You are on page 1of 85

 

   
 
      
 
 


 


 
 
   

     
   
   
  
   
 

  

 

   
        



        
       
      
  
    
   
        
 
 
   




    
  

        
  
    
     

  
  



 

  

  
 
 
  !  
 "

#

  "#! "
 #

 $

% &
 "#$%&!'   

'(!

 
  
  


)*" (


+
(# &#
,&
$ $**&-&
")
 
   

(# && +.'!/
&
.
 
 -

"  
$$$'

 0&
.

 



 







 
  
  




  


 
  
  




 
 



      


  

 
  
 


   
  
   
 
   
 



   
 


  
  
 
   
  

 


  
   


 
       
  
 

 
  
 

   

  

  
 
  
 
    

 


 
 
 


 

 !"#$
 
 
 
%
 " !
&

 

 ' &   
( 
)*!
 &

"
" *)+,-!




  


 

 ./
 &$

& 
  . 0,)* %
  "  !
 &

 

 
' &   ( 
 


"
" 0,)*
Intensity Based Medical Image Registration

Nagendra Pratap Singh∗, Rajeev Srivastava


Assistant Professor, Department of Computer Science and Engineering, M.M.M. University
of Technology, Gorakhpur-273010, U.P., India
Head, Department of Computer Science and Engineering, IIT (BHU), Varanasi-221005,
U.P., India

Abstract

The registration is the process to determine a geometrical transformation


that aligns points in one view of an image with corresponding points in an-
other view of that image or other image. Image fusion, matching or wrapping
can also be defined as a process of aligning two or more than two images.
The various design steps of a registration process include geometric transforma-
tion, similarity measure, and optimization methodology. Some review articles
available in literature mainly focused on either geometric transformation or op-
timization techniques briefly and some of them focused on all methodologies
but they have a lack of comprehensive presentation of review of each design
step. The objective of this book is to present a comprehensive review which
include a definitive documentation on mathematical as well as theoretical as-
pects of different methodologies used in each design step of intensity based
medical image registration system. This book also focuses on the limitations
of each methodology and discussed about recently proposed approaches of the
respective methodologies. Further, this book presents the summary and out-
comes of various intensity based medical image registration systems applied on
medical images of human body parts such as brain, heart, abdomen, retina,
breast etc. arising from various medical imaging modalities which include

∗ Corresponding author
Email addresses: npscs@mmmut.ac.in (Nagendra Pratap Singh ),
rajeev.cse@iitbhu.ac.in (Rajeev Srivastava )
Computerized Tomography-Computerized Tomography (CT-CT), Computer-
ized Tomography-Positron Emission Tomography (CT-PET), Positron Emis-
sion Tomography-Magnetic Resonance Imaging (PET- MRI), and Ultrasound-
Computerized Tomography (US-CT).
Keywords: Medical image registration, transformation, similarity measure,
optimization

1. Introduction

Digital image processing is applied on medical images and these images are
used for diagnosis, disease monitoring, treatment planning and guidance for
surgery. A large variety of medical imaging modalities exists that have been
used as primary inputs for medical image registration studies. The selection of
the imaging modality for a clinical study requires medical insights specific to
organs considered. It is impossible to capture all the details from one imag-
ing modality that would ensure clinical accuracy and robustness of the analy-
sis and resulting diagnosis. Some of the major modalities in clinical practice
include computerized tomography (CT), computerized tomography used for
angiography (CTA), quantitative computed tomography (QCT), quantitative
coronary angiography (QCA), quantitative vascular angiography (QVA), hip
structural analysis (HSA), magnetic resonance imaging (MRI), magnetic reso-
nance imaging used for angiography (MRA), dynamic contrast-enhanced mag-
netic resonance imaging (DCE-MRI), nuclear medicine using multi-gated acqui-
sition scan (MUGA), single photon emission computed tomography (SPECT),
positron emission tomography (PET), ultrasound for Abdominal/small parts ul-
trasound, echo-cardiography, intima-media thickness (IMT), contrast enhanced
ultrasound (CEUS), X-ray imaging for mammography. These imaging modal-
ities find a large range of application in diagnosis and assessment of medical
conditions affecting brain, breast, bone marrow, abdomen (liver, kidney and
spleen), cervical, chest, lung, entire body, pelvis, prostate, whole thorax , mouth,
teeth, intestines and soft tissues.

2
The images obtained by using different imaging modalities needs to be com-
pared with the one another and combined for analysis and decision making.
For disease progress monitoring and growth of abnormal structures, images are
acquired at different times and stages, different depths or with different modal-
ities. The misalignment between images is inevitable and reduces the accuracy
of further analysis. To deal with these issues image registration process plays
an important role in medical image analysis and provides a platform for group
analysis and statistical parametric mapping. The registration is the process
to determine a geometrical transformation that aligns points in one view of
an image with corresponding points in another view of that image or other
image. The other medical applications of medical image registration systems
include (i) monitoring of healing therapy and tumor evolution, (ii) combination
of sensors recording the anatomical body structure like MRI, CT with sensors
monitoring functional and metabolic activities like PET, SPECT, (iii) compar-
ison of patient’s image with digital anatomical atlases, specimen classification
(iv) Characterizing normal versus abnormal anatomical shape variations etc. In
addition to medical application of a registration system, the other application
areas where it is used are astrophysics (alignment of images from different fre-
quencies), military applications (target recognition), remote sensing, and many
others. A good overview of these applications of image registration is presented
in [1]. Any registration technique can be described with the help of four com-
ponents that are the feature space, the search space, the search strategy and
the similarity metric [2]. The feature space, where features to be matched are
selected in the image pair and the range of transformation or the search space re-
lates the source and target images. To find the optimum transformation within
the search space using an optimization algorithm, the similarity metric is used to
measure the similarity between source and target image which determines the
optimal transformation parameters to be used as a function of the similarity
measure.
Because of its importance the topic of medical image registration is useful
for both researchers and medical practitioners. Medical image registration has

3
been intensively investigated from previous two decades and numerous algo-
rithms have been proposed but only few short reviews are available on intensity
based medical image registration system which comments on the appropriate-
ness and usefulness of an available method. In this book, our concentration is
to present a literature review of previous two decades of intensity based med-
ical image registration systems that are used for registering brain images of
the same subject acquired with different modalities (e.g. MRI , CT and PET)
[3]-[4], retina [5], heart [6], [7], breast [8] and some others like abdomen, cer-
vical, chest, lung, pelvis, prostate, whole Thorax etc. Here we also describes
the advantages and limitations of the existing methods and comments on the
appropriateness and usefulness of an existing method which may be used in real
time applications.
Maintz and Vierger [1] had suggested the nine fundamental criterion to clas-
sify the various methodologies used for medical image registration that is dimen-
sionality, nature of registration basis, nature of transformation, domain of trans-
formation, degree of interaction, optimization procedure, modalities involved,
subject and object. In papers [9], [10], authors suggested the two classifica-
tion criteria of the image registration techniques a multi-modal registration and
temporal registration, where the multi-modal registration handle the registra-
tion of images of the same scene acquired from different sensors e.g. integrate
structural information from CT or MRI with functional information from ra-
dionucleic scanners such as PET or SPECT for anatomically locating metabolic
function. Where as the temporal registration can handles the registration of im-
ages of the same scene taken at different times or under different conditions e.g.
digital subtraction angiography registration of images before and after radio iso-
tope injections to characterize functionality, digital subtraction mammography
to detect tumors etc. In paper [11], authors presented a review of deformable
medical image registration system and discussed about deformable registration
methods with emphasis on the recent advances in the domain of multi-modality
fusion, longitudinal studies, population modeling, and statistical atlases. Fran-
cisco et al. [12] presented a review of medical image registration system and

4
classify the medical image registration system into intensity based and feature
based, where the author explored different methodologies used for the design of
the intensity and feature based medical image registration system. In literature
many authors presented the methodologies for the design and development of
medical image registration system for specific imaging modalities. M. Essadiki
presented a technique for combining panchromatic and multi-spectral spot im-
ages [13]. Flusser used moment based approach to correct affine distortion and
they had done medical image analysis for degraded images to locate invariants
[14]. J.P.W. Plum presented multi-modal image registration using generalized
survival exponential entropy medical image computing [15]. E.A. Sascha et al.
[16] proposed a DIRBoost algorithm which is inspired by the theory on hypoth-
esis boosting which is the well known field of machine learning.
Many authors presented their work on medical image registration accord-
ing to various applications in hand such as cardiac applications [17], nuclear
medicine [18], radiation therapy [19], digital subtraction angiography [20], and
brain warping applications [21]. In papers [22], [23] author focused on modeling
of soft-tissue deformation during imaging or surgery and papers [24]-[25] present
the model changes in anatomy of the object of interest.
There are some other image registration methods available in literature
which have been presented for non-medical applications but the methodolo-
gies presented in these papers may also be useful for medical image registra-
tion. For examples, Sabuneu et al. [26] presented various entropy based im-
age registration techniques, H.Manjunath et al. [27] explained various contour
based approaches for multi-spectral image registration, A. Quddus et al. [28]
proposed a novel and automatic multistage wavelet-based image registration
approach for image retrieval applications and used multi-scale wavelet represen-
tation with mutual information to facilitate matching of important anatomical
structures at multiple resolutions, F. Maes et al. [3] applied mutual information
to measure the information redundancy between the image intensities of corre-
sponding voxels in both source and target images and present novel histogram
based method for estimating and maximizing mutual information between two

5
multi-modal and possibly multi-band signals, Stone et al.[29] proposed fast im-
age registration by progressively registering wavelet representations at different
resolutions. Ghazaw [30] presented wavelet based image registration on parallel
computing and explained wavelet based image registration, W. Cao et al. [31]
proposed PCAT (principal component analysis transform) and WPT (wavelet
packet transform ) for remotely sensed image fusion.
From above presented information, it may be observed that there are various
methodologies available in literature which may be used for the design and
development of a medical image registration and each of them have their own
advantages and disadvantages. The registration methodologies may be broadly
classified in to two groups namely intensity based and feature based.
Here we emphasis to present a comprehensive literature review of intensity
based medical image registration system for various imaging modalities. The
literature review is categorized according to the design steps or methodologies
used for the design and development of an intensity based medical image regis-
tration system. The various methodologies used for the design of the intensity
based medical image registration system include the various types of geometric
transformations, similarity measures, and optimization techniques used. This
book provide a detail review of each of these three groups of methodologies for
various imaging modalities. The block diagram showing the various components
of an intensity based medical image registration system is shown in Figure 1.
The complete book is organized as follows: Section I presents the introduc-
tion; Sections II-IV presents the review on methodologies used for registration
of medical images wherein Section II presents the review on various geometric
transformations , Section III presents the review on various similarity measures,
and Section IV presents the review on optimization methods; Section V presents
the discussion of the review; and finally Section VI presents the conclusion.

6
Figure 1: Typical Image Registration Model

2. Geometric Transformation

Medical image registration approach highly depends on the geometric trans-


formation used. The transformation model is based on the nature of image
data to be registered. The geometric transformations used for registration can
be broadly classified into two group namely rigid and non-rigid geometric trans-
formations. The rigid geometric transformations classified into two groups in-
clude translation, and rotation. The non-rigid geometric transformation model
classified into five groups namely affine, projective, physical modal based, basis
function enpension based, and constraints on the transformation based. The
physical modal based geometric transformation model may include elastic, fluid
flow, and optical flow and elastic body model further classified in two groups
namely linear elastic body and non-linear elastic body model. The basis function
enpension based geometric transformation model divided in four groups include
radial basis functions (RBF), elastic body splines (EBS), free-form deformation
(FFD), piecewise affine. The constraints on the transformation based geomet-
ric transformation model divided in three groups include inverse consistency,
topology preservation, and diffeomorphic transformation. The classification of
geometric transformation available in literature is shown in Figure-2.

7
2.1. Rigid Transformation

Rigid transformation is a transformation of an image that preserves size and


shape by preserving the length of each segment and angle. The rigid transfor-
mation is equivalent to change the coordinate system of moving image from one
coordinate system to other by translation and rotation. These transformations
have been applied in the registration of image having rigid structure in nature
like bones or brain, when neither skull nor dura has been opened [32, 33, 34].
Some of the review papers available in literature such as [35],[1] mainly focusses
on the review of rigid transformations.

2.2. Non-rigid Transformation

Today the development of non-rigid registration techniques involves for the


applications of various modalities and soft tissue deformations of anatomical
structures. The non-rigid medical image registration is a challenging research
work due to its smoothness requirement and high degree of freedoms in the
deformation process of medical imaging and the computational time is also a
challenging task for many clinical applications [24].

2.2.1. Affine and Projective Transformation


Generally, rigid transformations combine only translations and rotations.
But in the literature, rigid transformations sometimes include scaling then the
corresponding transformation is known as affine transformation. Furthermore
an affine transformation can include shearing to maps straight lines to straight
lines for preserving the parallel lines [12],[15]. The affine transformation have
been used for the registration of ultrasound images [6],[7], [36], [37]. It is able to
extend the transformation parameter with a scaling factor [38],[39] and shearing
[40],[41],[42] for the image dimension. There are nine to twelve different param-
eters used in affine transformation that are able to correct the error in set of
pixels dimension [43],[44].
The projective transform can handle changes after a tilt of the image plane
relative to the object plane. It is usually applied in 2-D/3-D image registration.

8
Figure 2: Type of Geometric Transformation

It is differs from the affine transformation in the sense that it need not pre-
served the parallel lines [12]. In case of projective transformation the projective
parameters are fixed where as in other transformation (the rigid or affine) the
parameters are optimized. Images obtained from x-ray projection, endoscopy,
laparoscopy and microscopy are all two-dimensional views of three-dimensional
objects rendered by means of projecting light rays from a three-dimensional
scene onto a two-dimensional plane. This type of geometrical transformation is
known as perspective projective transformation. It is subset of the projective
transformations [45].

2.2.2. Physical Model


The non-rigid geometric transformations derived from physical model is fur-
ther classified in to three groups [46] namely elastic body model, viscous fluid
flow model, and optical flow. The elastic body model can further be classified in

9
two groups namely linear elastic body model and non-linear elastic body model.
Brief descriptions of these models are given in following subsections.

Elastic Models :-. In linear elastic body model the transformation is described
by Navier-Cauchy partial differential equation (Navier-Cauchy PDE), which is
shown in eq n -1,

μ∇2 d(x) + (μ + ψ)∇(∇.d(x)) + f (x) = 0 (1)

where d(x) is displacement vector at particular position x. μ and ψ are the


lame constants describe the material properties and f (x) represents the body
force per unit volume, which drives the registration. The Navier-Cauchy PDE
is used to optimize the balancing problem of an external forces with the internal
stresses, which can be solved by using finite difference [47], [48], finite element
method [49], Fourier transform method [50], basis function expansion [51]. This
model was initially proposed by Broit [47] for elastic membrane image that is
transformed under the influence of two forces and complete when equilibrium is
reached. The Bajcsy et al. [48] proposed an enhanced version of linear elastic
body model for low contrast anatomical brain structures in Positron Emission
Tomography (PET) and CT images, where before elastic registration author
corrected the global differences using a translation, rotation and scaling which
was determined by aligning the center of mass. Davatzikos et al. [52] proposed
an elastic algorithm for inter-subject registration of cortical grey matter and
modeled the brain cortex as a thin spherical shell of constant thickness. Da-
vatzikos in [53] further proposed a homothetic mapping approach to match both
the cortical and ventricular surfaces of segmented image of brain. These mod-
els are nowadays used by various authors [54, 55, 56, 57, 58].The linear elastic
registration algorithms generally use parametric models to represent the defor-
mation by a modifying number of parameters, in the multi scale setting. For
example a hierarchical basis functions by Moulin et al. [59], quad-tree-splines
[60], multi resolution subspaces [61] and wavelets [62].
Linear elastic body model is not able to handle large deformations, because
the Navier-Cauchy equation is valid only for small displacements. To solve this

10
problem a nonlinear elastic body model that preserves the topology of present
structure have been proposed by Bajcsy et al. [54] and Davatzikos [58], where
the author Bajcsy et al. [54] uses principal inertia axes and Davatzikos [58] uses
the stereotaxic space to provide a good initialization. Iteratively using a multi
resolution approach is another way to solve this problem [54]. Author Rab-
bitt et al. [63] proposed a deformable model based on hyper-elastic material
properties and the nonlinear equation was solved by using local linearization
and finite element method. The author Pennec et al. [64] cannot use local lin-
earity assumption and extend linear elastic model to the nonlinear by using St
Venant-Kirchoff elasticity energy and Yanovsky et al. [65] proposed a symmetric
registration model based on the St Venant-Kirchoff elasticity. Le Guyader et al.
[66] proposed an approach to combine the segmentation and registration based
on nonlinear elasticity. Burger et al. [67] focused on the numerical implemen-
tation of registration system by using poly-convex regularization term. Droske
et al. [68] used an hyper-elastic, poly-convex regularization term to handle the
length, area and volume deformations.

