AUTOMATED MULTI-MODALITY IMAGE REGISTRATION

BASED ON INFORMATION THEORY

A. COLLIGNON, F. MAES, D. DELAERE, D. VANDERMEULEN,

P. SUETENS AND G. MARCHAL
Laboratory for Medical Imaging Research, ESAT-Radiology
K. U. Leuven, Kardinaal Mercierlaan 94, 3001 Heverlee, and
U. Z. Gasthuisberg, Radiology, Herestraat 49, 3000 Leuven,
Belgium

Abstract. We propose an information theoretic approach to the rigid body

registration of 3D multi-modality medical image data. The "mutual infor-
mation" of grey-value pairs is proposed as a new matching criterion. The
new voxel similarity based (VB) registration algorithm is applied to real
world images of the human head (CT, MR, and PET volumes), and com-
pared to stereotactic and 2 other voxel based registration solutions. Our
results show that subvoxel accuracy can be obtained completely automati-
cally and without any pre-segmentation.

1. Introduction
Rigid body registration is a fundamental task in three-dimensional (3-D)
medical image processing that serves to align two or more 3-D scans that
were taken at a di erent time, or generated on di erent medical imaging
modalities. There are numerous applications. Medical diagnosis bene ts
from the complementarity of the information in images of di erent modal-
ities. During radiation therapy planning dose calculation needs to be done
based on the CT data while tumor outlining is often much better performed
in the corresponding MR scan. The analysis of brain function requires the
correlation of physiological data from PET data with anatomical data from
the corresponding MR data, etc.
A general overview of image registration techniques has been compiled by
Brown [1]. Two other articles survey image registration techniques and their
264 A. COLLIGNON ET AL.
applications in medical imaging [9], [13]. From these surveys we learn that
the bulk of registration algorithms are either frame based, point landmark
based (using external and/or internal landmarks), surface based, or voxel
based. Point landmark and surface based registration algorithms are typ-
ically used for the alignment of anatomical scans, while voxel similarity
based registration algorithms seem to be more robust for alignment prob-
lems involving functional scans ([7] and [15]), although many people are
using surface based registration algorithms in the latter case also.
Voxel based (VB) registration algorithms optimize a function measuring
the similarity of all possible pairs of voxel grey-values (or derived features,
e.g. gradients). The main advantage of VB methods is that feature calcula-
tion is straightforward or even absent when only grey-values are used and
thus, in theory, accuracy is not limited by segmentation errors as in surface
based algorithms. Simple correlation of grey-values works only for di erent
modalities if there exists a linear mapping between grey-values [1], which is
unfortunately not always the case. Nonetheless, interest in VB registration
algorithms has recently revived by the success of Woods' AIR package in
clinical practice [16], which has been made more exible by Hill et al [7].
Their basic assumption is that the grey-value mapping should be a func-
tion (e.g. to each MR grey-value correspondends at most one PET-value).
They accomplish this heuristically by demanding that a global measure of
the variance of the grey-value ratios, with the function's domain variable
as the ratios' denominator, be minimal. We recently compared these and
some other VB matching criteria [2]. All of them are heuristic in nature.
In the same work we rst suggested an information theoretic approach to
the design of a non-heuristic multi-modality matching criterion and we pro-
posed to use the entropy of the joint probability distribution p(g1 ; g2 ) of
the combined grey-values of all common voxel pairs in the two images to
be matched:
X
E (g1 ; g2 ) = , p(g1 ; g2 ) log (p(g1 ; g2 )) (1)
g1 ;g2

