the manual

Stefan Klein and Marius Staring

November 12, 2010
Contents
1 Introduction 1
1.1 Outline of this manual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Quick start . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Image registration 3
2.1 Registration framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4 Image samplers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.5 Interpolators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.6 Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.7 Optimisers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.8 Multi-resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.8.1 Data complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.8.2 Transformation complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.9 Evaluating registration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 elastix 15
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.1.1 Key features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Getting started . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2.1 Getting started the really easy way . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2.2 Getting started the easy way . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 How to call elastix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4 The parameter le . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4 transformix 20
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.2 How to call transformix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3 The transform parameter le . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.4 Some details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.4.1 Run-time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.4.2 Memory consumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5 Tutorial 23
5.1 Selecting the registration components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.2 Overview of all parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.3 Important parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
i
5.3.1 Registration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.3.2 Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.3.3 Sampler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.3.4 Interpolator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.3.5 Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.3.6 Optimiser . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.3.7 Image pyramids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
5.4 Masks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.5 Trouble shooting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.5.1 Common errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.5.2 Bad initial alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.5.3 Memory consumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
6 Advanced topics 32
6.1 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
6.1.1 Image registration with multiple metrics and/or images . . . . . . . . . . . . . . . . . 32
6.1.2 -mutual information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.1.3 Penalty terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.1.4 Bending energy penalty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6.1.5 Rigidity penalty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
6.1.6 DisplacementMagnitudePenalty: inverting transformations . . . . . . . . . . . . . . . . 35
6.1.7 Corresponding points: help the registration . . . . . . . . . . . . . . . . . . . . . . . . 35
6.1.8 VarianceOverLastDimensionMetric: aligning time series . . . . . . . . . . . . . . . . . 36
6.2 Image samplers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6.3 Interpolators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6.4 Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6.5 Optimisation methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
7 Developers guide 38
7.1 Relation to ITK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.2 Overview of the elastix code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
7.2.1 Directory structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
7.3 Creating new components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
A Example parameter le 41
B Example transform parameter le 43
C Software License 44
Bibliography 46
ii
Chapter 1
Introduction
This manual describes a software package for image registration: elastix. The software consists of a
collection of algorithms that are commonly used to solve medical image registration problems. A large part
of the code is based on the Insight Toolkit (ITK). The modular design of elastix allows the user to quickly
test and compare dierent registration methods for his/her specic application. The command-line interface
simplies the processing of large amounts of data sets, using scripting.
1.1 Outline of this manual
In Chapter 2 quite an extensive introduction to some general theory of image registration is given. Also, the
dierent components of which a registration method consists, are treated. In Chapter 3, elastix is described
and its usage is explained. Chapter 4 is dedicated to transformix, a program accompanying elastix. A
tutorial is given in Chapter 5, including many recommendations based on the authors experiences. More
advanced registration topics are covered in Chapter 6. The nal chapter provides more details for those
interested in the setup of the source code and gives information on how to implement your own additions to
elastix. In the Appendices A and B example (transform) parameter les are given. Appendix C contains
the software license and disclaimer under which elastix is currently distributed.
1.2 Quick start
Download elastix from http://elastix.isi.uu.nl/download.php. See Section 3.2 for details about
the installation. Do not forget to subscribe to the elastix mailing list, which is the main forum for
questions and announcements.
Read some basics about the program at http://elastix.isi.uu.nl/about.php and in this manual.
Try the example of usage. It can be found in the About section of the website. If you dont get the
example running at your computer take a look at the FAQ in the general section:
http://elastix.isi.uu.nl/FAQ.php
Read the complete manual if you are the type of person that rst wants to know.
Get started with your own application. If you need more information at this point you can now start
reading the manual. You can nd more information on tuning the parameters in Chapter 5. A list of all
available parameters can be found at http://elastix.isi.uu.nl/doxygen/pages.html. Also take
a look at the parameter le database at http://elastix.isi.uu.nl/wiki.php, for many example
parameter les.
1
When you are stuck, dont miss the tutorial in Chapter 5 of this manual. Also, take a look at the FAQ
again for some common problems.
When you are still stuck, do not hesitate to send an e-mail to the elastix mailing list. In general,
you will soon get an answer.
1.3 Acknowledgements
This manual has mainly been written while the authors worked at the Image Sciences Institute (ISI, http://
www.isi.uu.nl), Utrecht, The Netherlands. We thank the users of elastix, whose questions and remarks
helped improving the usability and documentation of elastix. Specically, we want to thank the following
people for proofreading (parts of) this manual when we constructed a rst version: Josien Pluim, Keelin
Murphy, Martijn van der Bom, Sascha M unzing, Jeroen de Bresser, Bram van Ginneken, Kajo van der Marel,
Adrienne Mendrik (in no specic order).
2
Chapter 2
Image registration
This chapter introduces primary registration concepts that are at the base of elastix. More advanced
registration topics are covered in Chapter 6.
Image registration is an important tool in the eld of medical imaging. In many clinical situations several
images of a patient are made in order to analyse the patients situation. These images are acquired with,
for example, X-ray scanners, Magnetic Resonance Imaging (MRI) scanners, Computed Tomography (CT)
scanners, and Ultrasound scanners, which provide knowledge about the anatomy of the subject. Combination
of patient data, mono- or multi-modal, often yields additional clinical information not apparent in the
separate images. For this purpose, the spatial relation between the images has to be found. Image registration
is the task of nding a spatial one-to-one mapping from voxels in one image to voxels in the other image,
see Figure 2.1. Good reviews on the subject are given in Maintz and Viergever [1998], Lester and Arridge
[1999], Hill et al. [2001], Hajnal et al. [2001], Zitova and Flusser [2003], Modersitzki [2004].
The following section introduces the mathematical formulation of the registration process and gives an
overview of the components of which a general registration method consists. After that, in Sections 2.3-2.8,
each component is discussed in more detail. For each component, the name used by elastix is given, in
typewriter style. In Section 2.9, methods to evaluate the registration results are discussed.
2.1 Registration framework
Two images are involved in the registration process. One image, the moving image I
M
(x), is deformed to
t the other image, the xed image I
F
(x). Registration is the problem of nding a displacement u(x) that
makes I
M
(x +u(x)) spatially aligned to I
F
(x). An equivalent formulation is to say that registration is the
problem of nding a transformation T(x) = x+u(x) that makes I
M
(T(x)) spatially aligned to I
F
(x). The
quality of alignment is dened by a distance or similarity measure S, such as the sum of squared dierences
(SSD), the correlation ratio, or the mutual information (MI) measure. Because this problem is ill-posed for
nonrigid transformations T, a regularisation or penalty term P is often introduced that constrains T.
p q
T
Figure 2.1: Image registration is the task of nding a spatial transformation mapping one image to another.
Adopted from Iba nez et al. [2005].
3
pyramid
optimiser
fixed image
transform
sampler
metric
multiresolution
interpolator pyramid moving image
resolution level
Figure 2.2: The basic registration components.
Commonly, the registration problem is formulated as an optimisation problem in which the cost function
C is minimised w.r.t. T:

T = arg min
T
C(T; I
F
, I
M
), with (2.1)
C(T; I
F
, I
M
) = S(T; I
F
, I
M
) +P(T), (2.2)
where weighs similarity against regularity. To solve the above minimisation problem, there are basically
two approaches: parametric and nonparametric. The reader is referred to Fischer and Modersitzki [2004]
for an overview on nonparametric methods, which are not discussed in this manual. The elastix software
is based on the parametric approach. In parametric methods, the number of possible transformations is
limited by introducing a parametrisation (model) of the transformation. The original optimisation problem
thus becomes:

= arg min
T
C(T

; I
F
, I
M
), (2.3)
where the subscript indicates that the transform has been parameterised. The vector contains the
values of the transformation parameters. For example, when the transformation is modelled as a 2D rigid
transformation, the parameter vector contains one rotation angle and the translations in x and y direction.
We may write Equation (2.3) also as:
= arg min

C(; I
F
, I
M
). (2.4)
From this equation it becomes clear that the original problem (2.1) has been simplied. Instead of optimising
over a space of functions T, we now optimise over the elements of . Examples of other transformation
models are given in Section 2.6.
Figure 2.2 shows the general components of a parametric registration algorithm in a block scheme. The
scheme is a slightly extended version of the scheme introduced in Iba nez et al. [2005]. Several components
can be recognised from Equations (2.1)-(2.4); some will be introduced later. First of all, we have the images.
The concept of an image needs to be dened. This is done in Section 2.2. Then we have the cost function
C, or metric, which denes the quality of alignment. As mentioned earlier, the cost function consists of
a similarity measure S and a regularisation term P. The regularisation term P is not discussed in this
chapter, but in Chapter 6. The similarity measure S is discussed in Section 2.3. The denition of the
similarity measure introduces the sampler component, which is treated in Section 2.4. Some examples of
transformation models T

are given in Section 2.6. The optimisation procedure to actually solve the problem
4
0 100 50 150 200
0
50
100
150
200
250
300
30.0
20.0
Size=7x6
Spacing=( 20.0, 30.0 )
Physical extent=( 140.0, 180.0 )
Origin=(60.0,70.0)
Image Origin
Voronoi Region
Pixel Coverage
Delaunay Region
Linear Interpolation Region
Pixel Coordinates
Spacing[0]
S
p
a
c
i
n
g
[
1
]
Figure 2.3: Geometrical concepts associated with the ITK image. Adopted from Iba nez et al. [2005].
(2.4) is explained in Section 2.7. During the optimisation, the value I
M
(T

(x)) is evaluated at non-voxel

positions, for which intensity interpolation is needed. Choices for the interpolator are described in Section
2.5. Another thing, not immediately clear from Equations (2.1)-(2.4), is the use of multi-resolution strategies
to speed-up registration, and to make it more robust, see Section 2.8.
2.2 Images
Since image registration is all about images, we have to be careful with what is meant by an image. We
adopt the notion of an image from the Insight Toolkit [Iba nez et al., 2005, p. 40]:
Additional information about the images is considered mandatory. In particular the information
associated with the physical spacing between pixels and the position of the image in space with
respect to some world coordinate system are extremely important. Image origin and spacing are
fundamental to many applications. Registration, for example, is performed in physical coordi-
nates. Improperly dened spacing and origins will result in inconsistent results in such processes.
Medical images with no spatial information should not be used for medical diagnosis, image
analysis, feature extraction, assisted radiation therapy or image guided surgery. In other words,
medical images lacking spatial information are not only useless but also hazardous.
Figure 2.3 illustrates the main geometrical concepts associated with the itk::Image. In this gure,
circles are used to represent the centre of pixels. The value of the pixel is assumed to exist as
a Dirac Delta Function located at the pixel centre. Pixel spacing is measured between the pixel
centres and can be dierent along each dimension. The image origin is associated with the
coordinates of the rst pixel in the image. A pixel is considered to be the rectangular region
surrounding the pixel centre holding the data value. This can be viewed as the Voronoi region of
the image grid, as illustrated in the right side of the gure. Linear interpolation of image values
is performed inside the Delaunay region whose corners are pixel centres.
Therefore, you should take care that you use an image format that is able to store the relevant information
(e.g. mhd, DICOM). Some image formats, like vpx, do not store the origin and spacing. This may cause
serious problems!
Up to elastix version 4.2, the image orientation (direction cosines) was not yet fully supported in
elastix. From elastix 4.3, image orientation is fully supported, but can be disabled for backward com-
patibility reasons.
5
2.3 Metrics
Several choices for the similarity measure can be found in the literature. Some common choices are described
below. Between brackets the name of the metric in elastix is given:
Sum of Squared Dierences (SSD): (AdvancedMeanSquares) The SSD is dened as:
SSD(; I
F
, I
M
) =
1
|
F
|

xiF
(I
F
(x
i
) I
M
(T

(x
i
)))
2
, (2.5)
with
F
the domain of the xed image I
F
, and |
F
| the number of voxels. Given a transformation T,
this measure can easily be implemented by looping over the voxels in the xed image, taking I
F
(x
i
),
calculating I
M
(T