Fluid Flow Model :-. In general the elastic transformations are not feasible to
handle the highly localized deformations because the elastic model based on
stress and stress is directly proportional to the strength of deformation. To
overcome this problem by using a viscous fluid that allows the restoring forces
to relax over time, which enables the modeling of highly localized deformations
including corners. The fluid flow transformation model is described by Navier-
Stokes partial differential equation (Navier-Stokes PDE) ,as shown in eq n -2.

μf ∇2 v(x, t) + (μf + ψf )∇(∇.v(x, t)) + f (x) = 0 (2)

where
∂d
v(x, t) = + v.∇d (3)
∂t
μf and ψf are the viscosity coefficients of the fluid and f (x) is body force per
unit volume. The Navier-Stokes partial differential equation is simplified by
considering a very low Reynold’s number flow. The Navier- Cauchy PDE of

11
linear elasticity (eq n -1) is identical to the Navier- Stokes PDE (eq n -2) except
that the PDE operates on displacement d rather than velocity v. Christensen
et al. in [51], [69], [70] proposed a well known fluid flow algorithm, where in
[51] the registration process start with affine then linear elasticity and finally
apply a fluid flow algorithm, in [70] the registration process is based on a trans-
formation hierarchy of successively increasing numbers of degrees of freedom.
To solving the Navier-Stokes PDE is computationally expensive and inefficient,
that is the major drawback of this model. So Christensen et al. in [69] proposed
an algorithm that was implemented on parallel hardware that could improve the
computational efficiency. There are some other author proposed an algorithm
for faster solutions such as Bro-Nielsen et al. [71] proposed a technique based
on a convolution filter with Green’s function in scale-space, Cahill et al. [72]
use Fourier methods to solve the linear PDE system efficiently for any boundary
condition and proposed an extended version that can be applied in the case of
diffusion, curvature and elastic regularization under periodic boundary condi-
tions, Crun et al. [73] proposed a multi-grid approach with a multi-resolution
scheme for handling anisotropic data, Wang et al. [74] used fluid deformation
models in an atlas-enhanced registration setting, and Chiang et al. [75] proposed
an inverse consistent variant of fluid registration for the registering a diffusion
tensor images.

Optical flow :-. Optical flow is a type of non-rigid transformation method used
for small scale deformations on temporal sequences of image frames [76, 77, 78].
It is based on the principle of intensity conservation between image frames. The
hypothesis of optical flow is to assume that the intensity of a moving object is
constant with time. The optical flow have been derived by using the first order
terms of the Taylor expansion of the intensity function in the target frame,
which is given in eq n -4.
d.∇g(x) = h(x) − g(x) (4)

where h(x) − g(x) is the change in intensity between frames and ∇g(x) is spatial
derivative of intensity in the target frame. Author Thirion [79] first proposed

12
an optical flow model for non-rigid transformation by using Maxwell’s demons
algorithms. The Maxwell’s demons optical flow, based non rigid registration
algorithm has been used to identify the changes in displacement field of image
frames and determined that changes and then re-samples the moving image for
successive iteration. The optical flow model gives a numerically stable expression
for d after approximation as shown in eq n -5.

[h(x) − g(x)]∇g
d(x) = (5)
(∇g)2 + [h(x) − g(x)]2

where ∇g represent the direction of displacement on the fixed image and its
orientation is positive if h(x) − g(x) > 0 otherwise negative. A major drawback
of this model is that there are no constraints of the displacement, it does not
necessarily preserve the topology and it requires a much computation time. Due
to these deficiencies this model is no much feasible for many clinical applications
[46].

2.2.3. Basis Function Expension Based


In general these types of geometric transformation are not derived from ei-
ther interpolation theory [80] or approximation theory [81]. In interpolation
theory author consider the known displacements in restricted set of location of
the sample and interpolated them in the rest of the image domain. Whereas
in approximation theory consider that there is an error to estimation of dis-
placements in the samples. Thus the transformation smoothly approximates
the known displacements rather than taking the exact values. [11]. Geometric
Transformations derived from interpolation or approximation theory are mainly
classified as radial basis functions, free-form deformation, elastic body splines,
basis function from signal processing and piecewise affine models. Out of these
a radial basis function and piecewise affine models are more stable and widely
used[46].

Radial basis function :-. According to M.J.D. Powell [82] and M.D. Buhmann
[83] the RBFs are the function of distance between the interpolation point x

13
and landmark position xi , defined as

N
f (x) = wi R(x − xi ) (6)
i=1

where i is a landmark pairs and N is total no of landmarks. wi are the weight of


ith landmark, which is determined by solving a set of linear equations such as
multiquadrics (MQ) [83], inverse multiquadrics (IMQ) [84] and Gaussians [84] ,
[85]. The MQ, IMQ and Gaussians is a type of radial basis function and defined
by following equations.
For MQ

R(x − xi ) = ri2 + s2 (7)

For IMQ
1
R(x − xi ) = [ri2 + s2 ]− 2 (8)

For Gaussians  
ri2
R(x − xi ) = exp − (9)
2σi2
where ri is an euclidean distance x − xi  and s is a parameter to control the
amount of smoothing, which is directly proportional to parameter s. According
to MQ equation the control parameter s of smoothing amount is s2 so the result
is more smoothing. When basis functions are logarithmic, then the obtained
transformation is similar to a thin-plate spline except for a linear term. After
comparative study Franke [86] found that multiquadrics followed by thin-plate
splines produced the best accuracy in the interpolation of randomly spaced data.
The basis functions used in radial basis functions are either monotonically
increasing or monotonically decreasing. When monotonically increasing basis
functions are used in image registration then farther away control points will
have a much influence on the transformation at a particular point with respect
to nearest control points. When monotonically decreasing basis functions are
used in image registration then a surface point will depend more on nearest
control points than on farther away so monotonically decreasing basis functions
used in radial basis functions have been more appropriate in image registra-
tion. The radial basis functions provides a global support and monotonically

14
decrease. On the basis of an extensive experiment author Franke [86] show
that the monotonically decreasing radial basis functions do not perform well in
comparison to monotonically increasing radial basis functions. These proper-
ties are useful for landmark based registration. Rohr et al., [87], [88], [55] and
Fornefett et al. [89] worked on landmark based medical image registration by
using RBF. Goshtasby [90] applied the TPS first time to the registration of re-
mote sensing images and Bookstein [91],[92] first time introduced a Thin-plate
splines (TPS) for modeling shape deformation in medical image registration.
Grimson [93] and Terzopoulos [94] described the TPS function as a variational
Euler-Lagrange equation which minimizes the bending energy. TPS generally
performs better in comparison to MQ, when the transformation functions to
be estimated linearly. However, when the transformation function is nonlinear,
MQ is slightly performs better. In literature, there are various author proposed
their worked on that such as Zagorchev et al. [95] proposed a comparative
study of RBFs used as transformation functions for nonrigid image registration,
Marsland and Twining [96] proposed a clamped-Plate Splines for group-wise
registration and analysis of deformable registrations, Camion et al. [97] intro-
duced Geodesic Interpolating Splines (GIS) , Younes [98] proposed an extended
method by combining combine GIS with affine transformations, Arad et al. [99]
suggested the use of Gaussian functions to parametrize the deformation, Yang
et al. [100] proposed compactly-supported RBFs for image registration to pre-
serve topology and recently Rohr et al. [101] proposed an extended approach
by introducing a variant of thin-plate splines for image registration with radial
basis functions.

Elastic Body Splines :-. Elastic body splines (EBS) for a class of 3D splines,
first proposed by Davis et al. [102] to landmark based registration and it is
also used in intensity based medical image registration. The EBS was inspired
by an interpolation scheme and may also be inspired by physical models. The
EBS is based on the theory of linear elasticity. In general EBS assume a poly-
nomial radially symmetric force f (x) = cr2k+1 . Whereas the author Davis et

15
al. [102] assumes polynomial forces f (x) = cr, r = ||x|| and f (x) = c/r where

r = x2 + y 2 + z 2 and c is constant vector. Author used these assumption
for deriving analytical solution of the Navier-Cauchy PDE to a homogeneous
isotropic elastic body with the help of Galerkin vector method, describe in [103].
Then these functions are used as a basis functions for interpolation based reg-
istration scheme. There are several drawbacks of the forces f (x) = cr and
f (x) = c/r used by Davis et al. [102]. Either they increase with increasing
distance from the origin, which is physically not possible or having a singularity
at the origin.
Kohlrausch et al. [104] proposed an extended approach known as Gaussian
elastic body splines (GEBS). The GEBS considering Gaussian forces that means
the forces are given by Gaussian function of the distance from the landmark.
Here the author used a Gaussian force, which is defined by eq n -10.

1 r2
f (x) = ci  e− 2σ2 (10)
( 2Πσ)3

In GEBS the size of Gaussian kernel can be used to parameterize the compact-
ness of the model’s support. The GEBS have big advantage in comparison to
EBS that used the more realistic forces (Gaussian forces) instead of polynomial
or rational forces.
Stefan Worz et al. [105] proposed a new approximation scheme for elas-
tic image registration using Gaussian elastic body splines. The approximation
scheme based on an extended energy function related to Navier-Cauchy PDE
under Gaussian forces and allows individual weight to the landmarks according
to their localization uncertainties. The uncertainties are depending on their
isotropic or anisotropic nature, which is represented as either scalar weights or
matrices. Author experimentally show that the approximating GEBS approach
gives better registration results with respect to previously proposed interpolat-
ing GEBS and TPS.

Free Form Deformations :-. Free-form deformation (FFD) is an important tool


for computer-assisted geometric design. It was commonly used as a transforma-

16
tion model in medical image registration. Alan Barr [106] proposed a method
to alter the transformation when it was being applied to the image. Author
uses a set of hierarchical transformations for deforming an image by including
stretching, taper, twisting, and bending operators. This technique uses the sur-
face normal vector of the undeformed surface and a transformation matrix to
calculate the normal vector of an arbitrarily deformed smooth surface. Seder-
berg et al. [107] proposed a more general approach which deformed images by
deforming the space where the object was embedded. This technique is based
on globally supported trivariate (means number of control points in three di-
rections) Bernstein polynomial basis functions βs,t , which provide the result of
the FFD shown in eq n -12 and the function βs,t is defined by eq n -11.

 
t s
βs,t (x) = x (1 − x)t−s (11)
s


cx ,cy ,cz
ϕ(x, y, z) = βi,cx (x)βj,cy (x)βk,cz (x)Pi,j,k (12)
i=0,j=0,k=0

where cx , cy , cz represents the number of control points in three directions x,y,


and z respectively. FFDs have a big advantages with respect to previous surface
based methods. Reason behind that it is independent of the objects surface and
the deformation model is object-independent in the sense that it applies to 3-D
space.
Coquillart et al. [108] proposed an extended free-form deformation (EFFD)
approach by using a lattice and B-spline control points to approximate the shape
of the intended deformation. The main objective of this approach is to change
the shape of existing image surface either by adding randomly shaped bumps
or by bending it along an arbitrarily shaped curve. Chang et al. [109] proposed
an approach to deform an image by repeatedly applying affine transformations
in space. It uses a single Bezier curve and a few affine maps and generalizes the
De Casteljau algorithm for curve evaluation. There are many other extensions
of FFDs exists in literature. Schnabel et al. [110] proposed an FFD approach
that use multi-level B-splines, where the overall transformation was given as

17
a summation of transformations of each level. The authors assign the status
of each control point, which is either active or passive in status. The active
control points are allowed to move, whereas the passive control points remained
fixed. Author Shi et al. [111] used the multi-level B-splines model of[110]
while imposing that only a sparse subset of the control points is active. Wang
et al. [112] proposed an approach based on non-uniform rational B-splines
(NURBS), which is used to perform medical image registration in an adaptive
focus manner. Noblet et al. [113] proposed a new symmetric extension of
FFDs by assuming the both images deform toward a common domain under
the influence of two isomorphic grids. The common domain was identified by
using an equal distance from the fixed and moving image. Feng et al. [114]
proposed an inverse consistent method based on FFDs. This method is used to
examine how the composition of the two transformations mapped back to the
image domain. Furthor more the FFDs have been extended to tackle multiple-
image registration [115],[116], where hard constraints are employed to define
a reference domain and the transformation model has been extended to the
spatio-temporal domain [117], [118], where B-splines are also used for temporal
axis due to their transparency, computational efficiency and applicability.

Affine based Models :-. This model is also known as locally affine model and pa-
rameterize the transformation by locally linear deformations. This model have
been classified in two groups, a piecewise affine models and poly-affine model. In
a piecewise affine models, the image is patched by a set of triangles or tetrahedra
and their nodes are responsible to parameterized the deformation by applying
affine interpolation inside each region. The main strengths of this method are
their efficiency and invertibility and the main limitation is a lack of smoothness
in the region boundaries [11]. The first method is usually credited to author
Collins et al. [119]. They proposed an algorithm named ANIMAL and used
for inter-subject brain registration, to determine the geometrical variability of
brain structures. There are various recent approaches exist in literature based
on a piecewise affine model. Hellier et al. [120] proposed a multi-resolution and

18
multi-grid approach, where the image was partitioned adaptively into cubes and
the affine transformation was inferred for each one. Zhang et al. [121] proposed
another similar approach that tackled diffusion tensor registration by consider-
ing tensor reorientation. The images were separated into contiguous blocks and
an affine transformation was recovered for each of them. Buerger et al. [122]
proposed a new hierarchical framework to adaptively separate the images into
regions. Splitting was formulated as an energy minimization problem and used
three criteria to group the regions. First criterion to group the regions with rich
structural information, second criterion to group the regions with significant
residual error in blocks, and the last criterion to grouped the regions with simi-
lar motion. Out of these three criterions the performance of second criterion was
found best [122]. The poly-affine model are used fuzzy regions that overcome
the limitation of piecewise affine model. Arsigny et al. [123] proposed a poly-
affine transformation model, where the transformation is parameterized by a set
of anchor points and the parameters are defines the importance of every point.
The fuzzy regions are defined by calculating the influence of an anchor point at
each position of the image. This approach is computationally expensive so au-
thor Arsigny et al. [124] again proposed an extended poly-affine transformation
so that it’s inverse is also poly-affine and the fusion of affine transformations
was rendered invariant to affine changes of the coordinate system.

