Vous êtes sur la page 1sur 4

2014 Middle East Conference on Biomedical Engineering (MECBME)

February 17-20, 2014, Hilton Hotel, Doha, Qatar

Medical Image Fusion Based on Joint Sparse Method


Anuyogam Venkataraman, Javad Alirezaie, Paul Babyn and Alireza Ahmadian

Abstract— In this paper, a novel joint image fusion Sparse representation of signals is now possible utilizing
algorithm which is the hybrid of Orthogonal Matching Pursuit many different Greedy approaches [3], including: 1.
(OMP) and Principal Component Analysis (PCA) is proposed Matching Pursuit (MP) [3] 2. Orthogonal Matching Pursuit
to properly utilize the advantages and to overcome the (OMP) [3], and 3. Stage wise Orthogonal Matching Pursuit
disadvantages of both OMP and PCA methods. Firstly, (St OMP) [4]. These techniques are used to represent
common and innovative images are extracted from the source
signals with the fewest number of non-zero coefficients.
images. Secondly, sparse PCA method is employed to fuse the
information of innovative features. Then weighted average Principal Component Analysis (PCA) [5] is one of the
fusion is used to fuse the sparse PCA result with the common powerful state-of-the-art image fusion approaches in terms
feature thereby preserving the edge information and high of visual inspection and quantitative evaluation metrics.
spatial resolution. We demonstrate this methodology on This fusion is carried out by integrating the principal
medical images from different sources and the experimental components of images to be fused. Both PCA and Sparse
results proves the robustness of the proposed method. fusion have specific advantages and disadvantages. PCA
fusion will enhance the spatial quality but have dense
Key words— image fusion, orthogonal matching pursuit, nonzero entries that might represent uninformative features.
principal component analysis, sparse representation Sparse fusion preserves important information but high
spatial resolution is lacking. This paper proposes a new
I. INTRODUCTION algorithm inspired by [6], which employs different fusion
rules for common and innovative sparse components of the
Multiple images of the same scene can be captured source images. The proposed algorithm utilizes the
simultaneously using different sensors. Perceiving the advantages of both PCA and Sparse representation for
complete picture of the scene from the captured images of fusing common and innovative features of the captured
multiple sensors is not possible. Image fusion algorithms images. This algorithm also overcomes the disadvantages of
allow the combination of multiple captured images to both PCA and Sparse representation. In this paper, we
generate a more informative composite image integrating demonstrate the effectiveness of our proposed method by
the complementary information from multiple sensors, even comparing its results with PCA and Sparse Fusion.
when they are out of focus and of differing resolution.
Medical Imaging has revolutionized the medical II. METHODS
diagnosis. The arrival of imaging modalities such as
magnetic resonance imaging (MRI) and computed As discussed in the previous section, sparsest approximation
tomography (CT) tend to give different perspectives of the is achieved through Greedy methods. In this section, we
same scene which can hinder clinical decision making and briefly explore the Orthogonal Matching Pursuit algorithm
the diagnostic process. For example MRI provides the best to achieve sparsest representation. These sparse coding
view of soft tissues, while CT is better for its assessment of algorithms are constructed based on the premise that
bone structures. This motivates the need for image fusion Dictionary D of size n×k is already known. For effective
for precise diagnosis by merging the complementary results, we use phase included DCT (Discrete Cosine
information. Three challenges to be addressed while fusing Transform) dictionary [7] for our experiment. The signal
the images [1] are: 1) The fused image should preserve all 𝑆 ∈ ℝ𝑑 is sparse represented 𝑠 ∈ ℝ𝑘 , as given dictionary
the important information needed for further processing. 2) 𝐷 ∈ ℝ𝑑×𝑘 . In this paper, we have customized OMP sparse
Artifacts should not be introduced in fused image. 3) Noise coding algorithm for fusion purposes.
and unimportant information should be suppressed. The ultimate aim of OMP algorithm is to achieve best
approximation. The mathematical formula for solving this
Anuyogam Venkataraman and Javad Alirezaie* are with the constraint problem is given by:
Department of Electrical and Computer Engineering, Ryerson University, argmin𝑠 ‖𝑆 − 𝐷𝑠‖ 22 ,subject to ‖𝑠‖ 0 ≤ 𝑁 (1)
Toronto, ON, M5B2K3, Canada (e-mails: anuyogam1.venkataram@ryers-
on.ca; javad@ryerson.ca, phone: 416-979-5000; fax: 416- 979-5280). argmin𝑠 ‖𝑠‖ 0 ,subject to ‖𝑆 − 𝐷𝑠‖ 22 ≤ 𝜖 (2)
Paul Babyn is with the Department of Medical Imaging, University of Where N is the number of non-zero coefficients. Equation 2
Saskatoon Health Region, Royal University Hospital, Saskatoon, SK, S7N
0W8 Canada (e-mail: paul.babyn@saskatoonhealthregion.ca). represents the definition for solving error constrained
Alireza Ahmadian is with the Department of Biomedical Systems and problem. Next atom to be added in this iterative framework
Biophysics, Tehran University of Medical Sciences, Poursina Ave., is the atom which has highest correlation to the residual at
Keshavarz Blvd., Tehran, Iran (e-mail: ahmadian@sina.tums.ac.ir, phone:
each stage.
98-21 66466383; fax: 98-21 66466383).
(Asterisk indicates corresponding author)