(x
i
)) by interpolation, and adding the squared dierence to the sum.
Normalised Correlation Coecient (NCC): (AdvancedNormalizedCorrelation) The NCC is dened
as:
NCC(; I
F
, I
M
) =

xiF

I
F
(x
i
) I
F

I
M
(T

(x
i
)) I
M

xiF

I
F
(x
i
) I
F

2

xiF

I
M
(T

(x
i
)) I
M

2
, (2.6)
with the average grey-values I
F
=
1
|F |

xiF
I
F
(x
i
) and I
M
=
1
|F |

xiF
I
M
(T

(x
i
)).
Mutual Information (MI): (AdvancedMattesMutualInformation) For MI [Maes et al., 1997, Viola and
Wells III, 1997, Mattes et al., 2003] we use a denition given by Thevenaz and Unser [2000]:
MI(; I
F
, I
M
) =

mLM

fLF
p(f, m; ) log
2

p(f, m; )
p
F
(f)p
M
(m; )

, (2.7)
where L
F
and L
M
are sets of regularly spaced intensity bin centres, p is the discrete joint probability,
and p
F
and p
M
are the marginal discrete probabilities of the xed and moving image, obtained by
summing p over m and f, respectively. The joint probabilities are estimated using B-spline Parzen
windows:
p(f, m; ) =
1
|
F
|

xiF
w
F
(f/
F
I
F
(x
i
)/
F
)
w
M
(m/
M
I
M
(T

(x
i
))/
M
),
(2.8)
where w
F
and w
M
represent the xed and moving B-spline Parzen windows. The scaling constants
F
and
M
must equal the intensity bin widths dened by L
F
and L
M
. These follow directly from the
grey-value ranges of I
F
and I
M
and the user-specied number of histogram bins |L
F
| and |L
M
|.
Normalized Mutual Information (NMI): (NormalizedMutualInformation)
NMI is dened by NMI = (H(I
F
) + H(I
M
))/H(I
F
, I
M
), with H denoting entropy. This expression
can be compared to the denition of MI in terms of H: MI = H(I
F
) + H(I
M
) H(I
F
, I
M
). Again,
with the joint probabilities dened by 2.8 (using B-spline Parzen windows), NMI can be written as:
NMI(; I
F
, I
M
) =

fLF
p
F
(f) log
2
p
F
(f) +

mLM
p
M
(m; ) log
2
p
M
(m; )

mLM

fLF
p(f, m; ) log
2
p(f, m; )
=

mLM

fLF
p(f, m; ) log
2
(p
F
(f)p
M
(m; ))

mLM

fLF
p(f, m; ) log
2
p(f, m; )
. (2.9)
6
Kappa Statistic (KS): (AdvancedKappaStatistic) KS is dened as:
KS(; I
F
, I
M
) =
2

xiF
1
IF (xi)>0,IM(T(xi))>0

xiF
1
IF (xi)>0
+1
IM(T(xi))>0
, (2.10)
where 1 is the indicator function.
The SSD measure is a measure that is only suited for two images with an equal intensity distribution,
i.e. for images from the same modality. NCC is less strict, it assumes a linear relation between the intensity
values of the xed and moving image, and can therefore be used more often. The MI measure is even more
general: only a relation between the probability distributions of the intensities of the xed and moving image
is assumed. For MI it is well-known that it is suited not only for mono-modal, but also for multi-modal image
pairs. This measure is often a good choice for image registration. The NMI measure is, just like MI, suitable
for mono- and multi-modality registration. Studholme et al. [1999] seems to indicate better performance
than MI in some cases. The KS measure is specically meant to register binary images (segmentations). It
measures the overlap of the segmentations.
2.4 Image samplers
In Equations (2.5)-(2.8) we observe a loop over the xed image:

xiF
. Until now, we assumed that the
loop goes over all voxels of the xed image. In general, this is not necessary. A subset may suce [Thevenaz
and Unser, 2000, Klein et al., 2007]. The subset may be selected in dierent ways: random, on a grid, etc.
The sampler component represents this process.
The following samplers are often used:
Full: (Full) A full sampler simply selects all voxel coordinates x
i
of the xed image.
Grid: (Grid) The grid sampler denes a regular grid on the xed image and selects the coordinates x
i
on
the grid. Eectively, the grid sampler thus downsamples the xed image (not preceded by smoothing).
The size of the grid (or equivalently, the downsampling factor, which is the original xed image size
divided by the grid size) is a user input.
Random: (Random) A random sampler randomly selects a user-specied number of voxels from the xed
image, whose coordinates form x
i
. Every voxel has equal chance to be selected. A sample is not
necessarily selected only once.
Random Coordinate: (RandomCoordinate) A random coordinate sampler is similar to a random sampler.
It also randomly selects a user-specied number of coordinates x
i
. However, the random coordinate
sampler is not limited to voxel positions. Coordinates between voxels can also be selected. The grey-
value I
F
(x
i
) at those locations must of course be obtained by interpolation.
While at rst sight the full sampler seems the most obvious choice, in practice it is not always used,
because of its computational costs in large images. The random samplers are especially useful in combination
with a stochastic optimisation method [Klein et al., 2007]. See also Section 2.7. The use of the random
coordinate sampler makes the cost function C a more smooth function of , which makes the optimisation
problem (2.4) easier to solve. This has been shown in Thevenaz and Unser [2008].
2.5 Interpolators
As stated previously, during the optimisation the value I
M
(T

(x)) is evaluated at non-voxel positions, for

which intensity interpolation is needed. Several methods for interpolation exist, varying in quality and speed.
Some examples are given in Figure 2.4.
7
(a) (b) (c) (d) (e)
Figure 2.4: Interpolation. (a) nearest neighbour, (b) linear, (c) B-spline N = 2, (d) B-spline N = 3, (e)
B-spline N = 5.
Nearest neighbour: (NearestNeighborInterpolator) This is the most simple technique, low in quality,
requiring little resources. The intensity of the voxel nearest in distance is returned.
Linear: (LinearInterpolator) The returned value is a weighted average of the surrounding voxels, with
the distance to each voxel taken as weight.
N-th order B-spline: (BSplineInterpolator or BSplineInterpolatorFloat for a memory ecient ver-
sion) The higher the order, the better the quality, but also requiring more computation time. In fact,
nearest neighbour (N = 0) and linear interpolation (N = 1) also fall in this category. See Unser [1999]
for more details.
During registration a rst-order B-spline interpolation, i.e. linear interpolation, often gives satisfactory
results. It is a good trade-o between quality and speed. To generate the nal result, i.e. the deformed
result of the registration, a higher-order interpolation is usually required, for which we recommend N = 3.
The nal result is generated in elastix by a so-called ResampleInterpolator. Any one of the above can
be used, but you need to prepend the name with Final, for example: FinalLinearInterpolator.
2.6 Transforms
The transformation model used for T

determines what type of deformations between the xed and moving

image you can handle. In order of increasing exibility, these are the translation, the rigid, the similarity,
the ane, the nonrigid B-spline and the nonrigid thin-plate spline like transformations.
Translation: (TranslationTransform) The translation is dened as:
T

(x) = x +t, (2.11)

with t the translation vector. The parameter vector is simply dened by = t.
Rigid: (EulerTransform) A rigid transformation is dened as:
T

(x) = R(x c) +t +c, (2.12)

with the matrix R a rotation matrix (i.e. orthonormal and proper), c the centre of rotation, and t
translation again. The image is treated as a rigid body, which can translate and rotate, but cannot be
scaled/stretched. The rotation matrix is parameterised by the Euler angles (one in 2D, three in 3D).
The parameter vector consists of the Euler angles (in rad) and the translation vector. In 2D, this
gives a vector of length 3: = (
z
, t
x
, t
y
)
T
, where
z
denotes the rotation around the axis normal to
the image. In 3D, this gives a vector of length 6: = (
x
,
y
,
z
, t
x
, t
y
, t
z
)
T
. The centre of rotation is
not part of ; it is a xed setting, usually the centre of the image.
8
Similarity: (SimilarityTransform) A similarity transformation is dened as
T

(x) = sR(x c) +t +c, (2.13)

with s a scalar and R a rotation matrix. This means that the image is treated as an object, which can
translate, rotate, and scale isotropically. The rotation matrix is parameterised by an angle in 2D, and by
a so-called versor in 3D (Euler angles could have been used as well). The parameter vector consists
of the angle/versor, the translation vector, and the isotropic scaling factor. In 2D, this gives a vector
of length 4: = (s,
z
, t
x
, t
y
)
T
. In 3D, this gives a vector of length 7: = (q
1
, q
2
, q
3
, t
x
, t
y
, t
z
, s)
T
,
where q
1
, q
2
, and q
3
are the elements of the versor. There are few cases when you need this transform.
Ane: (AffineTransform) An ane transformation is dened as:
T

(x) = A(x c) +t +c, (2.14)

where the matrix A has no restrictions. This means that the image can be translated, rotated, scaled,
and sheared. The parameter vector is formed by the matrix elements a
ij
and the translation vector.
In 2D, this gives a vector of length 6: = (a
11
, a
12
, a
21
, a
22
, t
x
, t
y
)
T
. In 3D, this gives a vector of
length 12.
B-splines: (BSplineTransform) For the category of non-rigid transformations, B-splines [Rueckert et al.,
1999] are often used as a parameterisation:
T

(x) = x +

x
k
Nx
p
k

x x
k

, (2.15)
with x
k
the control points,
3
(x) the cubic multidimensional B-spline polynomial [Unser, 1999], p
k
the
B-spline coecient vectors (loosely speaking, the control point displacements), the B-spline control
point spacing, and N
x
the set of all control points within the compact support of the B-spline at x.
The control points x
k
are dened on a regular grid, overlayed on the xed image. In this context we
talk about the control point grid that is put on the xed image, and about control points that are
moved around. Note that T

(x
k
) = x
k
+ p
k
, a common misunderstanding. Calling p
k
the control
point displacements is, therefore, actually somewhat misleading. Also note that the control point grid
is entirely unrelated to the grid used by the Grid image sampler, see Section 2.4.
The control point grid is dened by the amount of space between the control points = (
1
, . . . ,
d
)
(with d the image dimension), which can be dierent for each direction. B-splines have local support
(|N
x
| is small), which means that the transformation of a point can be computed from only a couple
of surrounding control points. This is benecial both for modelling local transformations, and for fast
computation. The parameters are formed by the B-spline coecients p
k
. The number of control
points P = (P
1
, . . . , P
d
) determines the number of parameters M, by M = (P
1
. . . P
d
) d. P
i
in
turn is determined by the image size s and the B-spline grid spacing, i.e. P
i
s
i
/
i
(where we use
since some additional control points are placed just outside the image). For 3D images, M 10000
parameters is not an unusual case, and M can easily grow to 10
5
10
6
. The parameter vector (for 2D
images) is composed as follows: = (p
1x
, p
2x
, . . . , p
P1
, p
1y
, p
2y
, . . . , p
P2
)
T
.
Thin-plate splines: (SplineKernelTransform) Thin-plate splines are another well-known representation
for nonrigid transformations. The thin-plate spline is an instance of the more general class of kernel-
based transforms Davis et al., Brooks and Arbel [2007]. The transformation is based on a set of
k = 1 . . . K corresponding landmarks in xed and moving image. The landmark displacements d
k
=
x
mov
k
x
x
k
form the parameter vector . The xed landmark positions x
x
k
are given by the user. The
transformation is expressed as a sum of an ane component and a nonrigid component:
T

(x) = x +Ax +t +

x
k
c
k
G(x x
x
k
), (2.16)
9
(a) xed (b) moving (c) translation
(d) rigid (e) ane (f) B-spline
Figure 2.5: Dierent transformations. (a) the xed image, (b) the moving image with a grid overlayed, (c)
the deformed moving image I
M
(T