2.2.4. Constraints on Transformation


The constraints play an important role for biomedical applications and that
may be applied to the transformation such that it exhibits special properties.
These properties include inverse consistency, topology preservation, and dif-
feomorphism. Mostly the existing registration algorithms are asymmetric that
means, when swapping the order of input images, then the asymmetric regis-
tration algorithm does not able to estimate the inverse transformation.

Inverse consistency :-. The main objective of inverse consistency methodology


is to overcome the problem of asymmetric registration algorithm by estimating
both the forward and the backward transformation simultaneously. Christensen

19
et al. [125] proposed a consistent image registration approach and Johnson et
al. [126] proposed a consistent landmark and intensity based image registra-
tion approach. Both authors have been focused on inverse consistency as reg-
ularization constraint and focused how well the images are aligned when one
image is deformed by the forward transformation and the other image by the
backward transformation that means both the forward and backward transfor-
mations must be inverse mappings of each other. This symmetric behavior is
achieved by introducing regularization constraint that penalizes the inconsis-
tency between the forward and backward transformations from the respective
inverse mappings. The inverse consistent methodology preserves topology and
not purely symmetric but only asymptotically symmetric.

Topology preservation :-. The Topology preservation is one of the most impor-
tant properties of registration algorithm and it is equivalent to the invertibility
of the deformation field. According to Musse et al. [61] topology must be pre-
served when two conditions are fulfilled first the determinant of the Jacobian
of the transformation is always positive and second the transformation is bijec-
tive. The authors Noblet et al. [127] and Christensen et al. [69] focused to
track the values of the Jacobian, when its value not positive then, an intermedi-
ate deformed image was created and the registration process was re-initialized.
There was some other approaches exists in literature for preserves the topology
by including an objective function that acts upon the Jacobian. Christensen et
al. [125] included an objective function that penalizes small and large Jacobian
values for both the forward and inverse transformation. The Rueckert et al.
[128] included another term in the objective function that penalizes values of
the Jacobian determinant, which is close to zero i.e. positive. Haber et al. [129]
proposed an inequality constraints based approach, where author used a variant
of a log-barrier method to solve the optimization problem rather than solving
the initial constrained problem.

Diffeomorphic transformations :-. Diffeomorphic preserves the topology of the


structures represented in the images to be registered and prevent folding which

20
is often physically impossible. A transformation function is a diffeomorphism,
if it is invertible and both the function and its inverse are differentiable [11].
It provides a good framework, when additional information about the spatial
transformation is not available. The registration methodologies based on diffeo-
morphic transformations are known as diffeomorphic image registration. The
Christensen et al. [69] proposed a diffeomorphic registration approach based
on the ”viscous fluid” registration method. In this model author used a finite
difference method to solve the differential equations that model one image as it
”flows” to match the shape of the other. The viscous fluid methods are used
to solve the large sets of partial differential equations but computationally ex-
pensive because solving the equations used successive over-relaxation. These
can be avoided by regularizing the velocity field [130]. To achieve this a diffeo-
morphic space-time mapping ϕ(x, t) is required and mapping is related to the
displacement field by ϕ(x, t) = x + d(x) that satisfies the following equations.

∂ϕ(x, t)
= v(x, t) (13)
∂t
∂ϕ−1 (y, t)
= [∇ϕ(y, t)]−1 v(y, t) (14)
∂t
ϕ(x, 0) = ϕ−1 (x, 0) = I (15)

where ϕ(x, t) describes a diffeomorphic flow through space-time mapping. The


vector y is a location in the target image and ∇ϕ(y, t) is the Jacobian of trans-
formation. Some author have been advised a faster ways to solve the differential
equations such as Fourier transforms to convolve with the impulse response of
the linear regularization operator by Bro Nielsen et al. [71], convolving with a
separable approximation by Thirion [79] and large deformation diffeomorphic
metric mapping (LDDMM) algorithm by M. F. Beg et al. [131]. The LD-
DMM algorithm allows for the definition of a distance between images or sets of
points. The distance between these elements is defined as a geodesic distance.
The geodesic distance is the distance measured along the shortest route between
two points. The geodesic distance of the set of points connects them and have
been used for studies of anatomical variability [132], [11]. The LDDMM algo-

21
rithms provides diffeomorphic transformation but not symmetric. So to achieved
the symmetric diffeomorphic transformation there are various approaches have
been exists in literature. Younes et al. [133] focused the alignment process of
symmetric property and Beg et.al. [134] focused on providing symmetric data.
The Avants et al. [135] proposed a symmetric LDDMM approach for a registra-
tion process based on cross-correlation. The Ashburner et al. [136] proposed a
geodesic shooting approach based on Gauss-Newton implementation, to calcu-
late an intermediate and final deformations on the basis of initial velocity and
starting position. Hernandez et al. [137] proposed a new approach by using
stationary Ordinary Differential Equations (ODEs) in the LDDMM framework.
The LDDMM algorithm have been extended to solve a various types of prob-
lem such as volume registration for scalar [134],[136] vector [138] and hybrid
registration [139],[140]. The author Pei Zhang et al. [141] proposed diffeomor-
phic registration algorithm for Diffusion-weighted imaging (DWI) data under
large deformation. This method generates spatially normalized DWI data and
hence it is possible to fit various diffusion models after registration for com-
parison purposes. Here author used a geodesic shooting mechanism to avoid
the huge computational resources that are needed to register high-dimensional
vector-valued data.

3. Similarity Measure

Similarity measure is most widely used registration basis for many appli-
cations. The intensity is referring as the scalar values of image pixels. The
pixel value depends on the modality involve in the registration process. In case
of intensity-base registration the transformation process is based on the pixel
intensity and it is determined by iteratively optimizing the similarity measure
calculated from all pixel values. In case of poor pixel value the optimization is
not able to provide better result so refer the similarity measure as a voxel similar-
ity measure. But the running time complexity of registration algorithm is high
when uses all voxel in similarity measure so in practice, generally use a subset of

22
Figure 3: Type of Similarity Measures

voxel and for that a little bit preprocessing are required. The subset of voxel is
either selected randomly or by using regular grid. In some algorithms the vox-
els are selected from defined region of interest (ROI), so the pre-segmentation
of the images are required. This preprocessing is depends on the modality of
medical images and body parts which is being studied. Sometimes the intensity
based similarity measures are applied on the derived parameter such as image
gradients of the image rather than the voxel values[45].
There are various types of similarity measures [45] that are used in intensity-
based image registration methods as shown in Figure-3. Before describing the
similarity measures, let us consider the images to be registered are fixed image
If and moving image Im , the set of voxel of these images are If (i) and Im (i)
respectively. Image Im is iteratively transformed to Im (T) by successively apply

23
transformation T i.e. Im (T) =T(Im ).

3.1. Sum of squares differences

Sum of squares differences (SSD) is a similarity measure also known as image


subtraction and used to identify the alignment of images If and Im . When the
value of SSD is zero then both images are aligned correctly otherwise there are
some misalignment. The misalignment is directly proportional to the increasing
value of SSD. The SSD is the optimum measure when two images only differ
by Gaussian noise [40]. Normally SSD is work well, when all used images in
registration process are nearly identical except for small change. For example
this approach was used in serial registration of MR images and fMRI exper-
iments [142], [143]. SSD is used by Friston’s statistical parametric mapping
software (SPM), which is based on a linear approximation [144], [145]. The
SSD is sensitive when small set of voxels having very large intensity differences
between images If and Im . The effect of these outlier voxels cannot be reduced
by SSD, so the sum of absolute differences (SAD) can be used to reduce the
effect of outliers [146]. SSD and SAD are calculated by using Eq n -1 and Eq n -2
respectively as show in Table-1.

3.2. Correlation Coefficient

The correlation coefficient (CC) is used as similarity measure, where the


intensities of images If and Im are linearly related [40]. For example ultrasound
image features of the wrist are linearly related to finger positions [147] and
supine magnetic resonance imaging Cobb measurements or idiopathic scoliosis
are linearly related to measurement from standing plain radiographs [148]. The
main characteristic of correlation techniques is carried out the correlation in
either the spatial domain or the spatial frequency domain (k-space) [146]. K-
space is widely used in magnetic resonance imaging first introduced by Likes in
1979 and then by Ljunggren and Twieg in 1983. According to the MRI physics,
k-space is use in 2-D or 3-D fourier transform of the MR image and having
complex values that are sampled during an MR measurement, in a premeditated

24
scheme controlled by a pulse sequence, i.e. an accurately timed sequence of
radio frequency and gradient pulses [149], [150]. This type of approach has
been applied on medical images, but the applicability appears to be limited by
the implicit assumption that the objects of interest are in the fields of view of
both images being registered [151],[152]. The CC is calculate by using Eq n -3 as
show in Table-1.

3.3. Variance of Intensity Ratio

The Variance of Intensity ratio (VIR) is also known as Ratio-Image Unifor-


mity (RIU). It is widely used in intra-modality registration that means, it is
apply for same image modality. The VIR algorithm was initially proposed by
Woods for the registration of serial PET images in 1992 [153] and then used
for serial MR image registration in 1998 [154].This algorithm is working with
a derived image ratio which is calculated by dividing the each pixel of image
If by each pixel in transform image Im . To find the transformation iteratively
that maximized the uniformity of the image ratio, determined by calculating
the normalized standard deviation of image ratio. The uniformity of the image
ratio is inversely proportional to the normalized standard deviation of image
ratio [45],[146]. The VIR is calculate by using Eq n -4 as show in Table-1.

3.4. Partitioned Intensity Uniformity

The partitioned intensity uniformity (PIU) is widely used for inter-modality


registration that means apply for different modality of image. Wood proposed
the modified version of own VIR algorithm for MR-PET registration, which is
known as PIU [155]. The PIU algorithm use an idealized assumption, that is
”all pixels with a particular MR pixel value represent the same tissue type so
that values of corresponding PET pixels should also be similar to each other”
[45]. The PIU algorithm partition the PET image into 256 separate groups
or iso-intensity sets, which is based on the values of PET image voxels then
maximize the uniformity of the MR voxel values within each groups. After
that uniformity is maximized within each group by minimizing the normalized

25
standard deviation. The PIU have two different versions depending on whether
image MR or image PET is partitioned [146]. The PIR is calculate by using
Eq n -5 as show in Table-1.

3.5. Information Theory based Similarity Measures

There are various types of similarity measures techniques based on informa-


tion theory like joint histogram and joint probability distribution, joint entropy,
mutual information (MI). The Shannon-Wiener entropy measure (Se ) is com-
monly used technique for measure of information in signal and image processing
[156], [157]. The Shannon-Wiener entropy (Se ) is measure by using a formula
as show in Eq n -6, Table-1 and is derived from three conditions that are able
to measure the uncertainty in a communication channel. First the Se should
1
be continuous in pi , second if all pi equal to n then Se should be monotoni-
cally increasing in n, where n is number of symbols. For example if there are
two images used in image registration then there are two symbols at each voxel
location then value of n is equals to two and the last condition, if a choice is
broken into sequence of choices then the Se should be the weighted sum of the
constituent Se . Shannon proved that Se is unique formula which satisfies all
these three conditions. The entropy having minimum value of zero when the
probability of occurring (pi ) one symbol is one and other is zero and the entropy
will be maximum if all symbols having equal probability of occurrence [45].
A joint histogram is useful for visualizing the relationship between the inten-
sity of corresponding voxel in two or more image. Joint histograms also consider
many features such as edge, density, gradient magnitude and rank. It is gener-
ally used in multi spectral data and becomes n-dimensional when n number of
images used to generate it. For two images If and Im the joint histogram is two
dimensional and it is constructed by plotting the intensity If (i) of each voxel in
image If against the intensity Im (i) of each voxel in image Im . So the axes of
histogram are the intensities or intensity partitions in each image. The value at
each point in the histogram is the number of corresponding voxel pairs with a
particular combination of intensities in the different spectral components. [158].

26
If the joint histogram is normalized, then it is use for the estimation of the joint
probability distribution function (PDF) of intensities in the images [159].
Mutual information as a registration measure is introduce in 1990 by Woods
et al. [153], [155] to defining the regions of similar tissue in between the different
modality of images. According to the book chapter [160], [161] mutual informa-
tion(MI) in the form of conditional entropy is defined for two images If and Im
by Eq n -8, Table-1 and defined in the form of joint entropy by Eq n -9, Table-1.
The MI has similar interpretation to the joint entropy. Both metrics reflect how
much information of one image has about another. Many researchers select the
MI rather than joint entropy for image registration because MI is avoid favor-
ing a transformation that forces the image so for apart that only background
is contained in the overlap region. According to the formula for joint entropy
Eq n -7, Table-1, the joint entropy is minimized at two points, when Se (If ) is
small and Im is dependent on If but in case of maximizing the MI by max-
imize Se (If ), while at same time limiting Se (If |Im ), where Se (If |Im ) is joint
Shannons entropy of the joint probability histogram. The maximal entropy of
Se (If ) ensures that the overlap region between two images contains most of the
images, including complex parts that increase individual entropy [162],[163].
MI is provide some improvement to solve the overlap problem but it is fail
for some type of clinical image that contains a large amount of noise or air
around the outside of the subject that means changes in overlap of very low
intensity regions of the image can disproportionately contribute to the mutual
information. Alternative option to improve the performance of mutual informa-
tion is normalization schemes. There are some schemes for normalization of MI
(NMI) have been proposed in journal articles [3],[164]. The Studholme et al.
[164] proposed a method for normalization of MI that overcome the sensitivity
of mutual information to change in image overlap and NMI is calculated by
using Eq n -10, Table-1

27
Table 1: Equation used for different similarity measure

Similarity Formula Eq n
measure
1

N
SSD SSD= N |If (i) − Im (i)|2 (1)
i
where ∀i  If ∩ Im

1

SAD SAD= N |If (i) − Im (i)| (2)
i
where ∀i  If ∩ Im


(If (i)−If ) (Im (i)−Im )
i
CC CC=   (3)
{ (If (i)−If )2 (Im (i)−Im )2 }1 /2
i i

σR
VIR VIR= μR (4)
1
2
where σR = (R(i) − μR ) N
i

μR = N1 R(i)
i
and R(i) = Im (i)/If (i)

nf (a) σm (a)


PIU PIU= N μm (a) (5)
a{a}

where f ≡ If and m ≡ Im
nf (a) is the number of voxels in image f
and a is partition in image f


SE Se = − pi log pi (6)
i

JSE JSe (If , Im ) = Se (If ) + Se (Im |If ) (7)

MI in the form of conditional entropy


MI MI(If , Im ) = Se (If ) − Se (If |Im ) (8)

28
Table 1: Equation used for different similarity measure

Similarity Formula Eq n
measure

MI in the form of joint entropy


MI MI(If , Im ) = Se (If ) + Se (Im ) − Se (If , Im ) (9)

Se (If )+Se (Im )


NMI NMI(If , Im ) = Se (If ,Im ) (10)

4. Optimization Procedure

Optimization is an important paradigm that is mostly used in all appli-


cations of engineering and trying to optimize something either to minimize the
cost, energy and complexity or maximize the performance, efficiency, and profit.
For better utilization of limited resource and time the optimization is practi-
cally important [165], [166]. In medical image registration, to minimize the
difference between two images and produce the best transformation to match a
moving image (Im ) to a fixed image (If ) is the main objective of optimization.
In other word registration is an optimization problem of finding a coordinate
transformation T(x) that makes Im (T(x)) spatially aligned with If (x).
To finding an optimal transformation could be much expensive with reference
to computation time, which is a big disadvantage for many clinical applications.
For example in case of lung cancer screenings with high-resolution 3-D images
require a fast registration algorithms [167], the brain shift estimation based
on intra-operatively acquired ultrasound [168], real-time and fast registration
also require in external radiotherapy. Optimization is an iterative procedure in
which iteration refines the parameter value based on the fitness computation
[169]. The cost function C(T;If , Im ) is used to define the quality of alignment
which measures the similarity of the fixed image and the moving image. The
similarity is inversely proportional to the cost function. The transformation of

