An intensity based medical image registration using genetic algorithm

Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.3, June 2014
DOI : 10.5121/sipij.2014.5305 53
AN INTENSITY-BASED MEDICAL IMAGE
REGISTRATION USING GENETIC ALGORITHM
Shanmugapriya.S1
, Dr.S.Poonguzhali2
, Umamaheshwari3
Department of Medical Electronics, CEG, Anna University, Chennai, India
ABSTRACT
Medical imaging plays a vital role to create images of human body for clinical purposes. Biomedical
imaging has taken a leap by entering into the field of image registration. Image registration integrates the
large amount of medical information embedded in the images taken at different time intervals and images
at different orientations. In this paper, an intensity-based real-coded genetic algorithm is used for
registering two MRI images. To demonstrate the efficiency of the algorithm developed, the alignment of the
image is altered and algorithm is tested for better performance. Also the work involves the comparison of
two similarity metrics, and based on the outcome the best metric suited for genetic algorithm is studied.
KEYWORDS
Image registration, Genetic algorithm, Similarity metric
1. INTRODUCTION
Image registration aligns the information in different images and to visualize as a combined
image. Image registration operates for different conditions such as images taken at different time,
view point and/ or different sensors [3]. The goal of image registration is to find the best
geometrical mapping to align the images. This geometrical alignment of medical images has its
rapid growth in diagnosis, treatment planning, computer-aided therapies and in biomedical
research. Image registration falls under two categories- voxel or intensity - based and feature –
based. In intensity-based method, the whole image is employed; where they operate directly on
image intensity. In feature-based method, features like contours, crest lines, points- have to be
extracted from the images in pre-processing step and has to be matched.
Intensity-based method works better than feature-based image registration [1][2]. Image
registration techniques generally depends on many criteria such as, nature of registration, ie,
extrinsic or intrinsic based, transformation model and optimisation procedure [4] .The most
applicable algorithms for medical image registration are based on optimising voxel-based
similarity measure. Surface registration techniques focus on three factors, namely, choice of
transformation, similarity criterion and global optimisation [6]. Many metrics such as, sum of
squared differences, sum of absolute differences, mutual information, normalised correlation can
be used for intensity-based medical image registration [7]-[9]. The drawbacks of traditional
numerical based optimisation algorithms like, gradient descent, Powell’s method, can be

54
overcome by genetic algorithm. It finds a global optimal solution, escaping from local optima
[5][10].
In this paper, image registration is done by genetic algorithm. The most crucial step in GA is to
choose the objective function which is used to evaluate the fitness of each chromosome. The
fitness value is determined by the similarity metrics which is the objective function. Here
normalised correlation is used as the objective function for GA process. The result obtained by
normalised correlation is compared with mutual information method which is estimated by mean
squared error.
2. METHODOLOGY
Fig1 shows the flow of image registration process. Image registration process involves two MRI
images, namely, reference image IR, and a scene image IS, which has to be aligned in accordance
with the reference image[1].
Fig1. Flow of image registration process
The result of registration process is that the reference IR and transformed scene image f(IS) are
similar possible.
Of those two images, one image is rotated to 180 degrees and this is taken as scene image.
Similarity transformations, ie, translation, rotation and scaling are applied to scene image. The
similarity between the reference and transformed scene image are determined by normalised
correlation which is the similarity metric. If A(x,y) and B(x,y) are two images, Normalised
correlation(NC) is defined.
(1)
Where n= number of pixels in two images, ‫,ܣ‬ ‫ܤ‬ are average intensity values.