(x)) with a translation transformation, (d) a rigid transformation, (e) an

ane transformation, and (f) a B-spline transformation. The deformed moving image nicely resembles the
xed image I
F
(x) using the B-spline transformation. The overlay grids give an indication of the deformations
imposed on the moving image. NB: the overlayed grid in (f) is NOT the B-spline control point grid, since
that one is dened on the xed image!
where x
k
are landmark positions in the xed image, G(r) is a basis function and c
k
are the coecients
corresponding to each landmark. The coecients c
k
and the elements of A and t are computed
from the landmark displacements. The specic choice of basis function G(r) determines the physical
behaviour. The most often used choice of G(r) leads to the thin-plate spline, but another useful
alternative is the elastic-body spline Davis et al.. The spline kernel transforms are often less ecient
than the B-splines (because they lack the compact support property of the B-splines), but they allow for
more exibility in placing the control points x
x
k
). The parameter vector is (for 2D images) composed
as follows: = (d
1x
, d
1y
, d
2x
, d
2y
, . . . , d
Kx
, d
Ky
)
T
. Note the dierence in ordering of compared to
the B-splines transform.
See Figure 2.5 for an illustration of dierent transforms. Choose the transformation that ts your needs:
only choose a nonrigid transformation if you expect that the underlying problem contains local deformations,
choose a rigid transformation if you only need to compensate for dierences in pose. To initialise a nonrigid
registration problem, perform a rigid or ane one rst. The result of the initial rigid or ane registration
T
0
is combined with a nonrigid transformation T
NR

in one of the following two ways:

addition: T

(x) = T
NR

(x) +T
0
(x) x (2.17)
composition: T

(x) = T
NR

(T
0
(x)) = (T
NR

T
0
)(x) (2.18)
10
Figure 2.6: Iterative optimisation. Example for registration with a translation transformation model. The
arrows indicate the steps a
k
d
k
taken in the direction of the optimum, which is the minimum of the cost
function.
The latter method is in general to be preferred, because it makes several postregistration analysis tasks
somewhat more straightforward.
2.7 Optimisers
To solve the optimisation problem (2.4), i.e. to obtain the optimal transformation parameter vector ,
commonly an iterative optimisation strategy is employed:

k+1
=
k
+a
k
d
k
, k = 0, 1, 2, , (2.19)
with d
k
the search direction at iteration k, a
k
a scalar gain factor controlling the step size along the search
direction. The optimisation process is illustrated in Figure 2.6. Klein et al. [2007] give an overview of
various optimisation routines the literature oers. Examples are quasi-Newton (QN), nonlinear conjugate
gradient (NCG), gradient descent (GD), and Robbins-Monro (RM). Gradient descent and Robbins-Monro
are discussed below. For details on other optimisation methods we refer to [Klein et al., 2007, Nocedal and
Wright, 1999].
Gradient descent (GD): (StandardGradientDescent or RegularStepGradientDescent) Gradient de-
scent optimisation methods take the search direction as the negative gradient of the cost function:

k+1
=
k
a
k
g(
k
), (2.20)
with g(
k
) = C/ evaluated at the current position
k
. Several choices exist for the gain factor a
k
.
It can for example be determined by a line search or by using a predened function of k.
11
Robbins-Monro (RM): (StandardGradientDescent or FiniteDifferenceGradientDescent) The RM
optimisation method replaces the calculation of the derivative of the cost function g(
k
) by an ap-
proximation g
k
.

k+1
=
k
a
k
g
k
, (2.21)
The approximation is potentially faster to compute, but might deteriorate convergence properties of
the GD scheme, since every iteration an approximation error g(
k
) g
k
is made. Klein et al. [2007]
showed that using only a small random subset of voxels ( 2000) from the xed image accelerates regis-
tration signicantly, without compromising registration accuracy. The Random or RandomCoordinate
samplers, described in Section 2.4, are examples of samplers that pick voxels randomly. It is important
that a new subset of xed image voxels is selected every iteration k, so that the approximation error
has zero mean. The RM method is usually combined with a
k
as a predened decaying function of k:
a
k
=
a
(k +A)

, (2.22)
where a > 0, A 1, and 0 1 are user-dened constants. In our experience, a reasonable choice
is 0.6 and A approximately 10% of the user-dened maximum number of iterations, or less. The
choice of the overall gain, a, depends on the expected ranges of and g and is thus problem-specic. In
our experience, the registration result is not very sensitive to small perturbations of these parameters.
Section 5.3.6 gives some more advice.
Note that GD and RM are in fact very similar. Running RM with a Full sampler (see Section 2.4),
instead of a Random sampler, is equivalent to performing GD. We recommend the use of RM over GD, since
it is so much faster, without compromising on accuracy. In that case, the parameter a is the parameter
that is to be tuned for your application. A more advanced version of the StandardGradientDescent is
the AdaptiveStochasticGradientDescent, which requires less parameters to be set and tends to be more
robust Klein et al. [2009].
Other optimisers available in elastix are: FullSearch, ConjugateGradient, ConjugateGradientFRPR,
QuasiNewtonLBFGS, RSGDEachParameterApart, SimultaneousPerturbation, CMAEvolutionStrategy.
2.8 Multi-resolution
For a good overview of multi-resolution strategies see Lester and Arridge [1999]. Two hierarchical methods
are distinguished: reduction of data complexity, and reduction of transformation complexity.
2.8.1 Data complexity
It is common to start the registration process using images that have lower complexity, e.g., images that are
smoothed and possibly downsampled. This increases the chance of successful registration. A series of images
with increasing amount of smoothing is called a scale space. If the images are not only smoothed, but also
downsampled, the data is not only less complex, but the amount of data is actually reduced. In that case,
we talk about a pyramid. However, confusingly, the word pyramid is used by us also to refer to a scale
space. Several scale spaces or pyramids are found in the literature, amongst others Gaussian and Laplacian
pyramids, morphological scale space, and spline and wavelet pyramids. The Gaussian pyramid is the most
common one. In elastix we have:
Gaussian pyramid: (FixedRecursiveImagePyramid and MovingRecursiveImagePyramid) Applies smooth-
ing and down-sampling.
Gaussian scale space: (FixedSmoothingImagePyramid and MovingSmoothingImagePyramid) Applies smooth-
ing and no down-sampling.
12
(a) resolution 0 (b) resolution 1 (c) resolution 2 (d) original
(e) resolution 0 (f) resolution 1 (g) resolution 2 (h) original
Figure 2.7: Two multi-resolution strategies using a Gaussian pyramid ( = 8.0, 4.0, 2.0 voxels). The rst
row shows multi-resolution with down-sampling (FixedRecursiveImagePyramid), the second row without
(FixedSmoothingImagePyramid). Note that in the rst row, for each dimension, the image size is halved
every resolution, but that the voxel size increases with a factor 2, so physically the images are of the same
size every resolution.
Shrinking pyramid: (FixedShrinkingImagePyramid and MovingShrinkingImagePyramid) Applies no smooth-
ing, but only down-sampling.
Figure 2.7 shows the Gaussian pyramid with and without downsampling. In combination with a Full
sampler (see Section 2.4), using a pyramid with downsampling will save a lot of time in the rst resolution
levels, because the image contains much fewer voxels. In combination with a Random sampler, or Random-
Coordinate, the downsampling step is not necessary, since the random samplers select a user-dened number
of samples anyway, independent of the image size.
2.8.2 Transformation complexity
The second multiresolution strategy is to start the registration with fewer degrees of freedom for the trans-
formation model. The degrees of freedom of the transformation equals the length (number of elements) of
the parameter vector .
An example of this was already mentioned in Section 2.6: the use of a rigid transformation prior to
nonrigid (B-spline) registration. We may even use a three-level strategy: rst rigid, then ane, then nonrigid
B-spline.
Another example is to increase the number of degrees of freedom within the transformation model. With
a B-spline transformation, it is often good practice to start registration with a coarse control point grid, only
capable of modelling coarse deformations. In subsequent resolutions the B-spline grid is gradually rened,
thereby introducing the capability to match smaller structures. See Section 5.3.5.
2.9 Evaluating registration
How do you verify that your registration was successful? This is a dicult problem. In general, you dont
know for each voxel where it should map to. Here are some hints:
13
The deformed moving image I
M
(T

(x)) should look similar to the xed image I

F
(x). So, compare
images side by side in a viewer. You can also display the two images on top of each other with a
checkerboard view or a dragable cross. Besides looking similar, also check that the deformed moving
image has the same texture as the moving image. Sudden blurred areas in the deformed image may
indicate that the deformation at that region is too large.
For mono-modal image data you can inspect the dierence image. Perfect registration would result in
a dierence image without any edges, just noise.
Compute the overlap of segmented anatomical structures after registration. The better the overlap,
the better the registration. To measure overlap, commonly the Dice similarity coecient (DSC) is
used:
DSC(X, Y ) =
2|X Y |
|X| +|Y |
, (2.23)
where X and Y represent the binary label images, and | | denotes the number of voxels that equal 1.
A higher DSC indicates a better correspondence. A value of 1 indicates perfect overlap, a value of 0
means no overlap at all. Also the Tannimoto (TC) coecient is used often. It is related to the DSC by
DSC = 2TC/(TC + 1). See also Crum et al. [2006]. It is important to realise that the surface-volume
ratio of the segmented structures inuences the overlap values you typically get [Rohlng et al., 2004].
A value of DSC = 0.8 would be very good for the overlap of complex vessel structures. For large
spherical objects though, an overlap < 0.9 is in general not very good. What is good enough depends
of course on your application.
Compute the distance after registration between points that you know correspond. You can obtain
corresponding points by manually clicking them in the xed and the moving image. A less time-
consuming option is the semi-automated approach of Murphy et al. Murphy et al. [2008], which is
designed for nding corresponding points in the lung. Ideally, the registration has found the same
correspondence as the ground truth.
Inspect the deformation eld by looking at the determinant of the Jacobian of T