978-1-4799-4799-7/14/$31.00 ©2014 IEEE 103


Fig.1. Flow of Sparse PCA joint fusion algorithm

Iteration is performed until it meets the stopping criterion. 3) Combine the sparse coefficients of innovative
OMP is due to orthogonalization between atoms in the components using PCA fusion rule, for which the covariance
dictionary D and residual r. matrix Cs of innovative images is calculated as follows,
1
Pseudo Algorithm of OMP 𝐶𝑠 = cov(𝐼𝑠 ) = cov([𝑠1 (: ), 𝑠2 (: )])= 𝐼 ∙ 𝐼𝑠 ∗ (3)
𝑖−1 𝑠

Given: Dictionary D, signal S, and error threshold ϵ Where s1, s2 are the sparse vectors of the innovative
1) Initialize residual r0=S-Ds0 , index set I0={} and main components. Find the Eigen sparse and normalized Eigen
iteration is k = k+1 (initial k=0). sparse vector of maximum Eigen value. Eigen vector will be
used as weightings for innovative sparse vectors to be fused.
4) Fused PCA result Ip is reshaped into a block of 8×8 and
2) Using the ideal solution 𝑧𝑖 = 𝑑𝑖𝑇 𝑟 𝑘−1 /‖𝑑𝑖 ‖22 , Calculate
each pixel position is the sum of several block values.
the error 𝑒(𝑖) = min𝑥 �𝑑𝑖 𝑧𝑖 − 𝑟 𝑘−1 � for all i. Reconstructed image I1 is obtained by dividing each pixel
3) Update stage: Augmenting the index set 𝐼 𝑘 = 𝐼 𝑘−1 ∪ by number of addition operations performed at each pixel.
{𝑖0 } (find i0 of 𝑒(𝑖): ∀1 ≤ 𝑖 ≤ 𝑚 and 𝑒(𝑖0 ) ≤ 𝑒(𝑖).
4) Update the solution 𝑠 𝑘 (𝑖0 )+= 𝑧𝑖 and residual. 5) For fusing common component and fused innovative
component, the fuse rule of weighted average scheme
5) If stopping criterion is met, 𝑠 = 𝑠 𝑘 ; else, apply another proposed by Burt et al. [2] is adopted.
iteration.

