is robust against the monotonic
21. end
gray level bias fields, which commonly exist in brain
22. h=norm(H);
MRI images; its occurrence of histogram can be
23. end
treated as the signature for each voxel. However,
24. end
there are many types of BSSPs that can make the
histogram of BSSPs too sparse to reliably reflect the
anatomical properties of input images. Therefore, The multi resolution analysis can be achieved in
only a subset of BSSPs is considered, namely, the above algorithm by using different values of radius(r)
USSP, which represents fundamental image to calculate the USSP features.
structures. The USSP is definition as follows.
B. UGSP
Definition 2: The USSPs are BSSPs that have at most Then we go for UGSP to representing second order
two continuous regions of “0”s and “1”s. voxel wise interaction information. Which is
compliment of USSP. The UGSP considers the
The USSP has some important physical meanings gradient orientations of the neighbouring voxels of
with respect to the fundamental image structures. the USSP features.
USSPs have the dominant proportions among all the
BSSPs. Let consider input image as Im, for each voxel V
belongs to Im, spherical neighborhood Sv centered at
While performing ARD it includes five V with radius R is defined. Nv samples are taken on
computational algorithms. First of all the input image the surface of Sv. Let Dvi denote the image gradient
is read and converted into BSSP. Then checks that of the i th sample Vi(i =1,2,,,,,,,Nv) on the surface of
BSSP image is either USSP or not. USSP is a subset Sv. Then a rotation invariant measure δvi is defined
of BSSP, so USSP is termed as monotonic gray-level for each neighboring voxel Vi , which is the angle

International Journal of Electronics Signals and Systems (IJESS), ISSN: 2231- 5969, Vol-4, Iss-2

66
 A Novel Feature Expansion Algorithm for Nonrigid CT/MRI Brain Image Registration.

between the orientation of τvi , and the direction of Output: Vector image, each voxel is represented by a
V- Vi . δvi can be calculated by Eq(1) shown below. UGSP signature.
1. unf=gradient(double(im1));
2. n1=numel(find(c>100));
...(1) 3. if n1>1
4. disp('input BGSP does not have gradient
information')
where | τvi |and | V- Vi | denote the magnitudes of 5. else
τvi and V- Vi , respectively. 6. end
7. [q1,num]=bwlabel(im1,r);
The measure of δvi is invariant to image rotation 8.
because no matter how the images rotate, the relative 9. for h2=1:numel(q1)-1
angle between τvi and V- Vi the direction of 10. if q1(h2)==q1(h2+1);
remains the same. 11. cou=cou+1;
12. Disp ('voxels are flagged')
After calculating δvi for each neighbouring voxel Vi, 13. else
δvi is assigned with a label by uniformly partitioning 14. end
the angle space into four subspace regions. Such 15. End
operation can be expressed by equation 2
Using this algorithm we not only determine the BGSP
and also two neighbouring voxels are considered to
be connected if and only if they have the same label
calculated by equation (2). Then the ID assignment
for the UGSP is based on the size of the largest
connected component of the UGSP.

Then we go for a new anatomical region descriptor

by combining both USSP and UGSP.
.....(2)
C. ARD
Then the basic gradient spherical pattern (BGSP) is In anatomical region descriptor both USSP and
introduced. BGSP is defined as the labelled spherical UGSP features are combined and this ARD stands as
surface obtained from the original spherical a 2D joint histogram of the both methods.
neighborhood centered at the reference voxel. UGSP
subset of BGSP. BGSP having the property of As stated in earlier sections, the USSP and the UGSP
monotonic gray level transformation invariant. This represent the first- order and second-order voxel-wise
property affects the absolute gradient magnitude of interaction information, respectively. Therefore,
each neighbouring voxel, the gradient orientation intuitively these both can provide complementary
remains the same. information individually. This leads us to design a
new anatomical region descriptor (ARD) feature. The
The UGSP is defined as:- UGSPs are BGSPs whose procedure of extracting the ARD feature as signatures
surfaces can be covered by at most two uniform for each voxel is stated in following algorithm.
regions. These uniform regions are the areas on the
surfaces of the corresponding BGSP where all the Algorithm 3
voxels belongs to the area having the same label.
Input: Input image (im), local cubic square window
The UGSP feature will be more sensitive to noise, for each voxel(w), radius(r), and the number of
and the histogram of the UGSP features will become neighbouring samples(N).
more sparse as it is more difficult for a BGSP to be a
UGSP when the number of possible labels to Output: Feature image , each voxel is represented by
represent the angle space is increased. a ARD signature.