29
Figure 4: Type of Optimization Method

coordinate that relates the both images is estimated by iteratively minimizing


the cost function with respect to T. The registration problem is mathematically
formulated as an optimization problem in which the cost function C is minimized
with respect to T. The optimization problem is formulated as [50]:

δ
= arg min C(δ; If , Im ) (16)
δ

where the vector δ


contains the set of transformation parameters and subscript
δ indicates which parameter is used from set of parameters to optimize. The
optimization problem as shown in Eq n -16 is solved with the help of an iterative
optimization method. There are various optimization methods are used in image
registration as shown in figure-4.
There is various cost function C have been proposed in the literature. Com-
monly used intensity-based cost functions are the mean squared difference (MSD)

30
[170], [171], normalized correlation (NC), [172], mutual information (MI) [3],
[173], [40], [15], and normalized mutual information (NMI) [8], [164]. Some-
times, a regularization term is added to the cost function, in order to penalize
undesired deformations [8],[174], [175]. The cost function is formulated accord-
ing to which similarity measure used. For example cost function for MSD

1 
M SD(Tδ ; If , Im ) = (If (x) − Im (Tδ (x)))2 (17)
N
x∈Υf

where Υf denotes the fixed image domain, and N the number of voxels x sampled
from the fixed image domain. This cost function is only suited for two images
with equal intensities, i.e. for images from the same modality. NC is less strict;
it assumes an affine relation between the intensity values of the fixed and moving
image. MI and NMI assume only a statistical relation between the intensities of
the images. It is widely used for nonrigid registration technique which is based
on maximization of the mutual information similarity measure, in combination
with a deformation field parameterized by cubic B-splines [34],[176]. Therefore
it is suited not only for mono-modal, but also for multi-modal image pairs. The
approach can be formulated by Eq n -16 In this case δ
represents the parameter
vector containing the B-spline coefficients which define the deformation field.
This cost function contains multiple local minima and negated mutual informa-
tion similarity metric, So for reduce the number of local minima a regularization
term (R) can be added to the cost function as show in eq n -18:

C(δ; If , Im ) = −M I(δ; If , Im ) +ω Rδ (18)

where ω is a weighting factor for the regularization term R and the R is defined
in term of volume preserving penalty [8], the elastic energy [125] and in term
of curvature [174]. Minimization of the cost function is depends on the set of

with the help of
parameters so to determine the optimal set of parameters δ,
iterative optimization strategy as follows:

δk+1 = δk + βk Θk , k = 0, 1, 2, 3, ...... (19)

31
where Θk is the search direction at kth iteration and βk is scalar gain factor
which is used for controlling the step along the search direction. The Θk and
βk jointly able to select the sequence of parameters δk that converges to local
minima of the cost function. The optimization methods are differentiated by
how Θk and βk computed. The iterative optimization strategy and their formula
for computing Θk and βk are described in respective section.

4.1. Gradient Descent

The optimal search method converges slowly towards the local minimum.
Because they are use the incomplete objective function at the current sampling
point [177]. GD approach is also known as steepest descent. For obtaining
the local optimum, GD travels in the opposite direction to the gradient of the
objective function [178]. An iterative optimization strategy for GD is shown in
Eq n -20. GD method is defined in two ways based on their gain factor [179].
One is gradient descent decaying (GDD) where the gain factor βk is defined as
a decaying factor as shown in Eq n -21 , which makes GD is more comparable to
the stochastic gradient descent method. Other is known as GDL where the gain
factor βk is determined by an inexact line search routine, which makes GD, is
more comparable to the QN and NCG methods.

δk+1 = δk − βk Θ(δk ) (20)

Where Θ(δk ) is derivative of cost function evaluated at δk , the current position


δk .
β
βk = (21)
(k + A)α
Where β > 0, A >= 1 and 0 <= α <= 1 are user defined constants.

4.2. Quasi-Newton

Quasi-Newton (QN), Gauss-Newton (GN), Levenberg-Marquardt (LM) are


a class of nonlinear optimization methods and based on the Newton-Raphson
algorithms in which the Hessian matrix is used for cost function and evaluated

32
at different parameters. The Newton-Raphson algorithm is given by:

δk+1 = δk − [H(δk )]−1 Θ(δk ) (22)

where the H(δk ) is Hassian matrix of cost function and evaluated at δk . The hes-
sian matrix is a square matrix of second-order partial derivatives of a function.
It provides a better theoretical convergence property with respect to gradient
descent but the computation of hessian and its inverse can be difficult in prac-
tice especially in case of high dimensional optimization problems like nonrigid
registration. The Gauss-Newton method approximates the hessian and the con-
vergence of GN to a local minimum can be rapid if the approximated hessian
term dominates the value of the full Hessian evaluation by Newton method [179].
The Levenberg-Marquardt optimization method is improvement over GN
method by resolving the weakness of GN, namely it takes into account whether
the error in approximating the hessian gets better or worse after each iteration
[180],[181]. In LM methods using the trust regions approach instead of step-size
line search. The rate of convergence for LM method can be as similar as GN
method and the comparable cost is the same.
Quasi-Newton method based on an approximation to the inverse of the hes-
sian, where in each iteration the approximation is updated and second-order
partial derivatives are not needed for this update. The QN method is imple-
mented with combination of an inexact line search routine and a gain factor βk .
The QN method follows the recurrence:

δk+1 = δk + βk Lk Θ(δk ) (23)

where Lk is an approximation to the inverse of hessian and the βk is determine


in such a way that satisfied the strong Wolfe condition:

C(δk+1 ) ≤ C(δk ) + c1 βk ΘTk Θ(δk ) (24)

|ΘTk Θ(δk+1 )| ≤ c2 |ΘTk Θ(δk )| (25)

where c1 and c2 are user defined constants satisfying 0 < c1 < c2 < 1. The Lk
sequences are constructed in many ways like Davidon-Fletcher-Powell (DFP),

33
Symmetric-Rank-1 (SR-1) and Broyden-Fletcher-Goldfarb-Shanno (BFGS) and
Limited memory BFGS (LBFGS) approach. Where the BFGS approach is very
efficient for many application which is experimentaly proved in [179] and it uses
the following recurrence for construction os Lk sequences.
   
θφT φθT θθT
Lk+1 = I − T Lk I − T + T (26)
θ φ θ φ θ φ
where I represents an identity matrix, θ = δk+1 −δk and φ = Θk+1 −Θk . But for
large problems it is much expensive with respect to memory so limited memory
BFGS is proposed by J. Nocedal [182]. The LBFGS stores only selected vectors
that represent the approximation implicitly and it requires a linear memory. So
the LBFGS method is well suited for optimizing the problem that having large
number of variables [179],[183].

4.3. Nonlinear Conjugate Gradient (NCG)

Conjugate gradient (CG) methods are a class of unconstrained optimization


algorithms which requires a low memory and having strong local and global
convergence properties. There are two type of CG methods one is linear CG
method [184] which was designed for solving a linear equation based optimiza-
tion problem. Other is nonlinear conjugate gradient method which is suitable
for optimizing nonlinear functions [179]. A nonlinear conjugate gradient method
generates a sequence Xk , k ≥ 1, starting from an initial guess X0 using the re-
currence:
δk+1 = δk + βk Θk , (27)

where the positive step size βk is obtained by a inexact line search, and the
directions Θk are generated by the rule:

Θk+1 = −hk+1 + Ωk Θk , where Θ0 = −h0 (28)

Here Ωk is the CG update parameter and the gradient hk = ∇C(δk ) of C at δk


is a row vector and hk is a column vector. Different CG methods correspond
to different choices for the scalar update parameter Ωk . If cost function C is
a strongly convex quadratic, then all choices used for the update parameter

34
are equivalent with an exact line search. But, for non-quadratic cost functions,
each choice for the update parameter gives different performance. So the hybrid
methods are proposed by Y.H.Dai [185] to improve the performance of CG
algorithms, which dynamically adjust the formula for Ωk as the iterations evolve,
and a method based on the recent update parameter Ωk which is as follows:

Ωk = max(0, min(ΩHZ DY
k , Ωk )) (29)

where
gk+1 2
ΩDY
k = ΘTk yk

 T
yk 2 gk+1
ΩHZ
k = yk − 2Θk ΘT ΘT
k yk k yk

yk = gk+1 − gk and   denote the Euclidean norm.


The jamming or loss of descent or convergence failure problems arises with
the many choices. But the scheme proposed by Dai and Yuan provides descent
for a Wolfe line search, and the scheme proposed by Hager and Zhang provides
sufficient descent independent of the line search. Hybrid schemes are very effi-
cient and having excellent performance as compared to several different methods
in practice.

4.4. Stochastic gradient Descent (SGD)


Stochastic gradient Descent (SGD) method is an important class of stochas-
tic approximation. Some well-known techniques like Kiefer-Wolfowitz (KW),
Robbins-Monro (RM) and simultaneous perturbation (SP) are special cases of
SGD. The KW, RM, SP follows the same recurrence as mention in Eq n -30
which is same as deterministic gradient descent with the distinction that the
k )
cost function Θ(δk ) is replaced by an approximation Θ(δ

k )
δk+1 = δk − βk Θ(δ (30)

A SGD method is normally applied when computation of the exact derivative


is very costly. An approximation of the exact derivative is decrease the com-
putation time per iteration, but may be have a negative effects on the speed of

35
convergence.There are three different criteria in stochastic gradient method for
k ) , which differentiate the KW, RM, and SP.
defining Θ(δ

4.4.1. Kiefer-Wolfowitz (KW)


This method proposed by Kiefer and Wolfowitz in 1952 [186], is based on
finite difference approximation of derivative and defined as:

k )]i = C(δk + βk ei ) − C(δk + βk ei )


[Θ(δ (31)
2βk

k )]i is the ith element of Θ(δ


where [Θ(δ k ) and βk is a small scalar and commonly

defined by expression:
β
βk = (32)
(k + 1)α
where β > 0 and 0 <= α <= 1 are a user-defined constants. The ei in Eq n -31
is an unit vector containing only zeros except for the ith element, which equals
to one. The KW method assumes the approximation of cost function values say
+ −
C̃ki and C̃ki is an approximation of cost function C(δk + βk ei ) and C(δk − βk ei )
respectively, which is calculated by:

+ +
C̃ki = C(δk + βk ei ) + ξki

and
+ −
C̃ki = C(δk − βk ei ) + ξki (33)
+ −
where ξki and ξki is an approximation errors which may arises by external source.
By substituting the values of cost function by Eq n -33 in Eq n -31 yield the KW
algorithm:
+ −
k )]i = C̃ki − C̃ki
[Θ(δ (34)
2βk
It is observe that in case of high dimensionality problem, the KW algorithm
the evaluation of cost function for each iteration is two times of their number
of parameters (P).

36
4.4.2. Simultaneous Perturbation (SP)
Simultaneous Perturbation is proposed by Spall [187] and it is based on a
efficient and easily implemented its approximation to the gradient and this gra-
dient approximation uses only two evaluation of cost function that means it is
independent from the number of parameters being optimized, whereas the KW
algorithm requires 2P times cost function evaluations per iteration to approxi-
mate the gradient [188]. Let C̃k+ and C̃k− is two approximation of cost functions
which is calculate by:
C̃k+ = C(δk + βk k ) + ξk+

and
C̃k+ = C(δk − βk k ) + ξk− (35)

where k denotes the random perturbation vector Generate by Monte Carlo and
each element of k assigned ±1 in each iteration by using Bernoulli distribution
with equal probability. The ξk+ and ξk− is represents an approximation errors.
k )]i of ith element is calculated by:
So the derivative vector [Θ(δ

k )]i = C̃k+ − C̃k−


[Θ(δ (36)
2βk [ k ]i

4.4.3. Robbins-Monro
The Robbins-Monro approach applies in root-finding/stochastic gradient-
based settings. The KW and SP is a special case of Robbins-Monro (RM)
because KW and SP are construct a derivative estimate on the basis of approxi-
mate evaluations of the cost function but RM can not specify how the derivative
is computed. So an approximation of the derivative of the cost function is cal-
culated by:
Θ˜k = Θ(δk ) + ξk (37)

The RM algorithm is equivalent to the deterministic gradient algorithm when


the value of ξk is zero in each iteration.

4.5. Adaptive Stochastic Approximation


The concept of adaptive stochastic approximation has been proposed by
Venter [189]. An adaptive SA algorithm is a simple method for estimating the

37
hessian matrix, which is concurrently estimating the primary parameters of in-
terest and based on the simultaneous perturbation idea. The ASA approach
applies in both the root-finding gradient-based (RM) and gradient-free opti-
mization (KW).The ASA algorithm requires only a small number of gradient
measurements per iteration and independent from the dimension for adaptively
estimate the parameters of primary interest and Hessian [190].

4.6. Evolutionary Approach

The evolutionary approach is a non-deterministic approach that guides for


searching process with objective to find optimal solution having a tendency to
get stuck in global optimum. It is prominent and successful technique for solving
a complex and computationally hard numerical optimization problems. The
evolutionary approach includes a genetic algorithms (GAs), simulated annealing
(SA), differential evolution (DE), and swarm intelligence (SI). Normally GA,
SA and SI based approaches are used for solving the optimization problem of
intensity based medical image registration. As mention in Table-2,3,4 and 5,
there are 16 authors out of 27 have been used the evolutionary approach to solve
the optimization problem for intensity-based medical image registration.
The genetic algorithm has been inspired by the principle of natural selection.
The GA address the optimization problems such as minimization of non-linear
objective function, by implementing the iterative process of mutation followed
by selection in each iteration and generate a new child from their parents then
their fitness value is calculated by using fitness function and finally select the
better candidates that becomes the parents for next generation [191],[192].
The simulated annealing is a neighborhood based optimization method and
inspired from annealing in metallurgy that involves heating process and con-
trolled cooling of material [193], [194]. In each iteration the SA consider a
solution S from the neighborhood of the current one C and calculate their qual-
ity, if the quality of S is higher than C then S becomes the current solution
and end the iteration. If S is having a poor quality than C then S is accepted
q(S)−q(C)
with probability e t . Where q(S) and q(C) is the quality of S and C

38
respectively and t is the temperature parameter, which controls the probability
of acceptance of poor quality solution. Initially the temperature parameter t
is set a high value and decreases in each iteration. The iteration is continuous
until the temperature parameter close to zero or reaches the stoping condition.

4.7. Particle Swarm Optimization

Particle swarm optimization (PSO) is stochastic optimization technique de-


veloped by Dr. Eberhart and Kennedy in 1995. PSO algorithm was inspired
by social behavior of bird flocking or fish schooling. PSO optimizes an objec-
tive function by attempting a population-based search. The population consists
of potential solutions, known as a particle, which represents the birds in flocks.
These particles are initialized randomly and freely move across the multi dimen-
sional search space. During movement, each particle updates its own velocity
and position based on the best experience of its own and the population for
achieving a better position for itself [195].All particles having its own fitness
value that is evaluated by fitness function to be optimized [196]. The position
corresponding to the best fitness is known as personal best value Pb and the
global best value Gb . The global best Gb means the overall best value out of all
the particles in the population.
In PSO values of particles are updated according to the similarity measure.
Particles movements are based on its velocities and current positions which up-
dates in each iterations. The velocity and position of each particle are calculated
by using following equations:

(t−1) (t−1) (t−1)


Vidt = ωVid + a1 υ1 [P bid − xid ] + a2 υ2 [Gbd − xid ] (38)

(t) (t−1)
xid = xid + Vidt (39)

Where t is number of iteration, the value of d varies from 1 to D (the number of


dimension in the search space) i.e. d = 1, 2, ...., D and the value of i varies from
1 to S (size of population) i.e. i = 1, 2, ...., S. The xid , Vid , P bid are the position,
velocity vector and personal best of the ith particle in dth dimension respectively

39
and Gbd is the global best of dth dimension. The acceleration constants a1 , a2
are helps to control the particle’s search towards its own position and best
neighboring positions and υ1 , υ2 to maintain the diversity of the population,
which is uniformly distributed random numbers lies between 0 to 1. ω is inertial
weight used to improve the performance of the PSO and it may be a positive
constant or even a positive linear or nonlinear function of time [197]. The value
of velocity vector in each dimension is lies between close interval [−Vmax , Vmax ]
which ensured that particle does not exit from allowed searching space. The
algorithm is run until it reaches to a predetermined number of iterations or the
minimum error criterion is reached [195].

