You are on page 1of 7

A Hybrid Approach for Medical Image Fusion using

SVD and PCA


Kusuma J K N Meera
Dept. of CSE, Dept. of Mathematics
Amrita School of Computing, Bengaluru Amrita School of Engineering, Bengaluru
Amrita Vishwa Vidyapeetham, India Amrita Vishwa Vidyapeetham, India
Kusumagowdaj@gmail.com kn_meera@blr.amrita.edu

ABSTRACT 1. INTRODUCTION

The process of combining the images into Fusion of images is an important task in
one image called as image fusion. This is different image processing applications
required to get the information from one where the images collected from different
image itself instead of getting the creators. The fusing of the images is done
information from multiple images and as we want the content of information in
making it as one. This also reduces the one single image rather than collecting the
randomness and redundancy which is information from many different images.
helpful in medical field and for human The images what we took as a input are
perception also. In this paper we proposed collected in varying timings and are taken
a new method called SVDPCA to raise the from different creators whose physical
information present in an image. Here we features are of different. As the images
selected two different types of images contain different features we have to
Magnetic Resonance Imaging (MRI) and extract that features and put it in one image
Positron Emission Tomography (PET) are which is helpful for both human vision and
taken for this experiment. Initially on 2 machine processing. The figure 1.1 shows
input images the SVD is applied then the how the image fusion looks.
PCA method is applied on S matrix that is
the output of SVD to get the resultant
image. To perform the experiment we used
30 image sets and are compared with other
techniques like PCA, SVD and maximum
method. The experiments show that the Input Image 1 Input Image 2 Fused Image
proposed method SVDPCA gives best
result.
Figure 1.1 Fusion of 2 images
Keywords: Singular Value Many methods have come up nowadays
Decomposition, Principal Component for this process. There are 2 fusion types:
Analysis. Magnetic Resonance Imaging, Spatial Domain (SD) and Transform
Positron Emission Tomography Domain (TD). The basic method used for
image fusion is PCA which belongs to SD
method. It is well known for feature
extraction and reducing the dimension and
is used for image fusion. SVD is another 2. Then from the covariance matric find
technique to get the features from an the eigenvectors and eigenvalues.
image which is used in different purposes 3. The eigenvector with highest eigenvalue
like recognition of the face, image fusion, represents the direction of the highest
registration of the image etc [1][2]. Here a variance. This helps in finding the 1st
new technique is induced by fusing SVD Principal component. The eigenvector with
and PCA. For experiments we have taken next eigenvalue represents the remaining
30 image sets and implemented on existing variance direction and it is orthogonal to
methods like PCA, maximum method and the 1st direction. So this helps in finding
SVD. These results are compared with the the 2nd Principal component. Like this we
proposed method. The results discussed in can find the n PC’s based on our
this paper show that the new induced requirement.
method gives the best result. The paper is 4. Finally the fusion is done using the
taken in the following way: Section 2 below equation:
presents the PCA, SVD and Maximum FusedImage=PC1*Image1+PC2*Image2
method. Section 3 gives the hybrid method
SVDPCA. Section 4 presents the result
analysis and Section 5 gives the 2.2 Singular Value Decomposition
conclusion of the work. SVD is a method to take out the features
present in the image. SVD on a matrix will
2. DIFFERENT FUSION METHODS
decompose the matrix into 3 matrices. In
2.1 Principal Component Analysis that 2 matrices are singular matrices and 1
matrix is diagonal matrix [2]. The singular
PCA is a popular technique for analyses of matrix gives description of the images like
large datasets which have high number of feature stability, scale invariance etc.
features for each observation, increasing When SVD is applied on a matrix A which
the understandable of data while keeping is of m x n will result as follows:
the maximum amount of information. It is
a statistical technique for dataset A = UASAVA T
dimension reduction. This can be done by
linearly transforming the data into new where UA is left singular vectors which
coordinate system where data variation is will be of size m x m, VA T is right
described with few dimensions than the singular vectors which is of size n x n and
initial data. Many authors have taken 1st 2 SA is diagonal matrix which is of size m x
PC’s to plot the data in 2-D and to visually n, the diagonal elements are called as
see the clusters of closely related data singular values, which is arranged in
points. It has many applications in many decreasing order. Below is the equation
areas like micro biome studies, which gives SVD of a matric:
atmospheric science and population U S V= SVD(A)
genetics.
For fusion of images we applied SVD on 2
Steps followed in PCA: images but the SVD accepts only two
1. Find the covariance matrix of an input dimensional image. So the coloured
image. images is decomposed into 3 channels that
is red, green and blue so we got it as three Same procedure is done for Green channel
2-D matrices from one single image. Same and Blue channel. Same way fuse the left
procedure for 2nd input image also, there singular vectors and right singular vectors
also we will get three 2-D matrices. So it of particular channel. Then we will end up
will look like: with 3 matrices of each channel. Finally
combine all 3 channels to get the colored
Input image1 fused image[2]. Fusing of all 3 channels is
done as below:
Fusedimage(:,:,1)=red channel matrix
Fusedimage(:,:,2)=green channel matrix
Matrix 1 Matrix 2 Matrix 3 Fusedimage(:,:,3)=blue channel matrix
Red channel Green channel Blue channel So finally the Fused image can be
displayed as imshow(Fusedimage).