55
2.1 Genetic Algorithm
If the best transformation is not obtained, an optimiser is used to find the optimal transformation.
Genetic algorithm is used to find the best transformation. The search strategy used is parameter-
based method [1][2]. The crucial step in GA is to encode the parameters into GA chromosomes.
Real encoding is used than binary encoding. Efficiency of GA is increased since there is no
conversion to binary type and different genetic operators can be used [11]. So the encoding
involves the transformation parameters directly. It is represented by five real numbers, namely
one rotation parameter (τ ), two translation parameters (t), and two scaling parameters (s). GA
cycle follows four process- fitness evaluation, selection, recombination and creation of new
population. Initial population is set to 50 chromosomes. Normalised correlation is used to find the
similarity for each transformation and evaluation of this objective function called fitness value is
calculated for each chromosome in the population. Next step involves selection of fittest
individuals. Tournament method requires more execution time and hence Roulette’s wheel
method is used. Highest fit individuals are selected and genetic operators such as; crossover and
mutation are applied to create next generation population. Here arithmetic crossover is used.
P = α*A + (1.0-α)*B (2)
Q = (1.0-α)*A + α*B (3)
Where, A, B - individual from the old generation, P, Q - individual from new generation, α is the
weight which governs dominant individual in reproduction and it is between 0 and 1.
So the succeeding generations will contain the offspring and also the best fit individuals from its
preceding generations. Mutation operators randomly pick one of the individual in the
chromosome and replace it by a random value in its range. Fitness value is calculated for the next
generation individuals. This process is continued for many generations. Stopping criteria chosen
is maximum generation. It is likely that as generation increases, best fitness, i.e. best
transformation can be obtained. The process is continued until fitness value reaches high so that
the registration completes. The global optimum is achieved by creating new population from
individuals of existing population.
3. EXPERIMENTAL STUDY
For this experiment, simulated real world MRI images were taken from a public database. Images
were obtained from brain web database at McGill University, which can be easily used to
evaluate the performance of various image analysis methods. It has been frequently used by
biomedical research community. The images taken are with and without noises, and presence or
absence of lesions. The images are at the size of 181×217 pixels.
Table 1: Types of images used
Cases Reference image Scene image
I Normal image Normal image with 180
degree rotated
II Normal image Lesion with 1% noise
III Normal image Lesion with 4% noise
IV Normal image Normal with noise 3%

56
3.1 Parameter settings
Transformation parameters are applied on the pair of images. The translation parameter, rotation
parameter and scaling parameter are noted as tx, ty, τ, Sx, Sy respectively. Rotation is taken as 0 to
2π, translation is taken as [100 300] and scaling as [0.5 2].
Initial population was taken as 50. Crossover probability is 0.6 and mutation probability is 0.3.
Stopping criteria is taken as the maximum generation. Maximum of 300 generations were done.
3.2 Analysis of results
Two pair of MRI images was taken. For each pair of images, GA was performed with NC and
with mutual information (MI). Fitness using NC is calculated by maximising the similarity
metric. For each generations, maximum of 10 runs were done. The results obtained are estimated
by mean square error (MSE).GA with NC showed better result than GA with MI.
Fig 2(a) Fig 2(b) Fig 2(c)
Fig 2(a) represents reference and scene image rotated to 180 degrees. Fig 2(b) represents
registered image using NC. Fig 2(c) represents registered image using MI. Fig 2(d) shows the
registered image using MSE. The output images shown are obtained in 300th
generation
Fig 3(a) Fig 3(b) Fig 3(c)
Fig 3(a) represents reference image with no noise and scene image with lesion 1% noise. Fig 3(b)
represents the registration image obtained using NC. Fig 3(c) shows the registration image using
MI. The results obtained were taken at 300th generation