5. Discussion

The objectives of the present discussion is to analyze the various geometric


transformation methods, similarity measures and optimization techniques pre-
sented in this review which is useful for intensity-based medical image registra-
tion. The short summary, outcomes and methodology used by various authors
are described in Table-2,3,4 and 5.
Initially start the discussion about the geometric transformation which is
the backbone of all registration techniques. The geometric transformation con-
tains important information about geometric differences between the medical
images. Some times due to lack of information, it is difficult to understand the
contents of medical images. There are two main factors to control the image
registration accuracy. One of them is control point correspondence and other
is the transformation function used for finding geometric differences between
the fixed and moving image. This book contains a survey of various types of
rigid and nonrigid geometric transformation used in medical image registration.
Generally the geometric transformation is selected on the basis of available in-
formation about the geometric difference between the images. In case, when
the sufficient information about geometric differences between the images is
available then mostly, the linear or projective transformation is used for image

40
registration. But, the deformable transformation is not feasible in this case due
to their computational cost and provides poor registration accuracy. If the in-
formation about the geometric differences between the images is sufficiently not
available then the transformation function plays an important role to identify
the geometric difference between the images. So in this case the deformable
transformation such as physical model, basis function based geometric transfor-
mations are used. The physical model based medical image registration provides
a realistic solution. The linear elasticity based approaches are most suited for
small deformations rather than a large deformation whereas the fluid flow is
more feasible for large deformation and preserves the topology but it is not
able to handle the elastic component of soft tissue deformation and also these
approaches are computationally expensive.
The basis function based models such as RBF, EBS, FFD and piecewise
methods are analyzed and found that the RBFs, including TPS, produce the
poor results due to some reasons. The first reason is related with use of basis
function. The basis function is responsible to determine the radially symmetric
component of transformation but if the component of transformation is asym-
metric then large number of errors is found. Second reason is related with the
monotonically increasing logarithmic function used in TPS. Due to this func-
tion the errors in correspondences or image deformations are scattered over the
entire re-sampled image. The last reason, when the number of correspondence
point is large enough then solving the PDE equations to find the component
of transformation is practically not feasible. Hence, the RBF including TPS
approach is feasible when small number of correspondence points exists. Ac-
cording to the comparative study by Franke [86], it was found that MQ basis
function followed by TPS produced the best accuracy in the interpolation of
randomly spaced data. It is difficult to decide which method is accurate and
best for a particular application due to lack of existing comparative analysis for
different transformation models.
The next component of discussion is the similarity measures that are used
to identify the alignment of the images. The similarity measures based on

41
intensity difference (SSD, SAD, MSD etc.) are computed from the voxel in-
tensity of the corresponding structures of both images so lower SSD indicates
better alignment. The CC and its related similarity measures are based on the
assumption that there is linear relationship between the intensities of the cor-
responding structure of both images so larger CC indicates better alignment.
The SSD, SAD, CC and its derived similarity measures are mostly appropriate
for mono-modal image registration and these measures are based on voxel to
voxel stationarity of the intensities and inaccuracy of independence [198]. The
VIR and PIU similarity measures are appropriate for serial intra-modality and
inter-modality image registration respectively. The VIR method maximizes uni-
formity by minimizes the normalized standard deviation. In some cases it gives
good result but preprocessing is required that remove some anatomy. The PIU is
an improved version of VIR, which overcome these problems. The VIR and PIU
can be useful for both the intensity based as well as feature based registration
whereas the information theory based similarity measures are appropriate for
intensity based image registration and applied for both intra-modality as well as
inter-modality image registration and provide maximum value when the input
images are registered correctly. The SSD, CC, VIR and PIU are the distance
based similarity measures and the information theory based similarity measures
are based on probabilistic approaches. Generally distance based approaches can
be used in feature based image registration whereas the probabilistic approach
based similarity measure is useful in intensity-based image registration.
In comparison to the CC and its derived similarity measure the MI have
big advantage because it is able for accurate registration of multi-modal images
and also able to measure the alignment of image signals which is helpful to
predict the other but CC fails for the same. If how to transform one image
signals to other is known then it is simple to apply that transformation and
then correlation performs well on the normalized image signals.
The joint entropy (probabilistic approach) based similarity measures has
some advantages over PIU such as it minimizes the spread of clusters in two
dimensions rather than one and minimizing entropy does not require that the

42
histograms are uni-modal in the way to minimizing variance. So the joint en-
tropy would be normally applicable to multi-modality registration. The MI is
reliable and robust similarity measure and used by many authors as shown in
Table-2,3,4 and 5. MI is based on the Shannon entropy which is computed from
joint probability distribution of the voxel intensities, but the other divergence
measures have been also used in literature in place of Shannon entropy like Reny
entropy [199], Havrda-Charval entropy [200] and Tasallis entropy [201], [202],
which may provide better result for some applications (excluding whether these
entropy is not used) of medical image registration in future.
Finally, discuss about the optimization technique that is basically used to
find the optimum transform parameters/function for aligning the images. There
are various optimization algorithms which exist in literature as described in
previous section. The good optimization algorithm is able to determine the
reliable transform parameters quickly. Generally, the intensity-based medical
image registration technique is used for non-rigid image registration and it is
more difficult to design an optimizer for non-rigid registration because the more
parameters are required to describe it. Due to more parameters the optimizer
takes much amount of time to determine the optimum. Due to local minima
problem there are more options to select a set of parameters. There are two
methods to determine the transform parameter, one is directly, that is explicit
way from available data and second is searching method that determine the
parameters by using optimum cost function in the search space.
The first derivative based optimization algorithm such as CG and LM is
normally used [3] but they are not able to obtain global optimum because the
intensity based similarity function and their cost functions are not optimal.
Hence, for global optimization the genetic algorithm and simulated annealing
are used but due to their low convergence rate these techniques are not highly
useful in intensity-based medical image registration. To obtain the fast and
robust convergence rate a multi resolution and multi scale optimization algo-
rithms are used in [173], [170], [203], [204]. The gradient and non-gradient based
optimization technique such as CG, QN, LM, Powell, simplex are multi reso-

43
lution optimization methods are compared by Maes et al. [203] and state that
these methods was normally used for affine transformation based registration
technique. The Klein et al. [183] compared the performance of all optimization
technique except the PSO and focuses to the computation time reduction per
iteration and experimentally show that the RM method is the best optimization
method for most application.
The particle swam optimization (PSO) technique for multi modal medical
image registration proposed by M. P.Wachowiak et al. [205],[206]. The PSO
method is fast in the earlier search stage but it easily converges to a local best
position instead of the global best position. To overcome these deficiencies the
various improved methods are proposed like Improved PSO [197], Hybrid PSO
[207], Niche Particle Swarm Optimization (NPSO) [206], quantum-behaved PSO
algorithm (QPSO) [208] and diversity-controlled QPSO (DQPSO) [209]. Author
D. Zhao et al. in 2013 [210] proposed an another novel algorithm RMI-SAPSO
(regional mutual information-simulated annealing particle swarm optimization)
which select regional mutual information as objective function and then SAPSO
used to optimize the RMI, which improve the accuracy of the medical image
registration in comparison to traditional MI-SAPSO and other PSO algorithms.
Shuihua Wang in 2011 proposed a rigid registration technique for non-medical
image based on artificial Bee colony optimization (ABC) algorithm and exper-
imentally show that their approach is efficient and robust in comparison to
GA and PSO [211]. Author Yang [212] proposed a bat optimization algorithm
(BOA) in 2011, which was based on bat behavior. This algorithm is efficient for
lower-dimensional optimization problems, but it may be problematic for higher-
dimensional problems due to initially fast convergence rate. So author Fister
at al. in 2013 [213] proposed a hybrid bat algorithm (HBA). The HBA have
improved the performance of the bat optimization algorithm. So in future the
ABC and HBA algorithm may be beneficial to solve the optimization problem
for intensity-based medical image registration.

44
Author/ ransfor- Similarity Optimiza- Modality/ Short summary and outcomes
tion
year mation Measure Algorithm Image Used
Studholme Rigid MI multi- MR,PET Fast efficient and robust multi-resolution optimization approach used for se
et al.(1996) resolution Brain the six dimensional parameter space of rigid body registration.
F. Maes et 3D Rigid MI Brent, MR, CT, Accurate, highly robust and completely automatic registration for multi-m
al. (1997) Powell PET Brain medical images.
D.Rueckert Affine NMI Gradient 3D-MRI Author proposed a new approach for the nonrigid registration of contrast-en
et al. Descent Breast breast MRI. It is better to recover the motion and deformation of the brea
(1999) respect to rigid/affine registration algorithms.
Alexandre rigid, MI LTS, MR, CT Performance was good and provides consistent results across modalities. If pu
G. et al. affine RLS Brain constraints on the intensity transformation that relates the images then pe
(2001) better search of the parameter space.
G. C. Ka- Affine MI GA Pow- SPECT, The author had been to develop an automatic technique for the medical imag
gadis et al. ell CT Brain tration of real CT and SPECT images based on either surface-based or volum


(2002) algorithms. Author show that the volume-based method consistently outp
with respect to surface-based method.
R. He et al. 3D Rigid MI DIRECT MRI Brain Author proposed a global optimization technique for intensity based medica
(2002) GA registration based on MI. It combines the genetic algorithm in continuou
search space with dividing rectangle (DIRECT), which theoretically gua
global optimization and was efficient in small search space.
R.Shekhar Rigid, MI Simplex US Brain Investigated the registration of ultrasound volumes based on the mutual infor
et al. affine abdomen measure. This technique originally applied to multi-modality registration
(2002) thorax capture range becomes smaller with the complexity of transformation an
enough for most practical applications.
D.Loeckx 2D affine Pattern SA 2D X-ray, Author proposed an approach for non-rigid alignment of thorax X-ray im
et al.(2003) B-spline intensity CR Thorax detect interval changes by temporal subtraction and provide good result
Author/ ransfor- Similarity Optimiza- Modality/ Short summary and outcomes
tion
year mation Measure Algorithm Image Used
T.Rohlfing Rigid NMI downhill pre and Author proposed intensity-based nonrigid registration algorithm by using
et al. Non- simplex post- regularization term to constrain the deformation.The incompressibility penal
(2003) rigid contrast can produce free-form deformations of MR breast images that substantially
MR Breast motion artifact with negligible volume loss from contrast enhancing structu
Wachowia 3D Rigid NMI PSO MRI, US, Author proposed a hybrid particle swarm optimization (HPSO) that provid
et al.(2004) Powell CT Brain, accurate registration in compare to the evolutionary approach in many ca
abdomen comparatively provide better convergence rate.
X. Xu et al. 3D Rigid MI, Gra- DE Pow- MRI, CT Author proposed the improved version of powell’s direction set method (
(2004) dient ell Brain that was strongly dependent on initial transformation so author compute th
Info. transformation without using a priori knowledge of final solution by using D
that improves the effectiveness of this approach.

46
Z. Xie et al. Affine CC optical MR,CT, Author proposed free-form deformation based registration approach by m
(2004) flow PET Brain them at increasing levels and also compared their results with Iterative close
algorithm and optical flow and implemented on both 2D and 3D image.
R.Shekhar Rigid, MI Downhill Real-time Author proposed a ”3-D stress echo-cardiography” that overcome the lim
et al.(2004) affine Simplex 3D US of stress echo-cardiography which is normally use for establish 3-D stres
Heart cardiography as a routine clinical procedure.
Talbi et 2D Rigid MI PSO MRI, Author proposed an algorithm to solve a problem of multi-modal medica
al.(2004) SPECT registration by using HPSO with differential evolution operator is a combin
CT, Brain robust entropy and power search strategy and provided bell-shaped mutation
on the population diversity, which overcome the problem of local minima.
Matsopoulos affine MI Simulated ICG Retina Author proposed an automatic method for registering multi-modal reti
et al. anneal- ages.The novelty of this approach are the vessel centerline detection, bifur
(2004) ing extraction process only on the reference image and proposed a novel imple
Author/ ransfor- Similarity Optimiza- Modality/ Short summary and outcomes
tion
year mation Measure Algorithm Image Used
X. Li et al. 3D rigid MI SA MRI Brain Author proposed a MI based registration method by using SA optimizatio
(2005) author apply SA based registration methods on 28 slices of MR images of
tients and show the successful registration for all patient images. Where as
and downhill simplex provide successful registration for 11 of 9 patients
So MI based registration method with SA optimization was robust and rel
comparison to Powell,s and DS optimization methods.
J.Du et al. 3D rigid DRPIU SA, PET, MRI Author proposed a novel similarity metric DRPIU for multi-modal image regi
(2006) Powell’s Brain and compare DRPIU with PIU and MI. The author combine forward and
transformation to reduce the negative effects of outliers in medical images.
Li et al. 2D rigid MI PSO MRI, Brain Author proposed four different mode (based on the velocity updates crite
(2007) PSO that globally optimize the issues of intensity-based medical image regis

47
Author also focuses on the evaluation and application of PSO optimization f
transformation of multi-modality images. The successful rate and compu
time of proposed PSO is relatively best and the efficient.
Winter et 3D-Rigid Application CMA-ES CT, 3D-US Author compared three gradient-based and one evolutionary optimization p
al.(2008) oriented Spine discussed their suitability to solve the problem in image registration. Experim
proved CMA-ES was usefull for general medical image registration.
Chen et 2D-Rigid MI PSO CT,MRI Author proposed a new approach for medical image registration using particle
al.(2008) Liver optimization and describe the effectiveness of PSO for both rigid and n
medical image registration.
X.Huang et Rigid MI Gradient MR/CT , Author proposed an efficient method for registering real-time 3D ultrasound
al.(2009) Heart US) to high-quality dynamic 3D MR/CT images of the beating heart. This te
reduces major limitation of registration problem such as the correct regi
Author/ ransfor- Similarity Optimiza- Modality/ Short summary and outcomes
tion
year mation Measure Algorithm Image Used
R. Bha- Non- SSD, MI IS-SA CT, MRI Author used an important sampling framework that was based on an edge-dep
galia et al. rigid Brain adaptive sampling distribution which have been used in intensity-based regi
(2009) algorithms and focuses on registration methods that estimate parameterize
models by solving an optimization problem.
Z. Ding et Rigid, Euclidean MAP ap- MR, Brain Author proposed an unified fiber bundling and registration (UFIBRE) fram
al. (2009) affine distance proach This approach is highly efficient and providing a potential routine for group a
of WM fibers in brain.
A. Mayer et Affine MSD PBT MRI, Brain Author proposed a supervised framework for the automatic registration and s
al. (2011) tation of white matter (WM) tractographies extracted from brain DT-MRI.
the first approach that automatically segment the Optic Radiation (OR) and

48
a quantitative validation.
Wang et al. 2D simi- NMI GA CT,MRI Author proposed an improved GA approach by combining operation migrati
(2011) larity Brain Nelder-Mead simplex method. This approach was used for preventing pre
convergence and improve their accuracy. It is successfully implemented on in
based medical image registration.
D.Zhou et 3D rigid MI QPSO CT, PET, Author proposed a diversity controlled revised quantum-behaved PSO (DR
al. (2011) MRI Brain that prevent the tendency of evolutionary algorithm to be trapped into local
as a result of decline in diversity. Author also investigates QPSO , RQP
show the improved performance of DRQPSO over PSO and QPSO for in
based medical image registration.
Valsechi et 3D affine MI GA MRI Brain Author proposed a novel intensity-based medical image registration approac
al. (2013) on r-GA and tested on 16 real MRI images registration. This methodology co
with four image registration techniques based on I-ICP, SS, Dyn-GA and G
6. Conclusion

The comprehensive review of methodologies involved in the design and de-


velopment of an intensity based medical image registration was presented. This
book focuses on the review of various methodologies available in literature for the
various steps used for the design of the intensity based registration system which
include geometric transformations, similarity measures, and optimization tech-
niques used. Here we also include a definitive documentation on mathematical as
well as theoretical aspects of different methodologies used in each design step of
the said system and also highlights the merits and limitations of each. Further,
this book presented the summary and outcomes of various intensity based med-
ical image registration systems applied on medical images of human body parts
such as brain, heart, abdomen, retina, breast etc. arising from various medi-
cal imaging modalities which include Computerized Tomography-Computerized
Tomography (CT-CT), Computerized Tomography-Positron Emission Tomog-
raphy (CT-PET), Positron Emission Tomography-Magnetic Resonance Imaging
(PET- MRI), and Ultrasound-Computerized Tomography (US-CT). Finally, the
observations and discussions presented in the book are useful for taking deci-
sions for the selection of an appropriate methodology for the design of a specific
step of an intensity based medical image registration system.