(x). Values smaller

than 1 indicate local compression, values larger than 1 indicate local expansion, and 1 means volume
preservation. If this value deviates substantially from 1, you may be worried (but maybe not if this is
what you expect for your application). In case it is negative you have foldings in your transformation,
and you denitely should be worried.
Inspect the convergence, by computing for each iteration the exact metric value (and not an approxi-
mated value, when you do random sampling), and plot it. For example for the SSD measure, the lower
the metric value, the better the registration.
14
Chapter 3
elastix
3.1 Introduction
The development of elastix started half to late 2003, and was intended to facilitate our registration research.
After some initial versions we decided to put the separate components of elastix in separate libraries. This
resulted in major version 3.0 in November 2004. elastix 3.0 was also the rst version that was made
publicly available on the elastix website, around the same time. The continued development brings us
today (November 12, 2010) to version 4.4.
what where
website http://elastix.isi.uu.nl
SVN repository https://svn.bigr.nl/elastix/trunkpublic
WIKI http://elastix.isi.uu.nl/wiki.php
FAQ http://elastix.isi.uu.nl/FAQ.php
Mailing list [email protected] (subscription via website)
The website also contains a doxygen
1
generated part that provides documentation of the source code.
An overview of all available classes can be found at
http://elastix.isi.uu.nl/doxygen/classes.html.
For each class a description of this class is given, together with information on how to use it in elastix. See
http://elastix.isi.uu.nl/doxygen/modules.html
for an overview of all available components.
3.1.1 Key features
elastix is
open source, freely available from http://elastix.isi.uu.nl;
based on the ITK, so the code base is thoroughly tested. Quite some modications/additions are made
to the original ITK code though, such as the use of samplers, a transformation class that combines
multiple transformation using composition or addition, and more.
1
http://www.doxygen.org
15
suitable for many image formats. The use of ITK implies that all image formats supported by ITK
are supported by elastix. Some often used (medical) image formats are: .mhd (MetaIO), .hdr (Ana-
lyze/NIfTI), .gipl, .dcm (DICOM slices). DICOM directories are not directly supported by elastix;
multi-platform (at least Windows, Linux and Mac OS), multi-compiler (at least Visual C++ 2005,
2008, gcc 3.2, 4.02, 4.1.2), and supports 32 and 64 bit systems. The underlying ITK code builds on
many more platforms, see www.itk.org/Wiki/ITK_Prerequisites. So, it is highly portable to the
platform of the users choice;
highly congurable: there is a lot of choice for all the registration components. Choosing the congu-
ration that suits your needs is easy thanks to human readable and editable parameter le;
easy to use for large amounts of data, since elastix can be called easily in a script;
fast, thanks to stochastic subsampling, if desired;
relatively easy to extend, i.e. to add new components, so it is very suited for research also.
3.2 Getting started
This section describes how you can install elastix, either directly from the binaries, or by compiling elastix
yourself.
3.2.1 Getting started the really easy way
The easiest way to get started with elastix is to use the pre-compiled binaries.
1. Download the compressed archive from the website:
http://elastix.isi.uu.nl/download.php
2. Extract the archive to a folder of your choice.
3. Make sure your operating system can nd the program:
(a) Windows XP: Go to the control panel, go to System, go to the tab Advanced, click Envi-
ronmental variables, add folder to the variable path.
(b) Linux: Add the following lines to your .bashrc le:
export PATH=folder/bin:$PATH
export LD LIBRARY PATH=folder/lib:$LD LIBRARY PATH
or call elastix with the full path: fullPathToFolder\elastix. Note that in Linux, you will have to
set the LD LIBRARY PATH anyway.
3.2.2 Getting started the easy way
It is also possible to compile elastix yourself, since the source code is freely available. In this section, we
assume you use the Microsoft Visual C++ 2003 (or higher) compiler under Windows, and the GCC compiler
under Linux.
1. Download and install CMake: www.cmake.org.
16
2. Download and compile the ITK version 3.20: www.itk.org. Make sure to set the following (advanced)
CMake variables to ON: ITK USE REVIEW, ITK USE REVIEW STATISTICS, ITK USE ORIENTED IMAGE DIRECTION,
ITK IMAGE BEHAVES AS ORIENTED IMAGE, ITK USE CENTERED PIXEL COORDINATES CONSISTENTLY. On
Linux systems with the GCC compiler: set CMAKE BUILD TYPE to Release.
3. Obtain the sources. There are three possibilities:
(a) Download the compressed sources from the website. Extract the archive to <your-elastix-folder>
of your choice.
(b) Use subversion (https://subversion.tigris.org) to check out the release from the subversion
repository:
svn co --username elastixguest --password elastixguest
https://svn.bigr.nl/elastix/tagspublic/elastix_XX_X <your-elastix-folder>
where XX X is the major version number (rst 2 digits) and the minor version number (1 digit).
So, for example, for version 4.5 this would be 04 5.
(c) Use subversion to check out the latest development version. NB: this version might be unstable!
svn co --username elastixguest --password elastixguest
https://svn.bigr.nl/elastix/trunkpublic <your-elastix-folder>
4. Run CMake for elastix:
(a) Windows: start CMake. Find the src folder with the source code. Set the folder where you
want the binaries to be created. Click Congure and select the compiler that you use. Set the
CMAKE INSTALL PREFIX to the directory where you want elastix installed. Click Congure
until all cache values are no longer red, and click OK.
(b) Linux: run CMake from the folder where you want the binaries to be created, with as command
line argument the folder in which the sources were extracted: ccmake <src-folder>. Set the
CMAKE BUILD TYPE to Release and the CMAKE INSTALL PREFIX to the directory where you want
elastix installed.
CMake will create a project or solution or make le for your compiler.
5. Compile elastix by opening the project and selecting compile in release mode, or on Linux by
running make install. Your compiler will now create the elastix binaries. On Windows you can
perform the installation step (which copies the binaries to the CMAKE INSTALL PREFIX directory) by
also compiling the INSTALL project.
6. Make sure your operating system can nd the program, see above.
For developers: When running CMake, you may toggle the display of the advanced options. In this
list you will nd several options like USE BSplineTransform ON/OFF. By default only the most commonly
used components are ON. To reduce compilation time, you may turn some components OFF, which you do
not plan to use anyway. Be careful though to not turn o essential components. The released binaries are
compiled with all components ON.
3.3 How to call elastix
elastix is a command line program. This means that you have to open a command line interface (a DOS-
box, a shell) and type in an appropriate elastix command. This also means that there is no graphical user
interface. Help on using the program can be acquired as follows:
17
elastix --help
which will give a list of mandatory and optional arguments. The most basic command to run a registration
is as follows:
elastix -f fixedImage.ext -m movingImage.ext -out outputDirectory -p parameterFile.txt
where ext is the extension of the image les. The above arguments are mandatory. These are minimally
needed to run elastix. The parameter le is an important le: it contains, in normal text, what kind of
registration is performed (i.e. what metric, optimiser, etc.) and what the parameters are that dene the
registration. It gives a high amount of exibility and control over the process. More information about the
parameter le is given in Section 3.4. All output of elastix is written to the output directory, which needs
to be created before running elastix. The output consists of a log le (elastix.log), the parameters of
the transformation T

that relates the xed and the moving image (TransformParameters.?.txt), and,
optionally, the resulting registered image I
M
(T

(x)) (result.?.mhd). The log le contains all messages

that were print to screen during registration. Also the parameterFile.txt is copied into the log le, and
the contents of the TransformParameters.?.txt les are included. The log le is thus especially useful for
trouble shooting.
Besides the mandatory arguments, there are some optional arguments. Mask images can be provided by
adding -fMask fixedMask.ext and/or -mMask movingMask.ext to the command line. An initial transfor-
mation can be provided with a valid transform parameter le by adding -t0 TransformParameters.txt
to the command line. With the command line option -threads unsigned int the user can specify the
maximum number of threads that elastix will use.
Running multiple registrations in succession, each possibly of a dierent type, and with the output of a
previous registration as input to the next, can be done with elastix in several ways. The rst one is to run
elastix once with the rst registration, and use its output (the TransformParameter.0.txt that can be
found in the output directory) as input for a new run of elastix with the command line argument -t0. So:
elastix -f ... -m ... -out out1 -p param1.txt
elastix -f ... -m ... -out out2 -p param2.txt -t0 out1/TransformParameters.0.txt
elastix -f ... -m ... -out out3 -p param3.txt -t0 out2/TransformParameters.0.txt
and so on. Another possibility is combine the registrations with one run of elastix:
elastix ... -p param1.txt -p param2.txt -p param3.txt
The transformations from each of the registrations are automatically combined, using one of the equations
(2.17) and (2.18).
On the elastix-website, in the About section, you can nd an example on how to use the program.
Maybe now is the time to try the example and see a registration in action.
3.4 The parameter le
The parameter le is a text le that denes the components of the registration and their parameter values.
Supplying a parameter works as follows:
(ParameterName value(s))
So parameters are provided between brackets, rst the name, followed by one or more values. If the value is
of type string then the values need to be quoted:
(ParameterName "value1" ... "valueN")
18
Comments can be provided by starting the line with //. A minimal example of a valid parameter le is
given in Appendix A. A list of available parameters for each class is given at http://elastix.isi.uu.nl/
doxygen/parameter.html. Examples of parameter les can be found at the wiki: http://elastix.bigr.
nl/wiki/index.php/Parameter_file_database.
Since the choice of the several components and the parameter values dene the registration, it is very
important to set them wisely. These choices are what make the registration a success or a disaster. Therefore,
a separate chapter is dedicated to the ne art of tuning a registration, see Chapter 5.
19
Chapter 4
transformix
4.1 Introduction
By now you are able to at least run a registration, by calling elastix correctly. It is often also useful
to apply the transformation as found by the registration to another image. Maybe you want to apply the
transformation to an original (larger) image to gain resolution. Or maybe you need the transformation to
apply it to a label image (segmentation). For those purposes a program called transformix is available. It
was developed simultaneously with elastix.
4.2 How to call transformix
Like elastix, transformix is a command line driven program. You can get basic help on how to call it, by:
transformix --help
which will give a list of mandatory and optional arguments.
The most basic command is as follows:
transformix -in inputImage.ext -out outputDirectory -tp TransformParameters.txt
This call will transform the input image and write it, together with a log le transformix.log, to the output
directory. The transformation you want to apply is dened in the transform parameter le. The transform
parameter le could be the result of a previous run of elastix (see Section 3.3), but may also be written
by yourself. Section 4.3 explains the structure and contents that a transform parameter le should have.
Besides using transformix for deforming images, you can also use transformix to evaluate the trans-
formation T

(x) at some points x. If you want to deform a set of user-specied points, the appropriate call
is:
transformix -def inputPoints.txt -out outputDirectory -tp TransformParameters.txt
This will create a le outputpoints.txt containing the input points x and the transformed points T

(x)
(both given as voxel indices of the xed image and as physical coordinates), the displacement vector T

(x)x
(in physical coordinates), and, if -in inputImage.ext is also specied, the transformed output points as
indices of the input image
1
. The inputPoints.txt le should have the following structure:
1
The downside of this is that the input image is also deformed, which consumes time and may not be needed by the user.
If this is a problem, just run transformix without -in and compute the voxel indices yourself, based on the T(x) physical
coordinate data.
20
<index, point>
<number of points>
point1 x point1 y [point1 z]
point2 x point2 y [point2 z]
. . .
The rst line indicates whether the points are given as indices (of the xed image), or as points (in
physical coordinates). The second line stores the number of points that will be specied. After that the
point data is given.
Instead of the custom .txt format for the input points, transformix also supports .vtk les:
transformix -def inputPoints.vtk -out outputDirectory -tp TransformParameters.txt
The output is then saved as outputpoints.vtk. The support for .vtk les is still a bit limited. Currently,
only ASCII les are supported, with triangle meshes. Any meta point data is lost in the output le.
If you want to know the deformation at all voxels of the xed image, simply use -def all:
transformix -def all -out outputDirectory -tp TransformParameters.txt
The deformation eld is stored as a vector image deformationField.mhd. Each voxel contains the displace-
ment vector T

(x) x in physical coordinates. The elements of the vectors are stored as float values.
In addition to computing the deformation eld, transformix has the capability to compute the spatial
Jacobian of the transformation. The determinant of the spatial Jacobian identies the amount of local
compression or expansion and can be quite useful, for example in lung ventilation studies. The determinant
of the spatial Jacobian can be computed on the entire image only using:
transformix -jac all -out outputDirectory -tp TransformParameters.txt
The complete spatial Jacobian matrix can also be computed:
transformix -jacmat all -out outputDirectory -tp TransformParameters.txt
where each voxel is lled with a dd matrix, with d the image dimension, instead of a simply a scalar value.
With the command-line option -threads unsigned int the user can specify the maximum number of
threads that transformix will use.
4.3 The transform parameter le
The result of a registration is the transformation T