57
Table 2: Optimisation result using GA based on NC
Cases NC FV Rotation Translation Scaling
I 1.906 3.14 [177.89 153.00] [1 1]
II 1.963 6.28 [178 153.72] [1 1]
III 1.954 6.28 [177.87 154.26] [1 1]
IV 1.894 6.27 [177.45 153.00] [1 1]
Table 3: Optimisation result using GA based on MI
Cases MI FV Rotation Translation Scaling
I 1.456 3.13 [175.00 152.21] [1 1 ]
II 1.729 6.27 [178.00 153.72] [1 1 ]
III 1.698 6.27 [177.31 153.89] [1 1 ]
IV 1.685 6.27 [175.13 154] [0.99 0.99]
Table 4 : Comparison of NC and MI using MSE
Cases NC-MSE MI-MSE
I 8.193 10.988
II 8.274 11.78
III 8.110 9.634
IV 8.540 10.324
4. CONCLUSION
It can be seen that in both the cases, normalised correlation showed better convergence and also
the MSE value to be less when compared with MI. It is observed that as generations’ increases,
the fitness value increases. Higher the fitness value and more constant it is, better the
convergence.
The objective functions were evaluated for GA using normalised correlation, mutual information
and mean square error method. Normalised correlation method shows better performance than the
other. Registration is nothing but the calculation of an image transformation T, achieved by
optimisation of some similarity measure computed directly from intensity values of the images.
In future work, performance of intensity-based and feature- based methods may be analysed.
REFERENCES
[1] Andrea Valsecchi and Sergio Damas,(2012) “ An Image Registration Approach using Genetic
Algorithms”, IEEE World Congress on Computational Intelligence.
[2] Andrea Valsecchi and Sergio Damas, Linda Marrakchi-Kacem,(2013) “ Genetic Algorithms for
Voxel-based Medical Image Registration” ,IEEE, CIMI,pp. 4673-5919
[3] Barbara Zitova,Jan Flusser,(2003) “Image registration methods: a survey”, Image and Vision
Computing 21 ,pp. 977-1000.
[4] Derek L G Hill, Philipp G Batchelor,(2001) “Medical image registration”, Phys. Med. Biol.46,pp.1–
45

58
[5] Mahua Bhattacharya And Arpita Das,(2007) “Multi Resolution Medical Image Registration Using
Maximization Of Mutual Information And Optimization By Genetic Algorithm”,IEEE Nuclear
science symposium, conference record, pp. 2961-2964
[6] Michel A.Audette, Frank P. Ferrie,(2000) “An Algorithmic Overview Of Surface Registration
Techniques For Medical Imaging” , Med Image Anal,vol 4,pp. 201-217
[7] Myung-Eun Lee, Soo-Hyung Kim,(2009) “Intensity-Based Registration Of Medical Images”,
International conference on test and measurement,pp.239-242
[8] C.Sasi varnan, A.Jagan, Jaspreet Kaur, Divya Jyoti, Dr.D.S.Rao,(2011) “Image Quality Assessment
Techniques on Spatial Domain” ,IJCST, Vol. 2, Issue 3,pp. 177-184.
[9] J. N. Ulysses, A. Conci,(2010) “Measuring Similarity in Medical Registration”, IWSSIP 17th
International Conference on Systems, Signals and Image Processing.
[10] Xiaoguang Pan, Kai Zhao, Jiren Liu, Yan Kang,(2010) “A Hybrid Optimization Method For
Maximum Mutual Information Registration”,International conference on biomedical engineering and
informatics,pp.18-22
[11] Yalcinoz T, Altun H and Uzam M,(2001) “Economic dispatch solution using a genetic algorithm
based on arithmetic crossover.” IEEE Porto PowerTech',Porto, Portugal, 10-13 September 2001.Paper
No. AIT1-166.
AUTHORS
Shanmugapriya.S completed B.E- Biomedical engineering in Rajiv Gandhi College of Engineering.
Chennai. Currently doing M.E – Medical Electronics in College of Engineering, Guindy, Anna University,
Chennai. Field of interest are image processing, tissue engineering and bio instrumentation.
Dr.S.Poonguzhali received her PhD degree (Medical Image Processing) from Anna University, Chennai in
2009. She has 17 years of experience in teaching and also has published more than 70 papers in National
and International Journals and conference. Her research interest is Biomedical Instrumentation, Medical
Image Processing. Currently, she is working as Assistant Professor in Department of ECE, College of
Engineering, Guindy, Anna University, Chennai.
Ms. Umamaheshwari, Research Scholar in Anna University, has published many papers in National
conference. Her research interest is Medical Image Processing

An intensity based medical image registration using genetic algorithm

More Related Content

What's hot (19)

Viewers also liked (19)

Similar to An intensity based medical image registration using genetic algorithm (20)

Recently uploaded (20)

An intensity based medical image registration using genetic algorithm