References

[1] JB Antoine Maintz and Max A Viergever. A survey of medical image


registration. Medical image analysis, 2(1):1–36, 1998.

[2] Hava Lester and Simon R Arridge. A survey of hierarchical non-linear


medical image registration. Pattern recognition, 32(1):129–149, 1999.

[3] Frederik Maes, Andre Collignon, Dirk Vandermeulen, Guy Marchal, and
Paul Suetens. Multimodality image registration by maximization of mu-
tual information. IEEE transactions on medical imaging, 16(2):187–198,
1997.

49
[4] Arnaldo Mayer, Gali Zimmerman-Moreno, Ran Shadmi, Amit Batikoff,
and Hayit Greenspan. A supervised framework for the registration and
segmentation of white matter fiber tracts. IEEE Transactions on medical
imaging, 30(1):131–145, 2011.

[5] George K Matsopoulos, Pantelis A Asvestas, Nicolaos A Mouravliansky,


and Konstantinos K Delibasis. Multimodal registration of retinal im-
ages using self organizing maps. IEEE Transactions on Medical Imaging,
23(12):1557–1563, 2004.

[6] Raj Shekhar and Vladimir Zagrodsky. Mutual information-based rigid


and nonrigid registration of ultrasound volumes. IEEE transactions on
medical imaging, 21(1):9–22, 2002.

[7] Raj Shekhar, Vladimir Zagrodsky, Mario J Garcia, and James D Thomas.
Registration of real-time 3-d ultrasound images of the heart for novel
3-d stress echocardiography. IEEE Transactions on Medical Imaging,
23(9):1141–1149, 2004.

[8] Torsten Rohlfing, Calvin R Maurer, David A Bluemke, and Michael A Ja-
cobs. Volume-preserving nonrigid registration of mr breast images using
free-form deformation with an incompressibility constraint. IEEE trans-
actions on medical imaging, 22(6):730–741, 2003.

[9] Lisa Gottesfeld Brown. A survey of image registration techniques. ACM


computing surveys (CSUR), 24(4):325–376, 1992.

[10] Barbara Zitova and Jan Flusser. Image registration methods: a survey.
Image and vision computing, 21(11):977–1000, 2003.

[11] Aristeidis Sotiras, Christos Davatzikos, and Nikos Paragios. Deformable


medical image registration: A survey. IEEE transactions on medical imag-
ing, 32(7):1153–1190, 2013.

50
[12] Francisco PM Oliveira and Joao Manuel RS Tavares. Medical image reg-
istration: a review. Computer methods in biomechanics and biomedical
engineering, 17(2):73–93, 2014.

[13] Mohammed Essadiki. New technique for combining panchromatic and


multispectral spot images for multipurpose image-maps. Mert Rory
Sabuncu,Entropy-Based Image Registration. Phd Thesis.

[14] Jan Flusser and Tomas Suk. Degraded image analysis: an invariant ap-
proach. IEEE Transactions on Pattern Analysis and Machine Intelligence,
20(6):590–603, 1998.

[15] Josien PW Pluim, JB Antoine Maintz, and Max A Viergever. Mutual-


information-based registration of medical images: a survey. IEEE trans-
actions on medical imaging, 22(8):986–1004, 2003.

[16] Sascha EA Muenzing, Bram van Ginneken, Max A Viergever, and


Josien PW Pluim. Dirboost–an algorithm for boosting deformable image
registration: Application to lung ct intra-subject registration. Medical
image analysis, 18(3):449–459, 2014.

[17] Timo Makela, Patrick Clarysse, Outi Sipila, Nicoleta Pauna, Quoc Cuong
Pham, Toivo Katila, and Isabelle E Magnin. A review of cardiac image
registration methods. IEEE Transactions on medical imaging, 21(9):1011–
1021, 2002.

[18] Brian F Hutton, Michael Braun, Lennart Thurfjell, and Dennys Y Lau.
Image registration: an essential tool for nuclear medicine. European jour-
nal of nuclear medicine and molecular imaging, 29(4):559, 2002.

[19] Julian G Rosenman, Elizabeth P Miller, Gregg Tracton, and Tim J Cul-
lip. Image registration: an essential part of radiation therapy treatment
planning. International Journal of Radiation Oncology* Biology* Physics,
40(1):197–205, 1998.

51
[20] Erik HW Meijering, Wiro J Niessen, and MA Viegever. Retrospective
motion correction in digital subtraction angiography: a review. IEEE
Transactions on Medical Imaging, 18(1):2–21, 1999.

[21] Paul M Thompson, Roger P Woods, Michael S Mega, and Arthur W


Toga. Mathematical/computational challenges in creating deformable and
probabilistic atlases of the human brain. Human brain mapping, 9(2):81–
92, 2000.

[22] David J Hawkes, D Barratt, Jane M Blackall, C Chan, Philip J Edwards,


K Rhode, Graeme P Penney, Jamie McClelland, and Derek LG Hill. Tissue
deformation and shape models in image-guided interventions: a discussion
paper. Medical Image Analysis, 9(2):163–175, 2005.

[23] Matthieu Ferrant, Arya Nabavi, Benoıt Macq, Peter M Black, Ferenc A
Jolesz, Ron Kikinis, and Simon K Warfield. Serial registration of intra-
operative mr images of the brain. Medical image analysis, 6(4):337–359,
2002.

[24] Petra A Van den Elsen, E-JD Pol, and Max A Viergever. Medical image
matching-a review with classification. IEEE Engineering in Medicine and
Biology Magazine, 12(1):26–39, 1993.

[25] Haili Chui and Anand Rangarajan. A new point matching algorithm
for non-rigid registration. Computer Vision and Image Understanding,
89(2):114–141, 2003.

[26] Mert Rory Sabuncu. Entropy-based image registration. PhD thesis, Prince-
ton University, 2004.

[27] Hui Li, BS Manjunath, and Sanjit K Mitra. A contour-based approach to


multisensor image registration. IEEE transactions on image processing,
4(3):320–334, 1995.

[28] Azhar Quddus and Otman Basir. Wavelet-based medical image registra-
tion for retrieval applications. In BioMedical Engineering and Informatics,

52
2008. BMEI 2008. International Conference on, volume 2, pages 301–305.
IEEE, 2008.

[29] Harold S Stone, Jacqueline Le Moigne, and Morgan McGuire. Image


registration using wavelet techniques. In 26th AIPR Workshop: Exploiting
New Image Sources and Sensors, pages 116–125. International Society for
Optics and Photonics, 1998.

[30] A El-Ghazaw and Prachya Chalermwa. Wavelet-based image registration


on parallel computers. In Supercomputing, ACM/IEEE 1997 Conference,
pages 20–20. IEEE, 1997.

[31] Wen Cao, Bicheng Li, and Yong Zhang. A remote sensing image fusion
method based on pca transform and wavelet packet transform. In Neural
Networks and Signal Processing, 2003. Proceedings of the 2003 Interna-
tional Conference on, volume 2, pages 976–981. IEEE, 2003.

[32] Stefan Heger, Frank Portheine, Jorg AK Ohnsorge, Erik Schkommodau,


and Klaus Radermacher. User-interactive registration of bone with a-
mode ultrasound. IEEE engineering in medicine and biology magazine,
24(2):85–95, 2005.

[33] Martin Auer, Peter Regitnig, and Gerhard A Holzapfel. An automatic


nonrigid registration for stained histological sections. IEEE Transactions
on Image Processing, 14(4):475–486, 2005.

[34] David Mattes, David R Haynor, Hubert Vesselle, Thomas K Lewellen,


and William Eubank. Pet-ct image registration in the chest using free-
form deformations. IEEE transactions on medical imaging, 22(1):120–128,
2003.

[35] Isaac N Bankman and Serban Morcovescu. Handbook of medical imaging.


processing and analysis. Medical Physics, 29(1):107–107, 2002.

[36] Andrew P King, Kawal S Rhode, Y Ma, Cheng Yao, Christian Jansen,
Reza Razavi, and Graeme P Penney. Registering preprocedure volumetric

53
images with intraprocedure 3-d ultrasound using an ultrasound imaging
model. IEEE Transactions on Medical Imaging, 29(3):924–937, 2010.

[37] Charles R Meyer, Jennifer L Boes, Boklye Kim, Peyton H Bland, Gerald L
Lecarpentier, J Brian Fowlkes, Marilyn A Roubidoux, and Paul L Carson.
Semiautomatic registration of volumetric ultrasound scans. Ultrasound in
medicine & biology, 25(3):339–347, 1999.

[38] Perry E Radau, Piotr J Slomka, Per Julin, Leif Svensson, and Lars-Olof
Wahlund. Evaluation of linear registration algorithms for brain spect and
the errors due to hypoperfusion lesions. Medical Physics, 28(8):1660–1668,
2001.

[39] Koenraad Van Laere, Michel Koole, Yves D.Asseler, Jan Versijpt, Kurt
Audenaert, Filip Dumont, and Rudi Dierckx. Automated stereotactic
standardization of brain spect receptor data using single-photon trans-
mission images. Journal of Nuclear Medicine, 42(2):361–375, 2001.

[40] Paul Viola and William M Wells III. Alignment by maximization of mu-
tual information. International journal of computer vision, 24(2):137–154,
1997.

[41] George C Kagadis, Konstantinos K Delibasis, George K Matsopoulos,


Nikolaos A Mouravliansky, Pantelis A Asvestas, and George C Nikiforidis.
A comparative study of surface-and volume-based techniques for the au-
tomatic registration between ct and spect brain images. Medical Physics,
29(2):201–213, 2002.

[42] Mark Jenkinson and Stephen Smith. A global optimisation method for ro-
bust affine registration of brain images. Medical image analysis, 5(2):143–
156, 2001.

[43] Derek LG Hill, Calvin R Maurer Jr, Colin Studholme, J Michael Fitz-
patrick, and David J Hawkes. Correcting scaling errors in tomographic

54
images using a nine degree of freedom registration algorithm. Journal of
Computer Assisted Tomography, 22(2):317–323, 1998.

[44] Mark Holden, Derek LG Hill, Erika RE Denton, Jo M Jarosz, Tim CS Cox,
Torsten Rohlfing, Joanne Goodey, and David J Hawkes. Voxel similarity
measures for 3-d serial mr brain image registration. IEEE transactions on
medical imaging, 19(2):94–102, 2000.

[45] Josien PW Pluim and J Michael Fitzpatrick. Image registration. IEEE


Transactions on Medical Imaging, 22(11):1341–1343, 2003.

[46] Mark Holden. A review of geometric transformations for nonrigid body


registration. IEEE transactions on medical imaging, 27(1):111–128, 2008.

[47] Chaim Broit. Optimal registration of deformed images. 1981.

[48] Ruzena Bajcsy, Robert Lieberson, and Martin Reivich. A computerized


system for the elastic matching of deformed radiographic images to ideal-
ized atlas images., 1983.

[49] James C Gee, David R Haynor, Martin Reivich, and Ruzena Bajcsy. Finite
element approach to warping of brain images. 1994.

[50] Wencheng Wu and Mu Qiao. Preserving user applied markings made to


a hardcopy original document, July 16 2013. US Patent 8,488,181.

[51] Gary E Christensen, Richard D Rabbitt, and Michael I Miller. 3d brain


mapping using a deformable neuroanatomy. Physics in medicine and bi-
ology, 39(3):609, 1994.

[52] Christos Davatzikos and N Bryan. Using a deformable surface model to


obtain a shape representation of the cortex. IEEE transactions on medical
imaging, 15(6):785–795, 1996.

[53] Christos Davatzikos. Spatial transformation and registration of brain im-


ages using elastically deformable models. Computer Vision and Image
Understanding, 66(2):207–222, 1997.

55
[54] David W Shattuck, Stephanie R Sandor-Leahy, Kirt A Schaper, David A
Rottenberg, and Richard M Leahy. Magnetic resonance image tissue clas-
sification using a partial volume model. NeuroImage, 13(5):856–876, 2001.

[55] Karl Rohr, H Siegfried Stiehl, Rainer Sprengel, Wolfgang Beil, Thorsten M
Buzug, Jürgen Weese, and MH Kuhn. Point-based elastic registration of
medical image data using approximating thin-plate splines. In Visualiza-
tion in Biomedical Computing, pages 297–306. Springer, 1996.

[56] Amer Butt, Raj Acharya, Claudio Sibata, and KH Shin. Surface matching
of multimodality image volumes by a fuzzy elastic registration technique.
Computerized medical imaging and graphics, 22(1):13–23, 1998.

[57] Mike Fornefett, Karl Rohr, and H Siegfried Stiehl. Elastic registration
of medical images using radial basis functions with compact support. In
Computer Vision and Pattern Recognition, 1999. IEEE Computer Society
Conference on., volume 1, pages 402–407. IEEE, 1999.

[58] Jan Kybic and Michael Unser. Multidimensional elastic registration of


images using splines. In Image Processing, 2000. Proceedings. 2000 Inter-
national Conference on, volume 2, pages 455–458. IEEE, 2000.

[59] Pierre Moulin, Ravi Krishnamurthy, and John W Woods. Multiscale mod-
eling and estimation of motion fields for video coding. IEEE Transactions
on Image Processing, 6(12):1606–1620, 1997.

[60] Richard Szeliski and Heung-Yeung Shum. Motion estimation with


quadtree splines. IEEE Transactions on pattern analysis and machine
intelligence, 18(12):1199–1210, 1996.

[61] Olivier Musse, Fabrice Heitz, and Jean-Paul Armspach. Topology preserv-
ing deformable image matching using constrained hierarchical parametric
models. IEEE Transactions on Image Processing, 10(7):1081–1093, 2001.

[62] Hiroyuki Yoshida. Removal of normal anatomic structures in radiographs


using wavelet-based nonlinear variational method for image matching. In

56
SPIE’s International Symposium on Optical Science, Engineering, and In-
strumentation, pages 174–181. International Society for Optics and Pho-
tonics, 1998.