relating the xed and moving image. The parameters

of this transformation are stored in a TransformParameters.?.txt-le. An example of its structure for a
2D rigid transformation is given in Appendix B. The text le contains all information necessary to resample
an input image (the moving image) to the region specied in the le (by default the xed image region).
The transform parameter le can be manually edited or created as is convenient for the user. Mul-
tiple transformations are composed by iteratively supplying another transform parameter le with the
InitialTransformParametersFileName tag. The last transformation will be the one where the initial
transform parameter le name is set to "NoInitialTransform".
An important parameter in the transform parameter les is the FinalBSplineInterpolationOrder.
Usually it is set to 3, because that produces the best quality result image after registration, see Sec 5.3.4.
However, if you use transformix to deform a segmentation of the moving image (so, a binary image), you
need to manually change the FinalBSplineInterpolationOrder to 0. This will make sure that the deformed
segmentation is still a binary label image. If third order interpolation is used, the deformed segmentation
image will contain garbage. This is related to the overshoot-property of higher-order B-spline interpolation.
21
4.4 Some details
4.4.1 Run-time
The run-time of transformix is built up of the following parts:
1. Computing the B-spline decomposition of the input image (in case you selected the FinalBSpline-
Interpolator);
2. Computing the transformation for each voxel;
3. Interpolating the input image for each voxel.
We have never performed tests to measure the computational complexity of each step, but we think that
step 1 is the least time-consuming task. This step can obviously be avoided by using a nearest neighbour or
linear interpolator. Step 2 is dependent on the choice of the transformation, where linear transformations,
such as the rigid and ane transform, are substantially faster than nonlinear transforms, such as the B-spline
transform. Step 3 depends on the specic interpolator. In order of increasing complexity: nearest neighbour,
linear, 1st-order B-spline, 2nd-order B-spline, etc.
4.4.2 Memory consumption
For more information about memory consumption, see Section 5.5.3 and also:
http://elastix.bigr.nl/wiki/index.php/Memory_consumption_transformix
22
Chapter 5
Tutorial
5.1 Selecting the registration components
When performing registration one carefully has to choose the several components, as specied in Chapter 2.
The components must be specied in the parameter le. For example:
(Transform "BSplineTransform")
(Metric "AdvancedMattesMutualInformation")
. . .
In Table 5.1 a list of the components that need to be specied is given, with some recommendations. The
Registration component was not mentioned in Chapter 2. The registration component serves to connect all
other components and implements the multiresolution aspect of registration. So, one may say that it actually
implements the block scheme of Figure 2.2. Also the Resampler component was not explicitly mentioned
in Chapter 2. It simply serves to generate the deformed moving image after registration. Currently there is
only one Resampler available in elastix: the DefaultResampler. The component is therefore not further
discussed.
Component Recommendation
Registration MultiResolutionRegistration
Metric AdvancedMattesMutualInformation
Sampler RandomCoordinate
Interpolator BSplineInterpolator
ResampleInterpolator FinalBSplineInterpolator
Resampler DefaultResampler
Transform Depends on the application
Optimizer StandardGradientDescent/
AdaptiveStochasticGradientDescent
FixedImagePyramid FixedSmoothingImagePyramid
MovingImagePyramid MovingSmoothingImagePyramid
Table 5.1: Some recommendations for the several components.
5.2 Overview of all parameters
A list of all available components of elastix can be found at:
23
http://elastix.isi.uu.nl/doxygen/modules.html
A list of all parameters that can be specied for each registration component can be found at the elastix
website:
http://elastix.isi.uu.nl/doxygen/parameter.html
At that site you can nd how to specify a parameter and what the default value is. We have tried to come
up with sensible defaults, although the defaults will certainly not work in all cases. A collection of successful
parameter les can be found at the wiki:
http://elastix.bigr.nl/wiki/index.php/Parameter_file_database
This may get you started with your particular application.
5.3 Important parameters
In the same order as in Section 2.3 we discuss the important parameters for each component and explain
the recommendations made in Table 5.1.
5.3.1 Registration
Just use the MultiResolutionRegistration method, since multi-resolution is a good idea. And if you still
think you dont need all this multi-resolution, you can always set the NumberOfResolutions to 1. You dont
have to set anything else. Section 5.3.7 discusses the number of resolutions in more detail.
5.3.2 Metric
The AdvancedMattesMutualInformation usually works well, both for mono- and multi-modal images. It
supports fast computation of the metric value and derivative in case the transform is a B-spline by exploiting
its compact support. You need to set the number of histogram bins, which is needed to compute the joint
histogram. A good value for this depends on the dynamic range of your input images, but in our experience
32 is usually ok:
(NumberOfHistogramBins 32)
5.3.3 Sampler
The RandomCoordinate sampler works well in conjunction with the StandardGradientDescent and
AdaptiveStochasticGradientDescent optimisers, which are the recommended optimisation routines. These
optimisation methods can be used with a small amount of samples, randomly selected in every iteration,
see Section 2.7, which signicantly decreases registration time. Set the NumberOfSpatialSamples to 3000.
Dont go lower than 2000. Compared to samplers that draw samples on the voxel-grid (such as the Random
sampler), the RandomCoordinate sampler avoids what is known as the grid-eect [Thevenaz and Unser,
2008].
An important option for the random samplers, discussed in Section 5.3.6 is:
(NewSamplesEveryIteration "true")
which enforces the selection of new samples in every iteration.
An interesting option for the RandomCoordinate sampler is the UseRandomSampleRegion parameter,
used in combination with the SampleRegionSize parameter. If UseRandomSampleRegion is set to "false"
(the default), the sampler draws samples from the entire image domain. When set to "true", the sampler
24
randomly selects one voxel, and then selects the remaining samples in a square neighbourhood around that
voxel. The size of the neighbourhood is determined by the SampleRegionSize (in physical coordinates). An
example for 3D images:
(ImageSampler "RandomCoordinate")
(NewSamplesEveryIteration "true")
(UseRandomSampleRegion "true")
(SampleRegionSize 50.0 50.0 50.0)
(NumberOfSpatialSamples 2000)
In every iteration, a square region of 50
3
mm is randomly selected. In that region, 2000 samples are selected
according to a uniform distribution. Eectively, a kind of localized similarity measure is obtained, which
sometimes gives better registration results. See Klein et al. [2008] for more information on this approach.
For the sample region size a reasonable value to try is 1/3 of the image size.
5.3.4 Interpolator
During the registration, use the BSplineInterpolator with rst order interpolation:
(BSplineInterpolationOrder 1)
We recommend this instead of the LinearInterpolator, since the BSplineInterpolator has a dedicated
function to compute the derivative of the moving image.
We recommend a higher quality third order B-spline interpolator for generating the resulting deformed
moving image:
(ResampleInterpolator "FinalBSplineInterpolator")
(FinalBSplineInterpolationOrder 3)
5.3.5 Transform
This choice depends on the application at hand. For images of the same patient where you expect no
nonrigid deformation, you can consider a rigid transformation, i.e. choose the EulerTransform. If you want
to compensate for dierences in scale, consider the ane transformation: AffineTransform. These two
transformations require a centre of rotation, which can be set by the user. By default the geometric centre
of the xed image is taken, which is recommended. Another parameter that needs to be set is the Scales.
The scales dene for each element of the transformation parameters a scaling value, which is used during
optimisation. The scaling serves to bring the elements of in the same range (parameters corresponding to
rotation have in general a much smaller range than parameters corresponding to translation). We recommend
to let elastix compute it automatically: (AutomaticScalesEstimation "true")
1
. Always start with a
rigid or ane transformation before doing a nonrigid one, to get a good initial alignment.
For nonrigid registration problems elastix has the BSplineTransform. The B-spline nonrigid trans-
formation is dened by a uniform grid of control points. This grid is dened by the spacing between the
grid nodes. The spacing denes how dense the grid is, or what the locality is of the transformation you can
model. For each resolution level you can dene a dierent grid spacing. This is what we call multi-grid. In
general, we recommend to start with a coarse B-spline grid, i.e. a more global transformation. This way
the larger structures are matched rst, for the same reason as why you should start with a rigid or ane
transformation. In later resolutions you can rene the transformation in a stepwise fashion; the idea is that
you subsequently match smaller structures, up to the nal precision. The nal grid spacing is specied with:
1
The implementation is given in elastix\src\Core\ComponentBaseClasses\elxTransformBase.hxx
25
(FinalGridSpacingInPhysicalUnits 10.0 10.0 10.0)
with as much numbers as there are dimensions in your image. The spacing is in most medical images specied
in millimetres. It is also possible to specify the grid in voxel units:
(FinalGridSpacingInVoxels 16.0 16.0 16.0)
If the nal B-spline grid spacing is chosen high, then you cannot match small structures. On the other
hand, if the grid spacing is chosen very low, then small structures can be matched, but you possibly allow
the transformation to have too much freedom. This can result in irregular transformations, especially on
homogenous parts of your image, since there are no edges (or other information) at such areas that can guide
the registration. A penalty or regularisation term, see Equation (2.2), can help to avoid these problems. It
is hard to recommend a value for the nal grid spacing, since it depends on the desired accuracy. But we can
try: if you are interested in somewhat larger structures, you could set it to 32 voxels, for matching smaller
structures you could go down to 16 or 8 voxels, or even up to 4. The last choice will maybe require some
regularisation term, unless maybe if you have carefully and gradually rened the grid spacing.
To specify a multi-grid schedule use the GridSpacingSchedule command:
(NumberOfResolutions 4)
(FinalGridSpacingInVoxels 8.0 8.0)
(GridSpacingSchedule 6.0 6.0 4.0 4.0 2.5 2.5 1.0 1.0)
The GridSpacingSchedule denes the multiplication factors for all resolution levels. In combination with
the nal grid spacing, the grid spacing for all resolution levels is determined. In case of 2D images, the above
schedule species a grid spacing of 6 8 = 48 voxels in resolution level 0, via 32 and 20 voxels, to 8 voxels
in the nal resolution level. The default value for the GridSpacingSchedule uses a powers-of-2 scheme:
(GridSpacingSchedule 8.0 8.0 4.0 4.0 2.0 2.0 1.0 1.0) (for 2D images).
As a side-note: the number of parameters that are minimised in (2.1) is determined by the size of , i.e.
in case of the B-spline deformable transform by the control point grid spacing. If you double the spacing, the
number of parameters are increased by a factor 8 for a 3D image. For a 256
3
image and a grid spacing of 16
voxels this will result in approximately (256/16)
3
3 12.000 parameters; for a grid spacing of 8 voxels this
is almost 100.000 parameters. The amount of parameters can be directly related to memory consumption
and registration time, depending on the specic implementation.
In most literature, cubic (3
rd
-order) B-splines are used for image registration. Other spline orders are also
possible of course. You may experiment with the BSplineTransformSplineOrder option. Orders 1, 2, and
3 are supported. A lower order will reduce the computation time, but may cause less smooth deformations.
With order 1, the deformation eld is not even dierentiable anymore, strictly speaking.
As an alternative to the B-spline transform, elastix includes the SplineKernelTransform, which imple-
ments a thin-plate spline type of transform. See Sections 2.6 and 6.4 for more information on this transform.
Lastly, it is wise to include the following line in every parameter le:
(HowToCombineTransforms "Compose")
Up to elastix version 4.2, by default, if this line was omitted, "Add" was used for backwards compatibility.
From elastix version 4.3, the default has been changed to "Compose", which is better in most applications.
See Section 2.6, Equations (2.17) and (2.18) for explanation.
5.3.6 Optimiser
The StandardGradientDescent method, see Equations (2.20) and (2.22) oers the possibility to perform
fast registration, see Klein et al. [2007]. The key idea is that you use a random subset of voxels (samples),
newly selected in each iteration, to compute the cost function derivatives. The number of samples to use is
specied using the parameter NumberOfSpatialSamples, see Section 5.3.3. Typically, 2000-5000 is enough.
It is important to tell the optimiser to select new samples in every iteration:
26
0 50 100 150 200 250 300 350 400 450 500
0
10
20
30
40
50
60
70
80
90
100