III. PROPOSED SCHEME IV. EXPERIMENTAL RESULTS


This work proposes a fusion method that employs PCA In order to test the performance of the proposed joint fusion
transformation and sparse transformation. An attempt is algorithm, we compared the quantitative and qualitative
made to efficiently utilize the advantages of PCA and Sparse results with two state the of art methods. Qualitative
fusion scheme. The proposed fusion framework has been measurement is done through visual inspection that
illustrated in Fig.1. Firstly, the common and innovative considers sharpness and noise suppression. Since the
components are extracted from geometrically aligned proposed joint fusion uses both PCA and sparse domain, we
multiple images of the same scene. Secondly, different use PCA, Sparse OMP fusion methods for comparison. For
fusion rules are adopted to fuse these coefficients. the evaluation, we adopted proposed joint strategy for Multi
resolution, Multi focus and Multimode images and
The algorithm is summarized as follows: compared the results with existing algorithms. Our
experiment is carried out with the assumption that the source
1) Registered multiple images in an ensemble have one images are registered already. Fig. 2, shows results of fusion
common component and multiple innovative components. for the case of multi focus images. Based on visual
Innovative components contain the complementary inspection, The Joint PCA Sparse algorithm performs the
information from different images of same scene best since the resultant image contain more geometric
i structures while sparse fusion comes the second. Result of
2) Innovative components {I i }i =1 of different images i are proposed method seems to contain sharp edges.
decomposed into sparse vectors s1,s2,s3,…si via orthogonal
matching pursuit method described in Figure.1.

104
(a) (b) (a) (b)

(c) (d) (e) (c) (d) (e)

Fig.2 Comparison of performance of different fusion Fig.4 Comparison of performance of different fusion
algorithms for case of Multi focus images (512×512) (a) algorithms for case of Multimodal medical images
Focus on right clock (b) Focus on left clock (c) Proposed (512×512) (a) CT Image ((b) MRI Image (c) Proposed Joint
Joint fusion (d) PCA fusion (e) Sparse Fusion fusion (d) PCA fusion (e) Sparse Fusion

Fig.4, illustrates the results of applying two multi modal


medical images to three image fusion algorithms. The
medical images are MRI and CT image of same scene which
have been registered already. CT image provides the
information on bone structures and MRI image contains
tissue information. Medical image fusion needs great
accuracy as it’s used for diagnosis. Hence, multimodal
(a) (b) image fusion would give sufficient details necessary for
diagnosis. Based on the visual inspection, the Joint Sparse
results contain more detail information. Results of PCA
seem to have high spatial resolution but they are
disappointing in terms of detailed information. Bone details
are not visible in PCA resultant image. Sparse result is better
but some artifacts are easily observed Reconstructing fused
image through joint fusion algorithms seems to be more
precise comparatively.
(c) (d) (e)
In order to analyze the quality of the algorithms
quantitatively, we consider 5 metrics: Mutual Information
Fig.3 Comparison of performance of different fusion (MI), PSNR, Correlation, Entropy and Structural Similarity
algorithms for case of Multi dosage images (512×512) (a) (SSIM) index. Mutual Information shows how much
Low dose Image, PSNR = 22.0994 db (b) Medium Dose information has been transferred from source images to the
Image, PSNR = 25.1981 db (c) Proposed Joint fusion (d) resultant images. Entropy shows the amount of important
PCA fusion (e) Sparse Fusion details available in the image.
PSNR is Peak Signal to Noise Ration which is used to
Fig.3, illustrates results of fusion of mutli dosage image measure the reconstruction quality of fused image. PSNR of
results based on three different fusion algorithms. Low dose the fused image If is calculated using the standard formula:
image seems to suffer from patchiness. As it can be
observed, visually, Joint Sparse result shows the details  M2  (4)
PSNR ( I f ) = 10 log10  
clearly than the Low dose image. Visually, joint fusion  MSE 
resultant seems to be better than other 2 methods. Increasing Where M is the maximum possible pixel value of the
the dosage might reduce the noise but harm patients. Low image and MSE is the Mean Square error. The SSIM [8]
dose images are prone to noise. Fusing low dose and provides structural information of objects and measures the
medium dose image should suppress the noise and enhance similarity between the two images.
the informative details for precise diagnosis.