Following algorithm describes the procedure of 1. for v11=0:size(im1,1)/w

determining whether a BGSP is a UGSP or not. 2. for v22=0:size(im1,2)/w
3. sub=im1(v11+1:v11+8,v22+1:v22+8);
Algorithm 2 4. H=zeros(1,n+2,1,n-(n/2)+1);
5. N=numel(H);
Input: Input image(im) , local cubic square window 6. bi1=im2bw(sub);
for each voxel(w), radius(r), and the number of 7. [q1,num]=bwlabel(bi1,r);
neighbouring samples(N). 8. pattid=0;
International Journal of Electronics Signals and Systems (IJESS), ISSN: 2231- 5969, Vol-4, Iss-2

67
 A Novel Feature Expansion Algorithm for Nonrigid CT/MRI Brain Image Registration.

9. for pr=1:numel(bi1)
10. if bi1(pr)==0
11. num=num+1;
12. else
13. num=n+1;
14. end
15. end
16. for pr=1:numel(bi1)-1
17. if q1(pr)==q1(pr+1)
18. s=max(max(q1));
19. id=s-(n/2);
20. else
21. id=n-(n/2)+1; Fig.2.(f).Histogram for Signatured image
22. end Fig 2.(a) shows the original image and its characteristics are
plotted by using a histogram shown in fig 2.(e)
Fig 2.(d) shows the ARD signature image and its characteristics
The robustness of the ARD is demonstrated in fig are plotted by using a histogram shown in fig 2.(f).
2.(a) and fig 2.(b) where Fig 2.(a) is a slice image
obtained from Cronnell university database without Table I
any field bias distortion. Fig 2.(b) is ARD signature IMAGE VALUES
image. Which is generated after applying the above
stated algorithms on fig 2(a). If u take a keen Parameter Input USSP ARD
observation on fig2.(b) the intensity of image is
slightly varied. All the parameters like CSF, Gray CSF 5.9985 6.8842 8.3886
value and White values of both input and signatured volume
images are tabulated in table 1. Gray 5.3870 3.2184 4.1684
volume
White 0.6115 3.6658 4.2202
volume
CSF & gray 0.8981 0.5746 0.4969
CSF & 0.1019 0.3387 0.5031
white

Table 1. shows the Parameter variations of both

USSP signature image and ARD signature image
along with input image for a corresponding image
shown in Fig 2.(a)

III. REGISTRATION PROCESS

The overall flow chart of the registration algorithm is

follows as shown below:-

Fig.2.(e).Histogram for Input image Fig. 3 shows the flow chart of a registration process.
International Journal of Electronics Signals and Systems (IJESS), ISSN: 2231- 5969, Vol-4, Iss-2