sequence 1
sequence 2
k
a
k
Figure 5.1: Sequence 1: a = 1000, A = 50, = 0.602. Sequence 2: a = 400, A = 10, = 0.602. Both
sequences start with the same step size, but sequence 2 decays much faster.
(NewSamplesEveryIteration "true")
A downside of the StandardGradientDescent method is that you need to tune the parameters of the
gain factor a
k
, see Section 5.3. Equation (2.21) needs a choice for the step size a
k
, which is in elastix
dened as in Equation (2.22). Figure 5.2 gives some examples.
The parameters and A dene the decay slope of the function. For the parameter we recommend the
value 0.6. For A, use something in the order of 50:
(SP alpha 0.6)
(SP A 50.0)
This leaves the parameter a, called SP a in elastix, as the most important parameter to tune. And it is
an important parameter, it can mean success for a good choice and failure if not! If a is set too high, the
iterative solving algorithm (2.21) becomes unstable, and you may deform your image beyond recognition.
If a is set too low, you will never make it to the optimum, or may get stuck in a very small nearby local
optimum. Figure 5.2 illustrates this. A good choice for a is dependent on the cost function that is used for
registration: the a that will give you a good result for SSD is not the same as the one that gives a good
result for MI. Finally, a also depends on the amount of deformation that you expect between the xed and
the moving image. So again, recommendations are hard to give. In general we advise you to think in orders
of magnitude, if a = 10 is too small, try a = 100 and not a = 11. For mutual information, normalised
correlation coecient, and normalised mutual information, you could start around a = 1000. For the mean
squared dierence metric you could try something smaller than 1. If a is chosen way too big, you may
encounter the error message Too many samples map outside moving image buer. This error may have
other causes as well though. The FAQ at the elastix website gives more information on this error message.
For every resolution you could specify a dierent value of SP a, but it might be easier to start with the
same value for every resolution.
(SP a 1000.0 1000.0 1000.0)
or, equivalently:
(SP a 1000.0)
The last important option related to the optimiser is the maximum number of iterations:
27
(a) (b)
(c) (d) (e)
Figure 5.2: The eect of the choice of the step size a (SP A). This example can be downloaded from the
elastix website. (a) the xed image, (b) the moving image. (c)-(e) show the registered images, with (c)
a = 320 is too small, (d) a = 3200 as in the downloadable example is good, (e) a = 32000 is too large.
(MaximumNumberOfIterations 500)
which is, in the case of StandardGradientDescent, not only the maximum, but also the minimum, since
there is no other stopping condition implemented for this optimiser. In general, the more iterations, the
better the registration result. But, of course, more iterations take more time. A value of 500 is a good start.
Use 2000 if computation time is not such an issue. You may try to go down to 200 iterations if you are in a
hurry. For small 2D images, and rigid registration, even less iterations may suce. A side benet of using
more iterations is that a wider range of SP a gives good results. Tuning SP a then becomes easier.
A relatively new development is that of an adaptive stochastic gradient descent algorithm Klein et al.
[2009], which does not require tuning a. This optimiser is covered in Chapter 6.
5.3.7 Image pyramids
The FixedImagePyramid and the MovingImagePyramid have identical options. What is said below about
the FixedImagePyramid works similarly for the MovingImagePyramid.
Use the FixedSmoothingImagePyramid, since it will not throw away valuable information, and since
you are not using the FullSampler anyway, down-sampling will not save you any time. It may consume
quite some memory though for large images and many resolution levels. Two parameters have to be set
to dene the multi-resolution strategy: the number of resolutions (NumberOfResolutions) and the specic
down-sampling schedule that is used in each resolution (FixedImagePyramidSchedule). If you only set the
NumberOfResolutions, a default schedule will be used that smoothes the xed image by a factor of 2 in
each dimension, starting from = 0.5 in the last resolution. That schedule is usually ne. In case you have
highly anisotropic data, you might want to blur less in the direction of the largest spacing.
In general 3 resolutions is a good starting point. If the xed and moving image are initially far away, you
28
can increase the number of resolution levels to, say, 5 or 6. This way the images are more blurred and more
attention is paid to register large, dominant structures.
The pyramid schedule denes the amount of blurring (and down-sampling in case a FixedRecursiveImagePyramid
is used), in each direction x, y, z and for each resolution level. It can be specied as follows:
(NumberOfResolutions 4)
(FixedImagePyramidSchedule 8 8 4 4 2 2 1 1)
In this example 4 resolutions for a 2D image are used. At resolution level 0 the image is blurred with = 8/2
voxels in each direction ( is half the pyramid schedule value). At level 1 = 4/2 is used, and nally at
the last level, level 4, the original images are used for registration. Specifying the xed and moving image
pyramids with an identical schedule can be done with one command:
(ImagePyramidSchedule 4 4 2 2 2 1 1 1 1)
for a 3D image with 3 resolution levels, where less smoothing is performed in the z-direction.
5.4 Masks
Sometimes you are specically interested in aligning only a part of the image. A possibility to focus on this
part is to crop the image. Cropping, however, restricts the region of interest (ROI) to be a square (2D) or
cube (3D) only. If you need an irregular shaped ROI, you can use masks. A mask is a binary image, lled
with 0s and 1s. If you use a mask, you only perform registration on the part of the image that is within
the masks, i.e. where the mask has 1s.
You can/should use a mask
when your image contains an articial edge that has no real meaning. The registration might be
tempted to align these articial edges, thereby neglecting the meaningful edges. The conic beam edge
in ultrasound images is an example of such an articial edge.
when the image contains structures in the neighbourhood of your ROI that may inuence the reg-
istration within your ROI. This is for example the case when matching lung data. Usually, you are
interested in the lungs, and not if the rib cage is well aligned. However, the ribs are structures that
for example in CT can have a strong inuence on the similarity metric, especially if you use the SSD
metric. In that case, the rib cage may be well aligned at the cost of vessels structures near the border
of the lung with the rib cage. In this case it will help you if you use a dilated lung segmentation as a
mask.
Masks can be used both for the xed and the moving image. A xed image mask is sucient to focus
the registration on a ROI, since samples are drawn from the xed image. You only want to use a mask for
the moving image when your moving image contains nonsense grey values near the ROI.
In case you are using a mask to prevent bad karma from an articial edge, you also need to set the
parameter:
(ErodeMask "true")
If not, then when performing multi-resolution, information from the articial edge will ow into you ROI
due to the smoothing step. In case the edge around your ROI is meaningful, e.g. in the lung example, you
should set it to false, because this edge will help to guide the registration.
A common exception that elastix throws when drawing samples is: Could not nd enough image
samples within reasonable time. Probably the mask is too small. The probable cause for this is that your
xed image mask is too small. See the FAQ for more information.
29
5.5 Trouble shooting
5.5.1 Common errors
Some common sources of confusion and questions have been gathered in a FAQ, which can be found at
http://elastix.isi.uu.nl/FAQ.php
5.5.2 Bad initial alignment
When the initial alignment between two images is very o, you cannot start a nonrigid registration. And
sometimes it can be a hassle to get it right. What factors can help to get it right?
Start with a transformation with a low degree of freedom, i.e. the translation, rigid, similarity or ane
transform. Sometimes the images are really far o, and have no overlap to begin with (NB: the position
of images in physical space is determined by the origin and voxel spacing; see Section 2.2). A solution
is then to add the following line to your parameter le:
(AutomaticTransformInitialization "true")
which aligns the centres of the xed and moving image.
You need a good multi-resolution strategy, i.e. quite a bit of resolution levels. This way a lot of
smoothing is going on, blurring away all the details and thereby focussing the registration on the
major structures.
Use more iterations.
Take larger steps. Set the parameter a so high that the translation component of the transformation
takes a step of several voxels in each iteration, up to 10. Maybe it will work, and in that case you
will get alignment pretty quick, but the step size a is still large, so you immediately jump away from
alignment again. If that happens, you should let the sequence a
k
= a/(A + k)

decay relatively fast.

This can be achieved by setting both a and A a bit lower, see Figure 5.1.
In case you need to nd a large rotation, you might want to take larger steps for the rotation, but not
for the translation. This can be achieved by modifying the scales parameter, see Section 5.3.5:
(Scales 10000.0)
You can set it lower to take larger rotation steps. If it is set to 1.0 you take as large steps for
the rotation as for the translation (but the rotation is dened in radials). If you set it really high
(> 10
6
) you wont rotate at all. You probably dont need to go lower than 1000.0. Note that the
AutomaticScalesEstimation option usually works ne, so specifying Scales is not necessary.
5.5.3 Memory consumption
The typical size of clinical images increases as a function of time. Therefore, memory eciency will become
more of an issue. elastix consumes 100 MB of memory for small images, for larger image pairs (256
3
)
and some common components, consumption can be about 1 - 1.5 GB. With very large images (400
3
and
above) memory consumption can rise above the 2 GB limit of most Windows systems. What to do with
large images?
Buy yourself a brand new computer with a lot of memory. Make sure that this computer is 64 bit,
otherwise you cannot address this much memory. Also make sure that your operating system supports
64 bit.
30
Images in elastix are internally by default represented as a bunch of voxels of oating type. You can
modify this to short images:
(FixedInternalImagePixelType "short")
(MovingInternalImagePixelType "short")
This way you save half the amount of memory that is used to store the xed and moving images, and
their multi-resolution pyramids. This will come at the cost of a loss of precision, but may not be that
harmful. This option is useful both for elastix and transformix.
Change the interpolator that is used during registration. By default a B-spline interpolator is used,
which stores a coecient image internally in double type. You can also specify a oat version:
(Interpolator "BSplineInterpolatorFloat")
which saves you another bit of memory the size of a short image. This option is useful for elastix
only. To save even more memory, use the LinearInterpolator. This may change the results a little
though, because of its dierent method to compute image derivatives.
Change the interpolator that is used when resampling an image:
(ResampleInterpolator "FinalBSplineInterpolatorFloat")
This option is useful both for elastix and transformix. However, for elastix it will only save you
some memory at the very end of the registration.
Use downsampled images during the registration. This probably will not eect the registration accuracy
too much. After registration you can apply the resulting transformation to the original full size moving
image, using transformix. See the FAQ for more information.
31
Chapter 6
Advanced topics
6.1 Metrics
6.1.1 Image registration with multiple metrics and/or images
Up till now we viewed image registration as the problem of nding the spatial relation between one xed
image and one moving image, using one similarity metric to dene the t. Sometimes, it is desirable to
combine multiple metrics, or multiple xed and moving images, or both. All these three generalisations are
available in elastix:
multi-metric In this case the registration cost function is dened as:
C(T

; I
F
, I
M
) =
1

N
i=1

i
N

i=1

i
C
i
(T

; I
F
, I
M
), (6.1)
with
i
the weights. This way the same xed and moving image is used for every sub-metric C
i
. This
way one can for example simultaneously optimise the SSD and MI during a registration.
elastix should be called like:
elastix -f fixed.ext -m moving.ext -out outDir -p parameterFile.txt
multi-image In this case the registration cost function is dened as:
C(T

; I
F
, I
M
) =
1

N
i=1

i
N

i=1

i
C(T

; I
i
F
, I
i
M
). (6.2)
This way one can simultaneously register all channels of multi-spectral input data, using a single type
of cost function for all channels.
elastix should be called like:
elastix -f0 fixed0.ext -f1 fixed1.ext -f<>... -m0 moving0.ext -m1 moving1.ext
-m<>... -out outDir -p parameterFile.txt
both In this case the registration cost function is dened as:
C(T

; I
F
, I
M
) =
1

N
i=1

i
N

i=1

i
C
i
(T

; I
i
F
, I
i
M
). (6.3)
32
This is the most general way of registration supported by elastix. This will make it possible for
example to register two lung CT data sets with MI, while simultaneously registering the ssure seg-
mentations with the kappa statistic. The two may help each other in getting a better registration
compared to only using a single channel.
All three scenarios use the multi-metric registration method, which is selected in the parameter le with:
(Registration "MultiMetricMultiResolutionRegistration")
Other parts of the parameter le should look like:
(FixedImagePyramid "FixedSmoothingImagePyramid" "FixedSmoothingImagePyramid" ...)
(MovingImagePyramid "MovingSmoothingImagePyramid" "MovingSmoothingImagePyramid" ... )
(Interpolator "BSplineInterpolator" "BSplineInterpolator" ...)
(Metric "AdvancedMattesMutualInformation" "AdvancedMeanSquareDifference" ...)
(ImageSampler "RandomCoordinate" "RandomCoordinate" ...)
(Metric0Weight 0.125)
(Metric1Weight 0.125)
(Metric2Weight 0.125)
etc
Another way of registering multi-spectral data is to use the -mutual information measure, described
below.
6.1.2 -mutual information
The -mutual information metric computes true multi-channel -mutual information. It does not use high-
dimensional joint histograms, but instead relies on k-nearest neighbour graphs to estimate -MI. Details can
be found in Staring et al. [2009]. It is specied in the parameter le with:
(Registration "MultiResolutionRegistrationWithFeatures")
(FixedImagePyramid "FixedSmoothingImagePyramid" "FixedSmoothingImagePyramid")
(MovingImagePyramid "MovingSmoothingImagePyramid" "MovingSmoothingImagePyramid")
(Interpolator "BSplineInterpolator" "BSplineInterpolator")
(Metric "KNNGraphAlphaMutualInformation")
(ImageSampler "MultiInputRandomCoordinate")
// KNN specific
(Alpha 0.99)
(AvoidDivisionBy 0.0000000001)
(TreeType "KDTree")
(BucketSize 50)
(SplittingRule "ANN_KD_STD")
(ShrinkingRule "ANN_BD_SIMPLE")
(TreeSearchType "Standard")
(KNearestNeighbours 20)
(ErrorBound 10.0)
A complete list of the available parameters can be found in the doxygen documentation
elx::KNNGraphAlphaMutualInformationMetric.
6.1.3 Penalty terms
This paragraph requires extension and modication.
33
In order to regularise the transformation T

often a penalty term P() is added to the cost function, so

it becomes:
C =
1
S +
2
P, (6.4)
where
1
,
2
user-dened constants that weigh similarity against regularity.
Penalty term are often based on the rst or second order spatial derivatives of the transformation. An
example is the bending energy of the transformation, which is arguably the most common penalty term, see
Section 6.1.4.
The derivative of the similarity measure usually involves computation of the spatial derivative of the
moving image:
IM
x
, and the derivative of the transformation to its parameters:
T