[63] Richard D Rabbitt, Jeffrey A Weiss, Gary E Christensen, and Michael I


Miller. Mapping of hyperelastic deformable templates using the finite el-
ement method. In PROCEEDINGS-SPIE THE INTERNATIONAL SO-
CIETY FOR OPTICAL ENGINEERING, pages 252–252. SPIE INTER-
NATIONAL SOCIETY FOR OPTICAL, 1995.

[64] Xavier Pennec, Radu Stefanescu, Vincent Arsigny, Pierre Fillard, and
Nicholas Ayache. Riemannian elasticity: A statistical regularization
framework for non-linear registration. Medical Image Computing and
Computer-Assisted Intervention–MICCAI 2005, pages 943–950, 2005.

[65] Igor Yanovsky, Carole Le Guyader, Alex Leow, Arthur Toga, Paul Thomp-
son, and Luminita Vese. Unbiased volumetric registration via nonlinear
elastic regularization. In 2nd MICCAI workshop on mathematical foun-
dations of computational anatomy, 2008.

[66] Carole Le Guyader and Luminita A Vese. A combined segmentation and


registration framework with a nonlinear elasticity smoother. Computer
Vision and Image Understanding, 115(12):1689–1709, 2011.

[67] Martin Burger, Jan Modersitzki, and Lars Ruthotto. A hyperelastic reg-
ularization energy for image registration. SIAM Journal on Scientific
Computing, 35(1):B132–B148, 2013.

[68] Marc Droske and Martin Rumpf. A variational approach to nonrigid


morphological image registration. SIAM Journal on Applied Mathematics,
64(2):668–687, 2004.

[69] Gary E Christensen, Richard D Rabbitt, and Michael I Miller. Deformable


templates using large deformation kinematics. IEEE transactions on im-
age processing, 5(10):1435–1447, 1996.

57
[70] Gary E Christensen, Sarang C Joshi, and Michael I Miller. Volumetric
transformation of brain anatomy. IEEE transactions on medical imaging,
16(6):864–877, 1997.

[71] Morten Bro-Nielsen and Claus Gramkow. Fast fluid registration of med-
ical images. In Visualization in Biomedical Computing, pages 265–276.
Springer, 1996.

[72] Nathan D Cahill, J Alison Noble, and David J Hawkes. Fourier meth-
ods for nonparametric image registration. In Computer Vision and Pat-
tern Recognition, 2007. CVPR’07. IEEE Conference on, pages 1–8. IEEE,
2007.

[73] WR Crum, C Tanner, and DJ Hawkes. Anisotropic multi-scale fluid reg-


istration: evaluation in magnetic resonance breast imaging. Physics in
medicine and biology, 50(21):5153, 2005.

[74] Yongmei Wang and Lawrence H Staib. Physical model-based non-rigid


registration incorporating statistical shape information. Medical image
analysis, 4(1):7–20, 2000.

[75] Ming-Chang Chiang, Alex D Leow, Andrea D Klunder, Rebecca A Dut-


ton, Marina Barysheva, Stephen E Rose, Katie L McMahon, Greig I
De Zubicaray, Arthur W Toga, and Paul M Thompson. Fluid registration
of diffusion tensor images using information theory. IEEE transactions on
medical imaging, 27(4):442–456, 2008.

[76] Berthold KP Horn and Brian G Schunck. Determining optical flow. Ar-
tificial intelligence, 17(1-3):185–203, 1981.

[77] JK Aggarwal and N Nandhakumar. On the computation of motion from


sequences of images-a review. Proceedings of the IEEE, 76(8):917–935,
1988.

58
[78] John L Barron, David J Fleet, and Steven S Beauchemin. Performance
of optical flow techniques. International journal of computer vision,
12(1):43–77, 1994.

[79] J-P Thirion. Image matching as a diffusion process: an analogy with


maxwell’s demons. Medical image analysis, 2(3):243–260, 1998.

[80] Erik Meijering. A chronology of interpolation: From ancient astronomy to


modern signal and image processing. Proceedings of the IEEE, 90(3):319–
342, 2002.

[81] Michael James David Powell. Approximation theory and methods. Cam-
bridge university press, 1981.

[82] M. J. D. Powell. Algorithms for approximation. chapter Radial Basis Func-


tions for Multivariable Interpolation: A Review, pages 143–167. Claren-
don Press, New York, NY, USA, 1987.

[83] Martin D Buhmann. Radial basis functions: theory and implementations,


volume 12. Cambridge university press, 2003.

[84] Detlef Ruprecht and Heinrich Müller. Free form deformation with scat-
tered data interpolation methods. In Geometric modelling, pages 267–281.
Springer, 1993.

[85] IP Schagen. The use of stochastic processes in interpolation and approx-


imation. International Journal of Computer Mathematics, 8(1):63–76,
1980.

[86] Richard Franke. Scattered data interpolation: tests of some methods.


Mathematics of computation, 38(157):181–200, 1982.

[87] Karl Rohr, Mike Fornefett, and H Siegfried Stiehl. Spline-based elastic im-
age registration: integration of landmark errors and orientation attributes.
Computer Vision and Image Understanding, 90(2):153–168, 2003.

59
[88] Karl Rohr, H Siegfried Stiehl, Rainer Sprengel, Thorsten M Buzug, Jürgen
Weese, and MH Kuhn. Landmark-based elastic registration using ap-
proximating thin-plate splines. IEEE Transactions on medical imaging,
20(6):526–534, 2001.

[89] Mike Fornefett, Karl Rohr, and H Siegfried Stiehl. Radial basis functions
with compact support for elastic registration of medical images. Image
and vision computing, 19(1):87–96, 2001.

[90] Ardeshir Goshtasby. Registration of images with geometric distortions.


IEEE Transactions on Geoscience and Remote Sensing, 26(1):60–64, 1988.

[91] Fred L. Bookstein. Principal warps: Thin-plate splines and the decomposi-
tion of deformations. IEEE Transactions on pattern analysis and machine
intelligence, 11(6):567–585, 1989.

[92] Fred L Bookstein. Thin-plate splines and the atlas problem for biomedical
images. In Biennial International Conference on Information Processing
in Medical Imaging, pages 326–342. Springer, 1991.

[93] William Eric Leifur Grimson. From images to surfaces: A computational


study of the human early visual system.

[94] Demetri Terzopoulos. Multiresolution computation of visible-surface rep-


resentations. PhD thesis, Massachusetts Institute of Technology, 1984.

[95] Lyubomir Zagorchev and Ardeshir Goshtasby. A comparative study of


transformation functions for nonrigid image registration. IEEE transac-
tions on image processing, 15(3):529–538, 2006.

[96] Stephen Marsland, Carole J Twining, and Chris J Taylor. Groupwise non-
rigid registration using polyharmonic clamped-plate splines. In Interna-
tional Conference on Medical Image Computing and Computer-Assisted
Intervention, pages 771–779. Springer, 2003.

60
[97] Vincent Camion and Laurent Younes. Geodesic interpolating splines. In
Energy Minimization Methods in Computer Vision and Pattern Recogni-
tion, pages 513–527. Springer, 2001.

[98] Laurent Younes. Combining geodesic interpolating splines and affine


transformations. IEEE transactions on image processing, 15(5):1111–
1119, 2006.

[99] Nur Arad, Nira Dyn, Daniel Reisfeld, and Yehezkel Yeshurun. Im-
age warping by radial basis functions: Application to facial expressions.
CVGIP: Graphical models and image processing, 56(2):161–172, 1994.

[100] Xuan Yang, Zhong Xue, Xia Liu, and Darong Xiong. Topology preserva-
tion evaluation of compact-support radial basis functions for image regis-
tration. Pattern Recognition Letters, 32(8):1162–1177, 2011.

[101] Karl Rohr and Stefan Wörz. An extension of thin-plate splines for image
registration with radial basis functions. In Biomedical Imaging (ISBI),
2012 9th IEEE International Symposium on, pages 442–445. IEEE, 2012.

[102] Malcolm H Davis, Alireza Khotanzad, Duane P Flamig, and Steven E


Harms. A physics-based coordinate transformation for 3-d image match-
ing. IEEE transactions on medical imaging, 16(3):317–328, 1997.

[103] Pei Chi Chou and Nicholas J Pagano. Elasticity: tensor, dyadic, and
engineering approaches. Courier Corporation, 1992.

[104] Jan Kohlrausch, Karl Rohr, and H Siegfried Stiehl. A new class of elas-
tic body splines for nonrigid registration of medical images. Journal of
Mathematical Imaging and Vision, 23(3):253–280, 2005.

[105] Stefan Wörz and Karl Rohr. Physics-based elastic registration using non-
radial basis functions and including landmark localization uncertainties.
Computer Vision and Image Understanding, 111(3):263–274, 2008.

61
[106] Alan H Barr. Global and local deformations of solid primitives. In ACM
Siggraph Computer Graphics, volume 18, pages 21–30. ACM, 1984.

[107] Thomas W Sederberg and Scott R Parry. Free-form deformation of solid


geometric models. ACM SIGGRAPH computer graphics, 20(4):151–160,
1986.

[108] Sabine Coquillart. Extended free-form deformation: a sculpturing tool for


3D geometric modeling, volume 24. ACM, 1990.

[109] Yu-Kuang Chang and Alyn P Rockwood. A generalized de casteljau ap-


proach to 3d free-form deformation. In Proceedings of the 21st annual
conference on Computer graphics and interactive techniques, pages 257–
260. ACM, 1994.

[110] Julia A Schnabel, Daniel Rueckert, Marcel Quist, Jane M Blackall,


Andy D Castellano-Smith, Thomas Hartkens, Graeme P Penney, Wal-
ter A Hall, Haiying Liu, Charles L Truwit, et al. A generic framework for
non-rigid registration based on non-uniform multi-level free-form defor-
mations. In International Conference on Medical Image Computing and
Computer-Assisted Intervention, pages 573–581. Springer, 2001.

[111] Wenzhe Shi, Xiahai Zhuang, Luis Pizarro, Wenjia Bai, Haiyan Wang,
Kai-Pin Tung, Philip Edwards, and Daniel Rueckert. Registration using
sparse free-form deformations. In International Conference on Medical
Image Computing and Computer-Assisted Intervention, pages 659–666.
Springer, 2012.

[112] Jianzhe Wang and Tianzi Jiang. Nonrigid registration of brain mri using
nurbs. Pattern Recognition Letters, 28(2):214–223, 2007.

[113] Vincent Noblet, Christian Heinrich, Fabrice Heitz, and Jean-Paul


Armspach. Symmetric nonrigid image registration: application to average
brain templates construction. Medical Image Computing and Computer-
Assisted Intervention–MICCAI 2008, pages 897–904, 2008.

62
[114] Wei Feng, Stanley J Reeves, Thomas S Denney, Steven Lloyd, Louis
Dell’Italia, and Himanshu Gupta. A new consistent image registration
formulation with a b-spline deformation model. In Biomedical Imaging:
From Nano to Macro, 2009. ISBI’09. IEEE International Symposium on,
pages 979–982. IEEE, 2009.

[115] Kanwal K Bhatia, Joseph V Hajnal, Basant K Puri, A David Edwards,


and Daniel Rueckert. Consistent groupwise non-rigid registration for at-
las construction. In Biomedical Imaging: Nano to Macro, 2004. IEEE
International Symposium on, pages 908–911. IEEE, 2004.

[116] Serdar K Balci, Polina Golland, Martha Shenton, and William M Wells.
Free-form b-spline deformation model for groupwise registration. In Med-
ical image computing and computer-assisted intervention: MICCAI...
International Conference on Medical Image Computing and Computer-
Assisted Intervention, volume 10, pages 23–30. NIH Public Access, 2006.

[117] Dimitrios Perperidis, Raad H Mohiaddin, and Daniel Rueckert. Spatio-


temporal free-form registration of cardiac mr image sequences. Medical
image analysis, 9(5):441–456, 2005.

[118] Jef Vandemeulebroucke, Simon Rit, Jan Kybic, Patrick Clarysse, and
David Sarrut. Spatiotemporal motion estimation for respiratory-
correlated imaging of the lungs. Medical physics, 38(1):166–178, 2011.

[119] D. Louis Collins and Alan C. Evans. Animal: validation and applica-
tions of nonlinear registration-based segmentation. International journal
of pattern recognition and artificial intelligence, 11(08):1271–1294, 1997.

[120] Pierre Hellier, Christian Barillot, Étienne Mémin, and Patrick Pérez. Hier-
archical estimation of a dense deformation field for 3-d robust registration.
IEEE transactions on medical imaging, 20(5):388–402, 2001.

[121] Hui Zhang, Paul A Yushkevich, Daniel C Alexander, and James C Gee.

63
Deformable registration of diffusion tensor mr images with explicit orien-
tation optimization. Medical image analysis, 10(5):764–785, 2006.

[122] Christian Buerger, Tobias Schaeffter, and Andrew P King. Hierarchical


adaptive local affine registration for fast and robust respiratory motion
estimation. Medical image analysis, 15(4):551–564, 2011.

[123] Vincent Arsigny, Xavier Pennec, and Nicholas Ayache. Polyrigid and
polyaffine transformations: a novel geometrical tool to deal with non-rigid
deformations–application to the registration of histological slices. Medical
image analysis, 9(6):507–523, 2005.

[124] Vincent Arsigny, Olivier Commowick, Nicholas Ayache, and Xavier Pen-
nec. A fast and log-euclidean polyaffine framework for locally linear reg-
istration. Journal of Mathematical Imaging and Vision, 33(2):222–238,
2009.

[125] Gary E Christensen and Hans J Johnson. Consistent image registration.


IEEE transactions on medical imaging, 20(7):568–582, 2001.

[126] Hans J Johnson and Gary E Christensen. Consistent landmark and


intensity-based image registration. IEEE transactions on medical imaging,
21(5):450–461, 2002.

[127] Vincent Noblet, Christian Heinrich, Fabrice Heitz, and J-P Armspach. 3-d
deformable image registration: a topology preservation scheme based on
hierarchical deformation models and interval analysis optimization. IEEE
Transactions on image processing, 14(5):553–566, 2005.

[128] Daniel Rueckert, Paul Aljabar, Rolf A Heckemann, Joseph V Hajnal, and
Alexander Hammers. Diffeomorphic registration using b-splines. In Inter-
national Conference on Medical Image Computing and Computer-Assisted
Intervention, pages 702–709. Springer, 2006.

64
[129] Eldad Haber and Jan Modersitzki. Image registration with guaran-
teed displacement regularity. International Journal of Computer Vision,
71(3):361–372, 2007.

[130] Alain Trouvé. Diffeomorphisms groups and pattern matching in image


analysis. International journal of computer vision, 28(3):213–221, 1998.

[131] M Faisal Beg, Michael I Miller, Alain Trouvé, and Laurent Younes. Com-
puting large deformation metric mappings via geodesic flows of diffeomor-
phisms. International journal of computer vision, 61(2):139–157, 2005.

[132] Michael I Miller, Alain Trouvé, and Laurent Younes. On the metrics
and euler-lagrange equations of computational anatomy. Annual review
of biomedical engineering, 4(1):375–405, 2002.

[133] Laurent Younes. Jacobi fields in groups of diffeomorphisms and applica-


tions. Quarterly of applied mathematics, pages 113–134, 2007.

[134] Mirza Faisal Beg and Ali Khan. Symmetric data attachment terms for
large deformation image registration. IEEE transactions on medical imag-
ing, 26(9):1179–1189, 2007.

[135] Brian B Avants, Charles L Epstein, Murray Grossman, and James C Gee.
Symmetric diffeomorphic image registration with cross-correlation: evalu-
ating automated labeling of elderly and neurodegenerative brain. Medical
image analysis, 12(1):26–41, 2008.