68
 A Novel Feature Expansion Algorithm for Nonrigid CT/MRI Brain Image Registration.

Where fixed image is source image and float image is square, symmetric and positive semi definite matrix
target image. That means the features of the source with eigen values ε.
image is transferred to the target image to achieve
registration. = d([l1,....,lN]) = , 1 ≥ li ≥0 ....(9)
By choosing particular regularization operator P,
Here in this paper ARD signature image is used as a
Then A is known and minimize the equation with
fixed image, and transform its features on to the float
respect to the transformation |τ.
image to produce a new image with features of fixed
image.
....(10)
Let consider two images Im and Refim to be aligned, The matrix A is of size N2, where N is a number of
assuming the following intensity relationship: image pixels. In above equation A represents a
I = J(τ ) + S + µ ..(4) covariance matrix. By....(1)
ignoring the independent terms
of τ a new similarity measure existed as shown below
Where S is an intensity correction field and µ is zero
mean Gaussian noise (note that for any two images
there always exists a correction field S). τ is the
geometric transformation that aligns Im and Refim.
The maximum a posterior (MAP) approach to ...(11)
estimate S and τ is to maximize the probability
IV. IMPLEMENTATION AND RESULTS
P(T , S| Im, Refim) α P(I, J | τ , S) P(τ )P(S) ...(5)
By define the basis eigen vectors Q and the
where we assume the independence of τ and S. The transformation τ using the FFD transformation with
first term is a joint likelihood of the images, which 3 hierarchical levels of B-spline control points. We
leads to the familiar similarity measure. P(τ) is a prior use the gradient descent optimization method to
used to regularize the transformation and P(S) is a iteratively update the transformation parameters τ.
prior on S that reflects our assumption on spatial
intensity interactions. Now, we also assume that RESULTS
pixel-wise probabilities are i.i.d., but only given the
correction field. MAP estimation in Equation (2) is
equivalent to minimization of the following objective
function:

E(S) = || Im – Refim – S||2 + δ||PS|| 2 ..(6)

where images and correction field are in column-

vector form, ||.|| is Euclidean norm and P is a
regularization operator for S. For now we omit the
transformation τ and its regularization term, to make
the derivation of S clear.

To get S let the gradient as zero.

S = (Id + δP^τ * P)-1(Im-Refim) ...(7)

Where Id is nothing but an identity matrix, P^τ * P is

a square, symmetric and positive semi defined matrix.
Substitute S along with P^τ * P in Eq(3).

After simplifying the objective functions because P is

orthogonal, and the majority of multiplications and
Fig 4. 4.(a) is a reference image, Fig 4.(b) is a float image and
inversions are only with diagonal matrices within. Fig 4.(c) is a registered image. Where the features of the float
image transforms on to the reference image and produced the
r ...(8) Registered /Deformed Image.

Here, evaluation of the proposed method by

where d( ) denotes a diagonal matrix. Where A = performing nonrigid image registration experiments
QLQ^ τ and eigen values of A are all nonnegative on the simulate database obtained from Brain web.
and bounded between [0, 1]. Here defined a new The proposed method is also compared Fig. 4.(d),

International Journal of Electronics Signals and Systems (IJESS), ISSN: 2231- 5969, Vol-4, Iss-2