In this article we use a basic theorem of statistical inference theory to de-

rive a new version of our entropy based VB matching criterion: the mutual
information or relative entropy of the joint probability distribution of the
grey-values of two (or more) scans. This criterion is more robust with re-
spect to the partial overlap problem and has some interesting mathematical
properties as well. Furthermore it is based on the original grey-values only
and requires no segmentation at all! Therefore, our new VB algorithm is
more robust than any of the surface based registration algorithms and than
any of the heuristic VB registration algorithms. We show results obtained
 AUTOMATED MULTI-MODALITY IMAGE REGISTRATION 265
with real world data in section 3. Those results are compared with that
of stereotactic and 2 other VB registration methods. First, however, we
will present the theory and discuss the implementation issues in detail.
In the discussion of the results we discuss robustness and accuracy of our
approach, and compare it to other VB algorithms.
2. Method
2.1. THEORY
In the sequel we restrict the geometric transformations to rigid body trans-
formations T ( consists of 3 translations and 3 rotations). The rigid body
assumption is well satis ed inside the skull in 3-D scans of the head, if
abstraction is made of scanner calibration problems (that can lead to inac-
curately known voxel sizes) and problems of geometrical distortions (e.g. in
MRI), both of which can be dealt with by careful calibration and scan pa-
rameter selection respectively. The applicability of the mutual information
matching criterion is not restricted to rigid body registration, however.
The selection of the new VB matching criterion (like in Woods [15], Hill [8],
and Collignon [2]) is based on the observation that the two-dimensional
grey-value histogram or scatter plot of the common voxel pairs of a pair
of registered images will be dispersed by misregistration as a function of
j  , j, where  is the registration solution.
We will now present a new measure of scatter-plot dispersion that also
has an interesting mathematical interpretation: relative entropy or mutual
information of the scatter-plot. Intuitively, we expect the dispersion to be
minimal when the interdependence of the grey-values of both images to
be registered is maximal. Two signals are maximally interdependent when
they are related by a one-to-one mapping, i.e. when g1 is a function of g2
and vice versa. This is seldom so for multi-modality data. As it happens,
the mathematical de nition of mutual information [3]:
X  p(x; y) 
S (X ; Y ) = p(x; y) log p(x)p(y) (2)
x;y