Same way the input image 2 also will have


3 matrices. Now SVD is applied on each 2.3 Maximum Method
matrix[3]. So total 6 matrices so 6 times
In this method, the maximum intensity
SVD is applied. Then combining the
pixel is selected to put in the fused image.
image1’s matrix1 (red channel) singular
This is done in each channel of the image.
matrix with image2’s matrix1 (red
For this each pixel of input image 1 is
channel) singular matrix to get singular
compared with the corresponding pixel of
matrix of fused image for red channel.
the input image 2 and the pixel whose
intensity value is maximum that we put in
the fused image, which results in, the
fused image’s every pixel is with
maximum intensity. The general
architecture diagram for all fusion methods
is shown in Figure 2.1.

Image RGB to
Preprocessing Gray

Fusion Gray
Input Image 1 to
Technique
RGB

RGB to Fused Image


Image
Preprocessing Gray

Input Image 2
Figure 2.1 General architecture diagram for all fusion methods.
3. PROPOSED FUSION METHOD 4. EXPERIMENTAL RESULTS

In this paper a new method is approached We took 30 PET and 30 MRI Input
called SVDPCA. A hybrid technique by images for experiment. We implemented
combining the SVD and PCA methods to existing methods also for comparison. For
get the best result compared to the existing measuring we took PSNR, Cross
methods mentioned in above section. Here correlation and Structural similarity
we applied SVD on input images as parameters. Out of 30 image sets 23
explained in above section. After applying image sets showed that the proposed
the SVD method we will be having 6 method is best. So we got the accuracy
singular matrices, 3 for input image 1 and 76.66%. The table 4.1 shows the result for
3 for input image 2. Then combined the 2 image sets for all fusion methods we
singular matrices of corresponding discussed in this paper. The highest values
channels[8]. Then we had 3 singular for each parameter are represented in bold
matrices. Upon that we applied PCA: In case. From the table we can see that our
PCA except the first step that is without proposed method is best. Figure 4.1 and
finding the covariance matrix the rest of 4.2 shows the output of different fusion
the steps are applied on three singular methods used in this paper with proposed
matrices what we got in previous step[9]. method output also.
At last we will be having three matrices for
each channel of fused image. Finally
combine all these channels as said in
section 2.2 to display the fused
image[13][14]. The Flow diagram of the
proposed method is shown in Figure 3.1.

Image Convert 3D Apply SVD on


Input Image1
Preprocessing to 2D each channel

Image Convert 3D Apply SVD on


Input Image2
Preprocessing to 2D each channel

Fuse singular matrices of each


Apply PCA method on singular channel correspondingly to get
matrices of each channel colored Image

Finally combine
all 3 channels Display fused image
Image2
Figure 3.1 Flow diagram of proposed method.
Table 4.1 Comparative Analysis.

Methods PSNR CC SSIM Val


PCA 11.8864 0.4743 0.5888
Maximum Method 11.0156 0.6867 0.7422

Set 1 SVD 15.6421 0.7208 0.745


SVDPCA(Proposed Method) 18.8252 0.8194 0.9747
PCA 13.7588 0.7118 0.6492
Maximum Method 10.5472 0.6619 0.6208
Set 2
SVD 15.9433 0.8061 0.7256
SVDPCA(Proposed Method) 17.6964 0.8599 0.9177