69
 A Novel Feature Expansion Algorithm for Nonrigid CT/MRI Brain Image Registration.

4.(e). Typical deformation field obtained by using the first- and second-order voxel-wise interaction
proposed method. It is shown that the generated information respectively. The ARD is rotation and
deformation field is smooth and does not have folding monotonic gray-level transformation invariant, which
effects, which preserves the topology of the are properties that are analyzed and proved here. This
anatomical structures with two registration ARD signatured images then fed to registration
algorithms. They are 1) SSD and 2) Correlation process. By reducing the residual complexities of the
coefficient. In above mentioned experiments, images. Experimental results demonstrate that the
subvolume is 3 and window size ’w’ was set to 5, and proposed method has the highest registration
the number of neighbouring samples was set to 80 in accuracy among all the compared methods.
ARD algorithm. Above experiments were run on a 2-
GHz duo core central processing unit 2-GB random REFERENCES
access memory computer.
[1] S. Liao and A. Chung, “Non-rigid image registration with
uniform spherical structure patterns,” in Proc. Inf. Process.
Med. Imag., 2009, pp. 163–175.
[2] S. Liao and A. Chung, “Non-rigid image registration with
uniform gradient spherical patterns,” in Proc. Conf. Med.
Image Comput. Comput.- Assist. Intervention, 2009, pp.
696–704.
[3] Rodhe, G., Aldroubi, A., and Dawant, B., "The Adaptive
Bases Algorithm for Intensity Based Non Rigid Image
Registration", IEEE Trans. on Medical Imaging, Vol. 22,
No. 11, November 2003.
[4] Christensen, G.E., Joshi, S.C., and Miller, M.I.,
"Transformation of Brain Anatomy", IEEE Trans. on
Medical Imaging 16:6; 864-877, 1997.
[5] Rodhe, G., Aldroubi, A., and Dawant, B., "The Adaptive
Bases Algorithm for Intensity Based Non Rigid Image
Registration", IEEE Trans. on Medical Imaging, Vol. 22,
No. 11, November 2003.
[6] Y.Wang and L. Staib, “Boundary finding with prior shape
and smoothness models,” IEEE Trans. Pattern Anal. Mach.
Intell., vol. 22, no. 7,Bp. 738–743, Jul. 2000.
[7] P. Roland, C. Graufelds, J. Whalin, L. Ingelman, M.
Andersson, A. Ledberg, J. Pedersen, S. Åkerman, A.
Dabringhaus, and K. Zilles, “Human brain atlas: For high-
resolution functional and anatomical mapping,” Human
Brain Map., vol. 1, no. 3, pp. 173–184, 1993.
[8] S. Liao and A. Chung, “Multi-modal image registration
using the generalized survival exponential entropy,” in Proc.
Above figures shows the reconstructed brain images Conf. Med. Image Comput. Comput.-Assist. Intervention,
using different methods. Fig 4.(d) is a resultant 2006, pp. 964–971.
registered image of the proposed method. Fig 4.(e) is
a resultant registered image using Correlation AUTHORS
coefficient method. Fig 4.(f) is a resultant registered
image using SSD method.

For comparison purpose, the registered image of the

proposed method is also shown in Fig.4(f). It is
shown that the brain images obtained via the
proposed method preserve most of the details of the
input images, particularly in the ventricular and
cortical regions. In Fig.4.(d), 4.(e), 4.(f), significant
differences are highlighted with blue circles.
Harsha Vardhan N. received B.Tech Degree in
Electronics & Communication Engg. from
V. CONCLUSION
JNTUniversity, Anantapur, India. He is currently
persuing M.Tech Degree in Electronics &
In this paper, a new feature expansion algorithm for
Communication Engg. Specialization in Embedded
nonrigid image registration has been proposed. The
Systems from Annamacharya Institute of
proposed method is based on image features, namely
Technology & Sciences, Rajampet, A.P. India. His
the ARD. The ARD has two kinds of complementary
research interests include embedded systems,
features, i.e., USSP and UGSP, which encode the
Networking and Image Processing.
International Journal of Electronics Signals and Systems (IJESS), ISSN: 2231- 5969, Vol-4, Iss-2

70
 A Novel Feature Expansion Algorithm for Nonrigid CT/MRI Brain Image Registration.

Presently he is with Annamacharya Institute of

Technology & Sciences, Rajampet, A.P. India
working as an Assistant Professor in Dept. of ECE.
Currently he is working on a research project
sanctioned by AICTE of 11.30 lacs on biomedical
engineering. He presented many research papers in
National & International Conferences. He is a
member of Professional societies like ISTE (India),
S. ASIF HUSSAIN received B.Tech & M.Tech BMESI (India).His research interests include
Degree in Electronics & Communication Engg. from embedded systems, Time Series Analysis and
JNTUniversity, Hyderabad, India. He is currently Image Processing.
working towards PhD Degree in Biomedical
Engineering at JNTU, Anantapur, India.



International Journal of Electronics Signals and Systems (IJESS), ISSN: 2231- 5969, Vol-4, Iss-2

71