also known as Shannon information of two random variables X and Y with

marginal and joint probability distributions p(x), p(y) and p(x; y) resp.,
satis es the following properties (for a proof see [11]):
1. Symmetry: S (X ; Y ) = S (Y ; X )
2. S (X ; Y ) = 0 , (X and Y are independent)
3. Non-negative and bounded: 0  S (X ; Y )  min(S (X ; X ); S (Y ; Y ))
4. Invariance: (T is a one-to-one mapping) ) S (X ; T (X )) = S (X ; X )
266 A. COLLIGNON ET AL.
Interpretation of these properties con rms our intuition. Mutual infor-
mation measures the degree of interdependence, which is bounded below
by complete independence and bounded above by one-to-one mappings. We
conclude that a mathematically sound matching criterion can be obtained
by substitution of X and Y for the grey-values g1 and g2 of the images to
be registered in (2). Moreover, it allows for much more general grey-value
relationships than do the matching criteria proposed by Woods and Hill et
al or than any of the other currently available heuristic ones (see [2] for an
overview).
2.2. IMPLEMENTATION
The algorithm can be summarized by the following equation:
 !
 = arg min , X p(f1 (T s); f2 (s)) log p(f1 (T s); f2 (s)) (3)
s
2
p(f1 (T s)) p(f2(s))
where s is a sample of coordinates inside the volume of feature image f2 ,
f1(T s) is the set of grey-values from feature image f1 sampled at the rigidly
transformed coordinates T s, and p(f1 ; f2 ) is obtained by normalizing the
2-D scatter-plot, H (f1 ; f2 ), of the grey-values of that part of sample s that
falls in the common volume of f1 and f2 , into a joint probability distribution
as follows:
p(f ; f ) = H (f1 ; f2 )
1 2
N (4)
where N is the number of coordinates in the overlapping part of sample s,
(gi ,gimin )
and where the features fi = ni are rescaled versions of the original
grey-values. ni are the rescaling factors. Due to the rigid body assump-
tion we have restricted T to a superposition of a 3-D rotation and a 3-D
translation: T s = R s + t .
2.2.1. Sample selection
Feature values for image f2 are known for a nite number of discrete voxel
positions only. We need to decide how many and which voxel grey-values
to include in our calculation of the mutual information matching criterion.
Either the coordinates of all voxels of f2 , or a subset, or a superset thereof
are used in the calculation of the mutual information for every considered
by the optimisation algorithm. In case a superset of the voxel coordinates
is used, we need to select an interpolation method to nd the grey-values at
coordinates that do not coincide with the voxel coordinates of f2 . In order to
keep the speed performance as high as possible we have simply used nearest
neighbour interpolation to obtain samples from f2 . The aim of sub-sampling
is to increase speed performance. It should be noted that some continuity
 AUTOMATED MULTI-MODALITY IMAGE REGISTRATION 267
of grey-values is assumed if we expect the mutual information to be well-
behaved in case of sub-sampling. The grey-value continuity assumption
implies that p(f1 ( )) will vary smoothly with . This additional assumption
is not a restriction as long as sub-sampling factors (de ned as the fraction
of samples taken per voxel size, in a certain direction) do not become too
small. The aim of super-sampling is to increase accuracy. This can only work
if the same continuity assumption is valid in both images, or equivalently
if a meaningful interpolation method can be de ned in both images.
2.2.2. Interpolation
Due to the continuity of the registration parameters , regardless of whether
the sample s is a sub-, or a super-set of all voxels in f2 , the transformed
sample coordinates will always fall in between voxel coordinates of f1 , and
thus interpolation is needed to obtain values from f1 . An interpolation
method other than nearest neighbour interpolation is required to be able
to guarantee subvoxel accuracy. The next best choice in speed is trilinear
interpolation. The problem of trilinear or for that matter any other classi-
cal interpolation method is that they may introduce new grey-values, i.e.
rescaled grey-values not present in the set of the original grey-values of
f1. Thus, any classical interpolation method other than nearest neighbour
interpolation may introduce unpredictable changes into p(f1 ) as a function
of changes in . To avoid this problem we propose to use trilinear partial
volume distribution which is dual to trilinear interpolation. It uses the same
interpolation weights but instead of using them to average grey-values from
f1, they are used to distribute the voxel volume of the grey-value from f2
in the scatter-plot over the grey-values of all nearest neighbours in f1 .
2.2.3. Probability Estimation
The evaluation of (2) requires the knowledge of grey-value probabilities
which can be estimated from the grey-value scatter-plot. If quantisation
of the scatter plot is too ne, application of parzen-windowing [5] can be
used to reduce the e ect of image noise and to increase the reliability of the
probability estimates. However, parzen-windowing is rather expensive from
a computational point of view. Clearly some compression of the data will
bene t both the memory requirements and the quality of the probability
distribution estimates. It is expected that small grey-value distortions of
the observations do not induce large deviations in the registration solution.
Based on these observations we propose to use simple rescalings of the im-
ages, f1 and f2 , instead of the original grey-values, g1 and g2 . For practical
purposes it should be noted that rescaling is equivalent to binning of the
grey-value histogram, which can be considered to be a crude approximation
of parzen-windowing.
268 A. COLLIGNON ET AL.
2.2.4. Search Strategy
In the search for the optimal value of we have taken a pragmatic ap-
proach: 1) we assume we are able to select a meaningful initial value inside
the search engine's attraction pool of the optimal value, 2) we then use a
locally converging optimisation algorithm as the search engine. The imple-
mentation details are as follows: 1) Put the origins of the physical world
coordinate systems in the image centres and assign zero values to those
registration parameters that represent the relative translations of these
coordinate systems as initial estimates. This is meaningful under the as-
sumption that both images represent approximately the same volume of
interest. 2) Apply rotations that are multiples of 90o so that the images
are approximately identically oriented. This can be automated based on
knowledge about the scanner and the scanning sequence. 3) We have used
Powell's locally converging optimisation algorithm [10] for maximization of
the mutual information matching criterion. The rst 2 to 3 iterations may
be performed using nearest neighbour interpolation for its speed, while later
iterations may be performed using more complex interpolation for more ac-
curacy. 4) For the same reasons we also suggest to use a subsample of the
oating image during the rst iterations, and a super-sample during the
nal iterations.
3. Results
3.1. AN ARTIFICIAL TEST USING A REAL MR IMAGE
We transformed a T1-weighted MR image (from the Heidelberg reference
data set, see [4] for a detailed description) over (10.39, -3.27, 2.03) degrees
and over (7.06, 1.21,18.26) mm using trilinear resampling. We then used
this transformed version MRT as a reference image to be registered with
the original MR image. Eventhough large parts of some slices (slices 21
to 29 of 30) of the reference image have been set to zero in MRT due to
the transformation, the images could be registered. Accuracy results are
summarized in table 1 for three di erent interpolation methods.
3.2. COMPARISON WITH STEREOTACTIC REGISTRATION:
CT/MRI/GD-ENHANCED MRI
The data set used in the previous test also contains corresponding CT
and Gd-enhanced MR (MRE) images. All the images were acquired under
stereotactic conditions. In this section we summarize our registration re-
sults relative to the stereotactic registration solution. Note that since MR
and MRE were registered by acquisition we used MRT instead of MR for
matching with the MRE image. This test is similar to, but much more re-
 AUTOMATED MULTI-MODALITY IMAGE REGISTRATION 269

TABLE 1. This table gives the registration error for our algorithm when
applied to an MR and its transformed version MRT . Transformation pa-
rameters were (10.39, -3.27, 2.03) degrees and (7.06, 1.21,18.26) mm.
Voxels are (1.33 x 1.33 x 4.0) mm3 in both images. Subsample \3.0/1.0"
means that only 1 in 3 voxels are sampled in X and Y direction of the
image planes, and that all voxels are used in the perpendicular Z di-
rection. The \error"-column gives us an arithmetic average registration
error for all voxels in the central quarter of all slicesT of the MR volume.
Binning parameter values were: ni = 16. For MR : gmax = 2550 and
gmin = 0. For MR: gmax = 3359 and gmin = 2.

f1 f2Interpolation sub-sampling factors error (mm)

MRT MR nearest neighb. 3.0/1.0 0.17
MRT MR trilinear 3.0/1.0 0.0048
MRT MR partial volume 1.0/1.0 0.17

alistic than the one in the previous section, because: 1) the noise will be
di erent now, 2) because of the Gd-enhancement, and 3) due to the visi-
bility of pathology (a tumor) in the MRE volume. A true multi-modality
test is of course the registration of CT with MR. Table 2 gives the results.