. In the ITK the

last derivative is implemented using transform->GetJacobian(), i.e. the derivative to the transformation
parameters is referred to as Jacobian.
Penalty terms usually consist of the rst and second order spatial derivatives of the transformation,
i.e.
T
x
and

2
T
xx
T
. We will refer to these derivatives as the SpatialJacobian and the SpatialHessian
to clearly distinguish between these derivatives and the Jacobian. In order to apply the gradient descent
optimisation routine (2.20), (2.21), we additionally need the derivatives

T
x
and

2
T
xx
T
. These we call
the JacobianOfSpatialJacobian and JacobianOfSpatialHessian, respectively.
The transform class as dened in the ITK does not support the computation of spatial derivatives T/x
and
2
T/x
2
, and their derivatives to . Initially, we created non-generic classes that combine mutual infor-
mation and the mentioned penalty terms specically (the MattesMutualInformationWithRigidityPenalty
component in elastix version 4.3 and earlier). In 2010, however, we created a more advanced version of
the ITK transform that does implement these spatial derivatives. Additionally, we created a bending energy
regularisation class that takes advantage of these functions, see Section 6.1.4. We also reimplemented the
rigidity penalty term, see Section 6.1.5; it currently however does not yet use these spatial derivatives. More
detailed information can be found in Staring and Klein [2010a].
This all means that it is possible in elastix to combine any similarity metric with any of the available
penalty terms (currently the bending energy and the rigidity penalty term).
6.1.4 Bending energy penalty
The bending energy penalty term is dened in 2D as:
P
BE
() =
1
P

xi

2
T
xx
T
( x
i
)

2
F
(6.5)
=
1
P

xi
2

j=1

2
T
j
x
2
1
( x
i
)

2
+ 2

2
T
j
x
1
x
2
( x
i
)

2
+

2
T
j
x
2
2
( x
i
)

2
, (6.6)
where P is the number of points x
i
, and the tilde denotes the dierence between a variable and a given
point over which a term is evaluated. As you can see it penalises sharp deviations of the transformation
(e.g. no high compression followed by a nearby high expansion). You can use it to regularize your nonrigid
transformation if you experience problems such as foldings. In our current implementation the computation
time of this term is relatively large, though.
It can be selected in elastix using
(Metric "AnySimilarityMetric" "TransformBendingEnergy")
(Metric0Weight 1.0)
(Metric1Weight <weight>)
and has no further parameters.
34
6.1.5 Rigidity penalty
Some more advanced metrics, not found in the ITK, are available in elastix: The rigidity penalty term
P
rigid
(T

; I
M
) described in Staring et al. [2007a]. It is specied in the parameter le with:
(Metric "AnySimilarityMetric" "TransformRigidityPenalty")
// normal similarity metric parameters
...
// Rigidity penalty parameters:
(RigidityPenaltyWeight 0.1)
(LinearityConditionWeight 10.0)
(OrthonormalityConditionWeight 1.0)
(PropernessConditionWeight 100.0)
(MovingRigidityImageName "movingRigidityImage.mhd")
A complete list of the available parameters can be found in the doxygen documentation
elx::TransformRigidityPenalty. See also Section 6.1.3.
6.1.6 DisplacementMagnitudePenalty: inverting transformations
The DisplacementMagnitudePenalty is a cost function that penalises ||T

(x) x||
2
. You can use this
to invert transforms, by setting the transform to be inverted as an initial transform (using -t0), setting
(HowToCombineTransforms "Compose"), and running elastix with this metric. After that you can man-
ually set the initial transform in the last parameter le to "NoInitialTransform", and voila, you have the
inverse transform! Strictly speaking, you should then also change the Size/Spacing/Origin/Index settings to
match that of the moving image. Select it with:
(Metric "DisplacementMagnitudePenalty")
Note that inverting a transformation becomes conceptually very similar to performing an image registration
in this way. Consequently, the same choices are relevant: optimisation algorithm, multiresolution etc...
Note that this procedure was described and evaluated in Metz et al. [in press].
6.1.7 Corresponding points: help the registration
Most of the similarity measures in elastix are based on corresponding characteristics of the xed and
moving image. It is possible, however, to register based on point correspondence. Therefore, in elastix
4.4 we introduced a metric that minimizes the distance of two point sets with known correspondence. It is
dened as:
S
CP
=
1
P

xFi
x
Mi
T

(x
Fi
) (6.7)
where P is the number of points x
i
, and x
Fi
, x
Mi
corresponding points from the xed and moving image
point sets, respectively. The metric can be used to help in a dicult image registration task that fails if
performed fully automatically. A user can manually click corresponding points (or maybe automatically
extract), and setup elastix to not only minimize based on intensity, but also taking into account that some
positions are known to correspond. The derivative of S
CP
reads:

S
CP
=
1
P

xFi
1
x
Mi
T

(x
Fi
)
(x
Mi
T

(x
Fi
))
T

(x
Fi
). (6.8)
In elastix this metric can be selected using:
35
(Metric "AnySimilarityMetric" "CorrespondingPointsEuclideanDistancePointMetric")
(Metric0Weight 1.0)
(Metric1Weight <weight>)
Note that this metric must be specied as the last metric, due to some technical constraints. The xed and
moving point set can be specied on the command line:
elastix ... -fp fixedPointSet.txt -mp movingPointSet.txt
The point set les have to be dened in a specic format, identical to supplying points to transformix, see
Section 4.2.
6.1.8 VarianceOverLastDimensionMetric: aligning time series
This metric is explained in Metz et al. [in press]. Example parameter les can be found on the wiki parameter
le database, entry par0012.
This metric should be used to estimate the motion in dynamic imaging data (time series). The variance
of intensities over time is measured. Two- to four-dimensional imaging data is supported.
6.2 Image samplers
RandomSparseMask This variant of the random sampler is useful if the xed image mask is sparse (i.e.
consists of many zeros).
6.3 Interpolators
ReducedDimensionBSplineInterpolator This is a variant of the normal B-spline interpolator, which
uses a 0
th
order spline in the last dimension. This saves time when aligning time-series, when you do
not have to interpolate in the last (time) dimension anyway. Its usage is illustrated in entry par0012
of the parameter le database.
6.4 Transforms
DeformationFieldTransform This transform serves as a wrapper around existing deformation eld vector
images. It computes the transformation by interpolating the deformation eld image. The relevant
tags in the transform parameter le are as follows:
(Transform "DeformationFieldTransform")
(DeformationFieldFileName "deformationField.mhd")
(DeformationFieldInterpolationOrder 1)
(NumberOfParameters 0)
The deformation eld images pixel type should be a vector of float elements. It could be a deformation
eld that is the result of transformix -def all for example! Since this transform does not have any
parameters (the has zero length), it makes no sense to use it for registration. It can just be used as
an initial transformation (supplied by the option -t0) or as input for transformix.
SplineKernelTransform As an alternative to the B-spline transform, elastix includes a SplineKernelTransform,
which implements a thin-plate spline type of transform; see also Section 2.6. This transformation re-
quires a list of xed image landmarks (control points) to be specied, by means of an input points
le which has the same format as the -def le used by transformix (Section 4.2. See the doxygen
documentation on the website for a list of its parameters.
36
WeightedCombinationTransform This is a transformation that is modelled as a weighted combination
of user-specied transformations: T