Input Image 1 Input image 2 PCA

Maximum method SVD SVDPCA

Figure 4.1 Results for Set I image


Input Image 1 Input Image 2 PCA

Maximum Method SVD SVDPCA

Figure 4.2 Results for Set II image

algorithms”, in: Proceedings of the IEEE-


5. CONCLUSION National Aerospace and Electronics
In this paper, initially the input images are Conference, pp. 231 235, 2009.
pre-processed then each image is divided [2]. H.Nasir, etal. “Singular value
into three matrices to make it two decomposition based fusion for super
dimensional matrix. Then SVD is applied resolution image reconstruction”, Signal
on each matrix for both the images. After Processing-Image Communication, 2011.
that the singular matrices are combined [3]. Indeevar Thakur , Hardeep Saini, “A
correspondingly for each channel. Then Review: Analysis of SVD based Image
PCA is applied and finally we got the Fusion Methods” , International Journal of
fused image. The results also show that the Engineering Research & Technology
fusion done by using the proposed method (IJERT) IJERT ISSN: 2278-0181
is best. The objective of fusion is to get the IJERTV3IS050981 www.ijert.org Vol. 3
more data from a single image. That we Issue 5, May - 2014 .
achieved here using the proposed method. [4]. Souparnika Jadhav, “Image Fusion
Based On Wavelet Transform”,
REFERENCES nternational Journal of Engineering
Research ISSN:2319-6890)(online),2347-
[1]. D.W.Repperger, A.R.Pinkus, K.A.
5013(print) Volume No.3, Issue No.7, pp :
Farris, R.G.Roberts, R.D.Sorkin,
442-445 , 2014.
“Investigation of image fusion procedures
[5].S hifali M Patil, “Image Fusion using
using optimal registration and SVD
Wavelet Transform”, International Journal
of Engineering and Advanced Technology
(IJEAT) ISSN: 2249 – 8958, Volume-5, Enhancement Technique for Clinical
Issue-4, April 2016 . Diagnosis”, in 2019 3rd International
[6]. Mostafa Amin-Naji, Pardis Ranjbar- Conference on Computing Methodologies
Noiey, and Ali Aghagolzadeh, “Multi- and Communication (ICCMC), Erode,
focus image fusion using Singular Value India, 2019.
Decomposition in DCT domain”, 10th [11] Ameel Ahmed Bhutto , Tian Lianfang
Iranian Conference on Machine Vision and (Senior Member, Ieee), Qiliang Du,
Image Processing, Nov. 22-23, 2017. Toufique Ahmed Soomro, “An Enhanced
[7]. Tian Lianfang D u Qiliang, Jameel Image Fusion Algorithm by Combined
Ahmed Bhutto, Bhawani Shankar, “Multi Histogram Equalization and Fast Gray
Focus Image Fusion using Combined Level Grouping Using Multi-Scale
Median and Average Filter based Hybrid Decomposition and Gray-PCA”, August
Stationary Wavelet Transform and 20, 2020, volumn 8 September 9, 2020.
Principal Component Analysis ” , [12] R. Revathy, S. Kumar, V. Achyuth,
(IJACSA) International Journal of V. Reddy, V. Bhaskar, and Bhavana V.,
Advanced Computer Science and “Medical Image Registration Using
Applications, Vol. 9, No. 6, 2018. Landmark Registration Technique and
[8] P. Prasad, Subramani, S., Bhavana V., Fusion”, in Computational Vision and Bio-
and Krishnappa, H. K., “Medical Image Inspired Computing, Cham, 2020.
Fusion Techniques Using Discrete [13].Harpreet Kaur1· Deepika Koundal2 ·
Wavelet Transform”, in 2019 3rd Virender Kadyan3, “Image Fusion
International Conference on Computing Techniques: A Survey”,Springer, Archives
Methodologies and Communication of Computational Methods in Engineering
(ICCMC), Erode, India, 2019. (2021).
[9] P. Sridhar and Sathiya, R. R., “Noise [14]. Multi-layer, multi-modal medical
standard deviation estimation for additive image intelligent fusion Nair, R.R., Singh,
white gaussian noise corrupted images T., asavapattana, A., Pawar,
using SVD domain”, International Journal M.M.Multimedia Tools and Applications,
of Innovative Technology and Exploring 2022 I.F 2.853
Engineering (IJITEE), vol. 8, 2019
[10] S. Asish K Reddy, K Kumar, K., K.
Kumar, N., Bhavana V., and Krishnappa,
H. K., “Multimodal Medical Image Fusion

You might also like