[136] John Ashburner and Karl J Friston. Diffeomorphic registration us-


ing geodesic shooting and gauss–newton optimisation. NeuroImage,
55(3):954–967, 2011.

[137] Monica Hernandez, Matias N Bossa, and Salvador Olmos. Registration


of anatomical images using paths of diffeomorphisms parameterized with
stationary vector field flows. International Journal of Computer Vision,
85(3):291–306, 2009.

65
[138] Yan Cao, Michael I Miller, Raimond L Winslow, and Laurent Younes.
Large deformation diffeomorphic metric mapping of vector fields. IEEE
transactions on medical imaging, 24(9):1216–1230, 2005.

[139] Brian B Avants, P Thomas Schoenemann, and James C Gee. Lagrangian


frame diffeomorphic image registration: Morphometric comparison of hu-
man and chimpanzee cortex. Medical image analysis, 10(3):397–412, 2006.

[140] Guillaume Auzias, Olivier Colliot, Joan Alexis Glaunes, Matthieu Perrot,
Jean-François Mangin, Alain Trouve, and Sylvain Baillet. Diffeomorphic
brain registration under exhaustive sulcal constraints. IEEE transactions
on medical imaging, 30(6):1214–1227, 2011.

[141] Pei Zhang, Marc Niethammer, Dinggang Shen, and Pew-Thian Yap. Large
deformation diffeomorphic registration of diffusion-weighted images. In
International Conference on Medical Image Computing and Computer-
Assisted Intervention, pages 171–178. Springer, 2012.

[142] Joseph V Hajnal, Nadeem Saeed, Elaine J Soar, Angela Oatridge, Ian R
Young, and Graeme M Bydder. A registration and interpolation procedure
for subvoxel matching of serially acquired mr images. Journal of computer
assisted tomography, 19(2):289–296, 1995.

[143] Karl J Friston, Steven Williams, Robert Howard, Richard SJ Frackowiak,


and Robert Turner. Movement-related effects in fmri time-series. Magnetic
resonance in medicine, 35(3):346–355, 1996.

[144] Karl Friston, John Ashburner, Christopher D Frith, J-B Poline, John D
Heather, Richard SJ Frackowiak, et al. Spatial registration and normal-
ization of images. Human brain mapping, 3(3):165–189, 1995.

[145] John Ashburner, Karl J Friston, et al. Nonlinear spatial normalization


using basis functions. Human brain mapping, 7(4):254–266, 1999.

[146] Derek LG Hill and Philipe Batchelor. Registration methodology: concepts


and algorithms. Medical Image Registration, pages 39–70, 2001.

66
[147] Claudio Castellini and Georg Passig. Ultrasound image features of the
wrist are linearly related to finger positions. In Intelligent Robots and Sys-
tems (IROS), 2011 IEEE/RSJ International Conference on, pages 2108–
2114. IEEE, 2011.

[148] Mark C Lee, Matthew Solomito, and Archit Patel. Supine magnetic
resonance imaging cobb measurements for idiopathic scoliosis are lin-
early related to measurements from standing plain radiographs. Spine,
38(11):E656–E661, 2013.

[149] Richard S Likes. Moving gradient zeugmatography, December 22 1981.


US Patent 4,307,343.

[150] Donald B Twieg. The k-trajectory formulation of the nmr imaging process
with applications in analysis and synthesis of imaging methods. Medical
physics, 10(5):610–621, 1983.

[151] Riadh Abdelfattah and Jean-Marie Nicolas. Sub-pixelic image registration


for sar interferometry coherence optimization. In practice, 9(10):17, 2004.

[152] Ariff Kassam and Michael L Wood. Fourier registration of three-


dimensional brain mr images: Exploiting the axis of rotation. Journal
of Magnetic Resonance Imaging, 6(6):894–902, 1996.

[153] Roger P Woods, Simon R Cherry, and John C Mazziotta. Rapid auto-
mated algorithm for aligning and reslicing pet images. Journal of computer
assisted tomography, 16(4):620–633, 1992.

[154] Roger P Woods, Scott T Grafton, Colin J Holmes, Simon R Cherry, and
John C Mazziotta. Automated image registration: I. general methods
and intrasubject, intramodality validation. Journal of computer assisted
tomography, 22(1):139–152, 1998.

[155] Roger P Woods, John C Mazziotta, Simon R Cherry, et al. Mri-pet reg-
istration with automated algorithm. Journal of computer assisted tomog-
raphy, 17(4):536–546, 1993.

67
[156] Claude E Shannon. A mathematical theory of communication. ACM
SIGMOBILE Mobile Computing and Communications Review, 5(1):3–55,
2001.

[157] Claude Elwood Shannon. Communication in the presence of noise. Pro-


ceedings of the IRE, 37(1):10–21, 1949.

[158] Allan H Friedman. Advanced techniques in image-guided brain and spine


surgery. Neuro-oncology, 6(1):77, 2004.

[159] Derek LG Hill, Colin Studholme, and David J Hawkes. Voxel similarity
measures for automated image registration. In Visualization in biomedical
computing, volume 2359, pages 205–216, 1994.

[160] János Aczél and Zoltán Daróczy. A mixed theory of information. i:


Symmetric, recursive and measurable entropies of randomized systems
of events. RAIRO. Informatique théorique, 12(2):149–155, 1978.

[161] Friedrich Liese and Igor Vajda. On divergences and informations in statis-
tics and information theory. IEEE Transactions on Information Theory,
52(10):4394–4412, 2006.

[162] Josien PW Pluim, JB Antoine Maintz, and Max A Viergever. Mutual in-
formation matching and interpolation artefacts. In SPIE Medical Imaging,
volume 3661, pages 56–65. Citeseer, 1999.

[163] Geoffrey Egnal and Kostas Daniilidis. Image registration using mutual
information. 2000.

[164] Colin Studholme, Derek LG Hill, and David J Hawkes. An overlap invari-
ant entropy measure of 3d medical image alignment. Pattern recognition,
32(1):71–86, 1999.

[165] Jasbir Arora. Introduction to optimum design. Academic Press, 2004.

68
[166] Xin-She Yang and Slawomir Koziel. Computational optimization, mod-
elling and simulation–a paradigm shift. Procedia Computer Science,
1(1):1297–1300, 2010.

[167] Ingrid Sluimer, Mathias Prokop, and Bram Van Ginneken. Toward auto-
mated segmentation of the pathological lung in ct. IEEE transactions on
medical imaging, 24(8):1025–1038, 2005.

[168] Xavier Pennec, Pascal Cachier, and Nicholas Ayache. Tracking brain de-
formations in time sequences of 3d us images. Pattern Recognition Letters,
24(4):801–813, 2003.

[169] Stefan Klein. Optimisation methods for medical image registration.


Utrecht University, 2008.

[170] Jan Kybic and Michael Unser. Fast parametric elastic image registration.
IEEE transactions on image processing, 12(11):1427–1442, 2003.

[171] Philippe Thevenaz, Urs E Ruttimann, and Michael Unser. A pyramid


approach to subpixel registration based on intensity. IEEE transactions
on image processing, 7(1):27–41, 1998.

[172] Colin Studholme, Derek LG Hill, and David J Hawkes. Automated 3-d
registration of mr and ct images of the head. Medical image analysis,
1(2):163–175, 1996.

[173] Philippe Thévenaz and Michael Unser. Optimization of mutual informa-


tion for multiresolution image registration. IEEE transactions on image
processing, 9(12):2083–2099, 2000.

[174] Bernd Fischer and Jan Modersitzki. A unified approach to fast image
registration and a new curvature based registration technique. Linear
Algebra and its applications, 380:107–124, 2004.

[175] Marius Staring, Stefan Klein, and Josien PW Pluim. A rigidity penalty
term for nonrigid registration. Medical physics, 34(11):4098–4108, 2007.

69
[176] Daniel Rueckert, Luke I Sonoda, Carmel Hayes, Derek LG Hill, Martin O
Leach, and David J Hawkes. Nonrigid registration using free-form defor-
mations: application to breast mr images. IEEE transactions on medical
imaging, 18(8):712–721, 1999.

[177] Steven C Chapra and Raymond P Canale. Numerical methods for engi-
neers, volume 2. McGraw-Hill New York, 1988.

[178] Renu Maria Mathews, D Raveena Judie Dolly, and Ann Therese Francy.
Comparison of optimization techniques for mutual information based real
time image registration. International Journal of Advanced Research in
Computer Engineering & Technology (IJARCET), 2(3):PP–1280, 2013.

[179] Stephen J Wright and Jorge Nocedal. Numerical optimization. Springer


Science, 35(67-68):7, 1999.

[180] Philip E Gill, Walter Murray, Michael A Saunders, and Margaret H


Wright. Recent developments in constrained optimization. Journal of
Computational and Applied Mathematics, 22(2):257–270, 1988.

[181] William H Press, Saul A Teukolsky, William T Vetterling, and Brian P


Flannery. Numerical recipes in C, volume 2. Cambridge university press
Cambridge, 1996.

[182] Jorge Nocedal. Updating quasi-newton matrices with limited storage.


Mathematics of computation, 35(151):773–782, 1980.

[183] Stefan Klein, Marius Staring, and Josien PW Pluim. Evaluation of op-
timization methods for nonrigid medical image registration using mu-
tual information and b-splines. IEEE transactions on image processing,
16(12):2879–2890, 2007.

[184] Magnus Rudolph Hestenes and Eduard Stiefel. Methods of conjugate gra-
dients for solving linear systems, volume 49. NBS, 1952.

70
[185] YH Dai and Yaxiang Yuan. An efficient hybrid conjugate gradient
method for unconstrained optimization. Annals of Operations Research,
103(1):33–47, 2001.

[186] Jack Kiefer, Jacob Wolfowitz, et al. Stochastic estimation of the maximum
of a regression function. The Annals of Mathematical Statistics, 23(3):462–
466, 1952.

[187] James C Spall. Multivariate stochastic approximation using a simulta-


neous perturbation gradient approximation. IEEE transactions on auto-
matic control, 37(3):332–341, 1992.

[188] James C Spall. Developments in stochastic optimization algorithms with


gradient approximations based on function measurements. In Proceedings
of the 26th conference on Winter simulation, pages 207–214. Society for
Computer Simulation International, 1994.

[189] JH Venter. An extension of the robbins-monro procedure. The Annals of


Mathematical Statistics, pages 181–190, 1967.

[190] James C Spall. Adaptive stochastic approximation by the simulta-


neous perturbation method. IEEE transactions on automatic control,
45(10):1839–1853, 2000.

[191] Ingo Rechenberg. Evolutionsstrategien. In Simulationsmethoden in der


Medizin und Biologie, pages 83–114. Springer, 1978.

[192] Hans-Paul Schwefel and Günter Rudolph. Contemporary evolution strate-


gies. In European conference on artificial life, pages 891–907. Springer,
1995.

[193] Scott Kirkpatrick, C Daniel Gelatt, Mario P Vecchi, et al. Optimization


by simulated annealing. science, 220(4598):671–680, 1983.

[194] Vladimı́r Černỳ. Thermodynamical approach to the traveling salesman


problem: An efficient simulation algorithm. Journal of optimization theory
and applications, 45(1):41–51, 1985.

71
[195] Chi Zhou, Hai-bing GAO, Liang Gao, and Wan-guo ZHANG. Particle
swarm optimization (pso) algorithm [j]. Application Research of Comput-
ers, 12:7–11, 2003.

[196] Chen-Lun Lin, Aya Mimori, and Yen-Wei Chen. Hybrid particle swarm
optimization and its application to multimodal 3d medical image regis-
tration. Computational intelligence and neuroscience, 2012:6, 2012.

[197] Qi Li and Hongbing Ji. Medical image registration based on maximization


of mutual information and particle swarm optimization. In Proc. of SPIE
Vol, volume 6534, page 65342M1, 2007.

[198] Andriy Myronenko and Xubo Song. Intensity-based image registration by


minimizing residual complexity. IEEE Transactions on Medical Imaging,
29(11):1882–1891, 2010.

[199] Yun He, A Ben Hamza, and Hamid Krim. A generalized divergence mea-
sure for robust image registration. IEEE Transactions on Signal Process-
ing, 51(5):1211–1220, 2003.

[200] Mark P Wachowiak, Renata Smolı́ková, and Terry M Peters. Multireso-


lution biomedical image registration using generalized information mea-
sures. In International Conference on Medical Image Computing and
Computer-Assisted Intervention, pages 846–853. Springer, 2003.

[201] Constantino Tsallis. Possible generalization of boltzmann-gibbs statistics.


Journal of statistical physics, 52(1):479–487, 1988.

[202] Mohammed Khader and A Ben Hamza. An entropy-based technique for


nonrigid medical image alignment. Combinatorial Image Analysis, pages
444–455, 2011.

[203] Frederik Maes, Dirk Vandermeulen, and Paul Suetens. Comparative eval-
uation of multiresolution optimization strategies for multimodality image
registration by maximization of mutual information. Medical image anal-
ysis, 3(4):373–386, 1999.

72
[204] Josien PW Pluim, JB Antoine Maintz, and Max A Viergever. Mutual in-
formation matching in multiresolution contexts. Image and Vision Com-
puting, 19(1):45–52, 2001.

[205] Mark P Wachowiak, Renata Smolı́ková, Yufeng Zheng, Jacek M Zurada,


and Adel Said Elmaghraby. An approach to multimodal biomedical image
registration utilizing particle swarm optimization. IEEE Transactions on
evolutionary computation, 8(3):289–301, 2004.

[206] Wang Anna, Wang Tingjun, Zhang Jinjin, and Xue Silin. A novel method
of medical image registration based on dtcwt and npso. In Natural Com-
putation, 2009. ICNC’09. Fifth International Conference on, volume 5,
pages 23–27. IEEE, 2009.

[207] Hassiba Talbi and Mohamed Batouche. Hybrid particle swarm with dif-
ferential evolution for multimodal image registration. In Industrial Tech-
nology, 2004. IEEE ICIT’04. 2004 IEEE International Conference on,
volume 3, pages 1567–1572. IEEE, 2004.

[208] Di Zhou, Jun Sun, Choi-Hong Lai, Wenbo Xu, and Xiaoguang Lee. An
improved quantum-behaved particle swarm optimization and its applica-
tion to medical image registration. International Journal of Computer
Mathematics, 88(6):1208–1223, 2011.

[209] Jun Sun, Bin Feng, and Wenbo Xu. Particle swarm optimization with
particles having quantum behavior. In Evolutionary Computation, 2004.
CEC2004. Congress on, volume 1, pages 325–331. IEEE, 2004.

[210] Danyang Zhao, Benqiang Yang, Libo Zhang, Xiangrong Zhou, Yenwei
Chen, Hiroshi Fujita, and Huiyan Jiang. Research of medical image regis-
tration based on rmi-sapso algorithm. In Proceedings of the 2nd Interna-
tional Conference on Green Communications and Networks 2012 (GCN
2012): Volume 2, pages 697–704. Springer, 2013.

73
[211] Prabhat Kumar Sharma, VS Bhavya, KM Navyashree, KS Sunil, and
P Pavithra. Artificial bee colony and its application for image fusion.
International Journal of Information Technology and Computer Science
(IJITCS), 4(11):42, 2012.

[212] Xin-She Yang. Bat algorithm for multi-objective optimisation. Interna-


tional Journal of Bio-Inspired Computation, 3(5):267–274, 2011.

[213] Iztok Fister Jr, Dušan Fister, and Xin-She Yang. A hybrid bat algorithm.
arXiv preprint arXiv:1303.6310, 2013.

74
Table of contents

1. Introduction ............................................2
2. Geometric Transformation .....................7
3. Similarity Measure .................................22
4. Optimization Procedure ........................29
5. Discussion ..............................................40
6. Conclusion .............................................49

You might also like