105
TABLE I

PERFORMANCE OF FUSION METHODS BY THE QUALITY EVALUATION METRICS

Image Type Fusion PSNR(db) Mutual Information Entropy Correlation SSIM


Strategy

Multi focus Joint PCA


34.1742 2.1733 7.3656 0.9990 1.000
Fig. 2. Sparse
PCA 31.6321 2.0177 7.2607 0.9981 .9999
OMP fusion 32.3392 2.0606 7.3654 0.9981 1.000
Multi dose
Joint PCA
Fig. 3. 25.2115 0.7887 4.8643 0.9997 1
Sparse
PCA 22.6994 0.7638 4.7905 0.9991 0.9997
OMP fusion 24.4680 0.7794 4.7937 0.9995 0.9998
Multimodal (MRI & Joint PCA
26.4111 0.9634 6.7409 0.9043 .9977
CT) Fig. 4. Sparse
PCA 20.8380 0.8096 6.5502 0.8690 .9919
OMP fusion 24.8056 0.9440 6.7376 0.8985 .9975

Experiment results are shown in Table.1. Tabulated result


demonstrates the effectiveness of the proposed algorithm
over existing methods in terms of Qualitative and [2] P. Burt, R. Kolczynski, "Enhanced image capture through
Quantitative methodologies. We can observe that the results fusion", Proceedings of the 4th International Conference on
of multimodal image fusion and multi focus image fusion Computer Vision, 1993, pp. 173-182.
utilizing our proposed fusion strategy outperforms PCA and [3] L. Rebollo‐Neira and D. Lowe, "Optimized orthogonal
Sparse fusion. The PCA by itself performs poorer results for matching pursuit approach, "IEEE Signal Processing Letters,
all cases. pp.137–140, 2002.

V. CONCLUSION [4] D. Donoho and Y. Tsaig, "Sparse solutions of


underdetermined linear equations by stagewise orthogonal
Medical Image fusion plays an important role in clinical matching pursuit," Stanford University, Tech. report 2006.
diagnosis. In this paper, a joint fusion modal is proposed
based on sparse representation theory and PCA for [5] M. R.Metwalli, A. H. Nasr, O. S. Farag Allah, and S. El-
multimodal and multi dose medical images. Visually and Rabaie,”Image fusion based on Principal Component Analysis
quantitatively, the experimental results show that the and High-pass Filter”, Proceedings of the IEEE/ ICCES 2009
proposed method has effectively expressed the geometric international Conference, DEC. 14-16, 2009, pp. 63-70.
structures and edges and has proved to perform better than
PCA and OMP fusion. This modal can also be extended to [6] H. Yin, S. Li,”Multimodal image fusion with joint sparsity
fuse multiple source images from multi resolution, multiple model”, Opt Eng., 50(6), (2011).
spectral frequencies and multiple modalities.
[7] Z. Sadeghipour, M. Babaie-Zadeh, and C. Jutten, "An
adaptive thresholding approach for image denoising using
REFERENCES redundant representations", IEEE international workshop on
[1] S.G. Nikolov., P.R. Hill., D.R. Bull., C.N. Canagarajah., Machine Learning for Signal Processing, 2009, pp. 1-6.
"Wavelets for image fusion", A. Petrosian, F. Meyer (Eds.), [8] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh and Eero P.
Wavelets in Signal and Image Analysis, Computational Simoncelli, “Image Quality Assessment: From Error Visibility
Imaging and Vision Series, Kluwer Academic Publishers, to Structural Similarity”, IEEE transactions on Image
Dordrecht, The Netherlands (2001). pp. 213–244. Processing, vol. 13, no. 4, April 2004.

106

Vous aimerez peut-être aussi