(x) =

i
w
i
T
i
(x). The weights w
i
form the parameter vector
. The sub-transforms T
i
(x) may for example follow from a statistical deformation model, obtained
by principal component analysis. See the doxygen documentation on the website for a list of its
parameters.
BSplineTransformWithDiusion This transform implements the work described in Staring et al. [2007b].
BSplineStackTransform This transformation model denes a stack of independent B-spline transforma-
tions. Its usage is illustrated in Metz et al. [in press]. Example parameter les can be found on the
wiki parameter le database, entry par0012.
6.5 Optimisation methods
AdaptiveStochasticGradientDescent This optimizer is very similar to the StandardGradientDescent,
but implements an adaptive step size mechanism and estimates a proper initial value for SP a. See
Klein et al. [2009] for more details. In practice this optimizer works in many applications with its
default settings. Only the number of iterations must be specied by the user:
(Optimizer "AdaptiveStochasticGradientDescent")
(MaximumNumberOfIterations 500)
(SigmoidInitialTime 4.0)
(MaximumStepLength 1.0)
The last two options are not mandatory. SigmoidInitialTime corresponds to t
0
in the article; its
default value is 0. The MaximumStepLength corresponds to in the article; its default value equals the
average voxel spacing of xed and moving image.
Conjugate gradient ConjugateGradientFRPR
CMAEvolutionStrategy
FiniteDierenceGradientDescent
Full search
Quasi Newton
RegularStepGradientDescent RSGDEachParameterApart
SimultaneousPerturbation
37
Chapter 7
Developers guide
7.1 Relation to ITK
A large part of the elastix code is based on the ITK Iba nez et al. [2005]. The use of the ITK implies that
the low-level functionality (image classes, memory allocation etc.) is thoroughly tested. Naturally, all image
formats supported by the ITK are supported by elastix as well. The C++ source code can be compiled
on multiple operating systems (Windows XP, Linux, Mac OS X), using various compilers (MS Visual Studio
up to version 2010, GCC up to version 4.3), and supports both 32 and 64 bit systems.
In addition to the existing ITK image registration classes, elastix implements new functionality. The
most important enhancements are listed in Table 7.1.
A modular framework for sampling strategies. See for more details Staring and
Klein [2010b].
Several new optimisers: Kiefer-Wolfowitz, Robbins-Monro, adaptive stochastic
gradient descent, evolutionary strategy. Complete rework of existing ITK op-
timisers, adding more user control and better error handling: quasi-Newton,
nonlinear conjugate gradient.
Several new or more exible cost functions: (normalised) mutual information,
implemented with Parzen windowing similar to Thevenaz and Unser [2000], mul-
tifeature -mutual information, bending energy penalty term, rigidity penalty
term.
The ability to concatenate any number of geometric transformations.
The transformations support computation of not only T/, but also of spatial
derivatives T/x and
2
T/x
2
, and their derivatives to , frequently required
for the computation of regularisation terms. Additionally, the compact support of
certain transformations is integrated more generally. See for more details Staring
and Klein [2010a].
Linear combinations of cost functions, instead of just a single cost function.
Table 7.1: The most important enhancements and additions in elastix, compared to the ITK.
38
7.2 Overview of the elastix code
The elastix source code consists roughly of two layers, both written in C++: A) ITK-style classes that
implement image registration functionality, and B) elastix wrappers that take care of reading and setting
parameters, instantiating and connecting components, saving (intermediate) results, and similar adminis-
trative tasks. The modular design enables adding new components, without changing the elastix core.
Adding a new component starts by creating the layer A class, which can be compiled and tested independent
of layer B. Subsequently, a small layer B wrapper needs to be written, which connects the layer A class to
the other parts of elastix.
The image samplers, for example, are implemented as ITK classes that all inherit from a base class
itk::ImageSamplerBase. These can be found in src/Common/ImageSamplers. This is layer A in elastix.
For each sampler (random, grid, full... ) a wrapper is written, located in src/Components/ImageSamplers,
which takes care of conguring the sampler before each new resolution of the registration process. This is
layer B of elastix.
7.2.1 Directory structure
The basic directory structure is as follows:
dox
src/Common: ITK classes, Layer A stu. This directory also contains some external libraries, unrelated
to ITK, like xout (which is written by us) and the ANNlib.
src/Core: this is the main elastix kernel, responsible for the execution ow, connecting the classes,
reading parameters etc.
src/Components: this directory contains the components and their elastix wrappers (layer B). Very
component-specic layer A code can also be found here.
In elastix 4.4 and later versions, it is also possible to add your own Component directories. These can
be located anywhere outside the elastix source tree. See Section 7.3 for more details about this.
7.3 Creating new components
If you want to create your own component, it is natural to start writing the layer A class, without bothering
about elastix. The layer A lter should implement all basic functionality and you can test in a separate
ITK program if it does what it is supposed to do. Once you got this ITK class to work, it is trivial to write
the layer B wrapper in elastix (start by copy-pasting from existing components).
With CMake, you can tell elastix in which directories the source code of your new components is
located, using the ELASTIX USER COMPONENT DIRS option. elastix will search all subdirectories of these
directories for CMakeLists.txt les that contain the command ADD ELXCOMPONENT( <name> ... ). The
CMakeLists.txt le that accompanies an elastix component looks typically like this:
ADD_ELXCOMPONENT( AdvancedMeanSquaresMetric
elxAdvancedMeanSquaresMetric.h
elxAdvancedMeanSquaresMetric.hxx
elxAdvancedMeanSquaresMetric.cxx
itkAdvancedMeanSquaresImageToImageMetric.h
itkAdvancedMeanSquaresImageToImageMetric.hxx )
39
The ADD ELXCOMPONENT command is a macro dened in src/Components/CMakeLists.txt. The rst argu-
ment is the name of the layer B wrapper class, which is declared in elxAdvancedMeanSquaresMetric.h.
After that, you can specify the source les on which the component relies. In the example above, the les
that start with itk form the layer A code. Files that start with elx are the layer B code. The le
elxAdvancedMeanSquaresMetric.cxx is particularly simple. It just consists of two lines:
#include "elxAdvancedMeanSquaresMetric.h"
elxInstallMacro( AdvancedMeanSquaresMetric );
The elxInstallMacro is dened in src/Core/Install/elxMacro.h.
The les elxAdvancedMeanSquaresMetric.h/hxx together dene the layer B wrapper class. That class
inherits from the corresponding layer A class, but also from an elx::BaseComponent. This gives us the
opportunity to add a common interface to all elastix components, regardless of the ITK classes from which
they inherit. Examples of this interface are the following methods:
void BeforeAll(void)
void BeforeRegistration(void)
void BeforeEachResolution(void)
void AfterEachResolution(void)
void AfterEachIteration(void)
void AfterRegistration(void)
These methods are automatically invoked at the moments indicated by the name of the function. This gives
you a chance to read/set some parameters, print some output, save some results etc.
40
Appendix A
Example parameter le
//ImageTypes
(FixedInternalImagePixelType "float")
(FixedImageDimension 2)
(MovingInternalImagePixelType "float")
(MovingImageDimension 2)
//Components
(Registration "MultiResolutionRegistration")
(FixedImagePyramid "FixedRecursiveImagePyramid")
(MovingImagePyramid "MovingRecursiveImagePyramid")
(Interpolator "BSplineInterpolator")
(Metric "AdvancedMattesMutualInformation")
(Optimizer "StandardGradientDescent")
(ResampleInterpolator "FinalBSplineInterpolator")
(Resampler "DefaultResampler")
(Transform "EulerTransform")
// ********** Pyramid
// Total number of resolutions
(NumberOfResolutions 3)
// ********** Transform
//(CenterOfRotation 128 128) center by default
(AutomaticTransformInitialization "true")
(AutomaticScalesEstimation "true")
(HowToCombineTransforms "Compose")
// ********** Optimizer
// Maximum number of iterations in each resolution level:
(MaximumNumberOfIterations 300 300 600)
//SP: Param_a in each resolution level. a_k = a/(A+k+1)^alpha
(SP_a 0.001)
41
//SP: Param_alpha in each resolution level. a_k = a/(A+k+1)^alpha
(SP_alpha 0.602)
//SP: Param_A in each resolution level. a_k = a/(A+k+1)^alpha
(SP_A 50.0)
// ********** Metric
//Number of grey level bins in each resolution level:
(NumberOfHistogramBins 32)
(FixedKernelBSplineOrder 1)
(MovingKernelBSplineOrder 3)
// ********** Several
(WriteTransformParametersEachIteration "false")
(WriteTransformParametersEachResolution "false")
(ShowExactMetricValue "false")
(ErodeMask "true")
// ********** ImageSampler
// Number of spatial samples used to compute the
// mutual information in each resolution level:
(ImageSampler "RandomCoordinate")
(NumberOfSpatialSamples 2048)
(NewSamplesEveryIteration "true")
// ********** Interpolator and Resampler
//Order of B-Spline interpolation used in each resolution level:
(BSplineInterpolationOrder 1)
//Order of B-Spline interpolation used for applying the final deformation:
(FinalBSplineInterpolationOrder 3)
//Default pixel value for pixels that come from outside the picture:
(DefaultPixelValue 0)
42
Appendix B
Example transform parameter le
(Transform "EulerTransform")
(NumberOfParameters 3)
(TransformParameters -0.000000 -4.564513 -2.091174)
(InitialTransformParametersFileName "NoInitialTransform")
(HowToCombineTransforms "Compose")
// Image specific
(FixedImageDimension 2)
(MovingImageDimension 2)
(FixedInternalImagePixelType "float")
(MovingInternalImagePixelType "float")
(Size 256 256)
(Index 0 0)
(Spacing 1.0000000000 1.0000000000)
(Origin 0.0000000000 0.0000000000)
// EulerTransform specific
(CenterOfRotationPoint 128.0000000000 128.0000000000)
// ResampleInterpolator specific
(ResampleInterpolator "FinalBSplineInterpolator")
(FinalBSplineInterpolationOrder 3)
// Resampler specific
(Resampler "DefaultResampler")
(DefaultPixelValue 0.000000)
(ResultImageFormat "mhd")
(ResultImagePixelType "short")
43
Appendix C
Software License
Overview:
Elastix was developed by Stefan Klein and Marius Staring under
supervision of Josien P.W. Pluim, initially under contract to the
Image Sciences Institute, University Medical Center Utrecht, The
Netherlands.
Elastix is distributed under the new and simplified BSD license
approved by the Open Source Initiative (OSI)
[http://www.opensource.org/licenses/bsd-license.php].
The software is partially derived from the Insight Segmentation and
Registration Toolkit (ITK), which is also distributed under the new
and simplified BSD licence. The ITK is required by Elastix for
compilation of the source code.
The copyright of the files in the Common/KNN/ann_1.1 subdirectory
is held by a third party, the University of Maryland. The ANN
package is distributed under the GNU Lesser Public Licence. Please
read the content of the subdirectory for specific details on this
third-party license.
Elastix Copyright Notice:
Copyright (c) 2004-2010 University Medical Center Utrecht
All rights reserved.
License:
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
* Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
44
* Redistributions in binary form must reproduce the above copyright notice,
this list of conditions and the following disclaimer in the documentation
and/or other materials provided with the distribution.
* Neither the name of the University Medical Center Utrecht nor the names of
its contributors may be used to endorse or promote products derived from
this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND
CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES,
INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS
BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR
TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF
THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
SUCH DAMAGE.
45
Bibliography
R. Brooks and T. Arbel. Improvements to the itk::KernelTransform and subclasses. The Insight Journal,
January - June, 2007. URL http://hdl.handle.net/1926/494.
W. R. Crum, O. Camara, and D. L. G. Hill. Generalized overlap measures for evaluation and validation in
medical image analysis. IEEE Trans. Med. Imag., 25(11):14511461, 2006.
M. H. Davis, A. Khotanzad, D. P. Flamig, and S. E. Harms. A physics-based coordinate transformation for
3-D image matching. IEEE Trans. Med. Imag.
B. Fischer and J. Modersitzki. A unied approach to fast image registration and a new curvature based
registration technique. Linear Algebra Appl., 380:107 124, 2004.
Joseph V Hajnal, Derek L G Hill, and David J Hawkes, editors. Medical Image Registration. CRC Press,
2001. ISBN 0849300649.
D. L. G. Hill, P. G. Batchelor, M. Holden, and D. J. Hawkes. Medical image registration. Phys. Med. Biol.,
46(3):R1 R45, 2001.
L. Iba nez, W. Schroeder, L. Ng, and J. Cates. The ITK Software Guide. Kitware, Inc. ISBN 1-930934-15-7,
second edition, 2005.
S. Klein, M. Staring, and J. P. W. Pluim. Evaluation of optimisation methods for nonrigid medical image
registration using mutual information and B-splines. IEEE Trans. Image Process., 16(12):2879 2890,
December 2007.
S. Klein, U. A. van der Heide, I. M. Lips, M. van Vulpen, M. Staring, and J. P. W. Pluim. Automatic
segmentation of the prostate in 3D MR images by atlas matching using localized mutual information.
Math. Program., 35(4):1407 1417, April 2008.
S. Klein, J. P. W. Pluim, M. Staring, and M.A. Viergever. Adaptive stochastic gradient descent optimisation
for image registration. International Journal of Computer Vision, 81(3):227 239, March 2009.
H. Lester and S. R. Arridge. A survey of hierarchical non-linear medical image registration. Pattern Recognit.,
32(1):129 149, 1999.
F. Maes, A. Collignon, D. Vandermeulen, G. Marchal, and P. Suetens. Multimodality image registration by
maximization of mutual information. IEEE Trans. Med. Imag., 16(2):187 198, 1997.
J. B. A. Maintz and M. A. Viergever. A survey of medical image registration. Med. Image Anal., 2(1):1
36, 1998.
D. Mattes, D. R. Haynor, H. Vesselle, T. K. Lewellen, and W. Eubank. PET-CT image registration in the
chest using free-form deformations. IEEE Trans. Med. Imag., 22(1):120 128, 2003.
46
C.T. Metz, S. Klein, M. Schaap, T. van Walsum, and W.J. Niessen. Nonrigid registration of dynamic medical
imaging data using nD+t B-splines and a groupwise optimization approach. Medical Image Analysis, in
press.
J. Modersitzki. Numerical Methods for Image Registration. Oxford University Press, 2004. ISBN 978-
0198528418.
K. Murphy, B. van Ginneken, J.P.W. Pluim, S. Klein, and M. Staring. Semi-automatic reference stan-
dard construction for quantitative evaluation of lung CT registration. In Medical Image Computing and
Computer-Assisted Intervention (MICCAI), volume 5242 of Lecture Notes in Computer Science, pages
1006 1013, 2008.
J. Nocedal and S. J. Wright. Numerical optimization. Springer-Verlag, New York, 1999. ISBN 0-387-98793-2.
T. Rohlng, R. Brandt, R. Menzel, and C. R. Maurer Jr. Evaluation of atlas selection strategies for atlas-
based image segmentation with application to confocal microscopy images of bee brains. NeuroImage, 21
(4):14281442, 2004.
D. Rueckert, L. I. Sonoda, C. Hayes, D. L. G. Hill, M. O. Leach, and D. J. Hawkes. Nonrigid registration
using free-form deformations: Application to breast MR images. IEEE Trans. Med. Imag., 18(8):712
721, 1999.
M. Staring and S. Klein. itk::transforms supporting spatial derivatives. The Insight Journal, 2010a. URL
http://hdl.handle.net/10380/3215.
M. Staring and S. Klein. An image sampling framework for the itk. The Insight Journal, 2010b. URL
http://hdl.handle.net/10380/3190.
M. Staring, S. Klein, and J. P. W. Pluim. A rigidity penalty term for nonrigid registration. Medical Physics,
34(11):4098 4108, 2007a.
M. Staring, S. Klein, and J. P. W. Pluim. Nonrigid registration with tissue-dependent ltering of the
deformation eld. Physics in Medicine and Biology, 52(23):6879 6892, 2007b.
M. Staring, U. A. van der Heide, S. Klein, M. A. Viergever, and J. P. W. Pluim. Registration of cervical
MRI using multifeature mutual information. IEEE Transactions on Medical Imaging, 28(9):14121421,
2009. In press.
C. Studholme, D. L. G. Hill, and D. J. Hawkes. An overlap invariant entropy measure of 3D medical image
alignment. Pattern Recognit., 32:7186, 1999.
P. Thevenaz and M. Unser. Halton sampling for image registration based on mutual information. Sampling
Theory in Signal and Image Processing, 7(2):141171, 2008.
P. Thevenaz and M. Unser. Optimization of mutual information for multiresolution image registration. IEEE
Trans. Image Process., 9(12):2083 2099, 2000.
M. Unser. Splines: A perfect t for signal and image processing. IEEE Signal Process. Mag., 16(6):22 38,
1999.
Paul Viola and William M. Wells III. Alignment by maximization of mutual information. Int. J. Comput.
Vis., 24(2):137 154, 1997.
B. Zitova and J. Flusser. Image registration methods: a survey. Image Vis. Comput., 21(11):9771000, 2003.
47