TABLE 2. Same as table 1, but for MRT combined with MRE. For MRE:
gmax = 3582 and gmin = 2. For CT: gmax = 4094 and gmin = 2.

f1 f2 Interpolation sub-sampling factors error (mm)

MRT MRE nearest neighb. 3.0/1.0 0.51
MRT MRE trilinear 3.0/1.0 0.54
MRT MRE partial volume 1.0/1.0 0.40
CT MR partial volume 1.1/0.3 2.50

3.3. COMPARISON WITH OTHER VB REGISTRATION ALGORITHMS

USING 3-D MRI/CT
We have also compared the outcome of our algorithm with that of two other
VB registration algorithms by van den Elsen ([12], [14]) on a CT/MR pair.
In both cases her registration was performed automatically, without user
interaction or use of ducial markers. In [12] hierarchical correlation tech-
niques were applied to \ridge"-feature images, obtained from the original
data using geometrical operators in scale space. In [14] grey-value correla-
tion is proposed between a grey-value mapping of the CT and the original
270 A. COLLIGNON ET AL.
MR volumes. The di erence between the solutions found are summarized
in table 3. By visual inspection it was impossible to see the di erences be-
tween our solution and the grey-value correlation based method of van den
Elsen unless extremely large zooming factors were used. Fig. 1 visualizes
our result.

TABLE 3. This table gives registration di erences between our mutual

information based registration algorithm and both VB algorithms pro-
posed by van den Elsen et al. Voxels are (0.9765625 by 0.9765625 by 1.0)
mm3 for the MRI and (0.9375 by 0.9375 by 1.55) mm3 for the CT volume.
As before the average error
avg was calculated over the central quarter
of all planes of f2 , i.e. of the CT-volume.
t (c) is the distance between
the transformed coordinates of the volume centroid of f2 .
 is the an-
gle between the unit eigenvectors corresponding to the unit eigenvalue of
the rotation matrix. The solutions using nearest neighbour interpolation
were found in less than 7 minutes and sub-sampling (1 voxels in 4, in all
directions) while the solutions using partial volume distribution and no
sub-, nor super-sampling were found in approximately 4 hours on an IBM
RS/6000 workstation.
reference interpolation
avg (mm)
t (c) (mm)
 (deg)
[12] nearest neighbour 2.72 2.2398 3.6776
partial volume 2.02 1.6245 0.8957
[14] nearest neighbour 1.62 0.8563 3.8177
partial volume 1.26 0.0582 0.8674

3.4. TESTS WITH MRI/PET

The algorithm has been applied to MRI and PET data of Multiple Sclerosis
patients for improved diagnosis, and to MRI and PET data of volunteers. In
the latter case the registration serves to enhance the accuracy of a functional
analysis of the human brain that has been performed earlier [6] using PET
data alone. Due to the relatively low resolution of the PET images, even
full sample application of our algorithm is fast (approximately 6 minutes
for the data shown in g. 1).
4. Discussion
There never exists a perfect one-to-one grey-value mapping between im-
ages to be registered, even in the registered position, due to noise and often
also due to the nature of the data involved in multi-modality registration,
e.g. skull and background have overlapping intensities in MR while they
are completely di erent in CT. The algorithm proposed by Woods is not
 AUTOMATED MULTI-MODALITY IMAGE REGISTRATION 271

Figure 1. Left: MRI/CT combination registered using mutual information with partial
volume distribution, ni = 16, sampling factors equal to 1. Right: PET/MRI combination
of an MS patient. The exact registration solution is unknown. This result was found to
be clinically acceptable.

robust enough to handle this type of deviations from the one-to-one map-
ping assumption. Therefore, his algorithm requires prior segmentation of
the brain from the MR image before application of his matching criterion to
that part of the image. After careful inspection of the dispersion behaviour
of the grey-value scatter plot Hill et al propose to select only speci c bins
of the scatter plot to obtain meaningful dispersion measurements. Their
approach can be automated but in a data dependent manner only. Our
rst attempt in [2] to increase robustness in a data independent way was
based on an intuitive information theoretic reasoning only: the scatter-plot
entropy as a measure of multi-modality information content or signal com-
plexity has all the properties of a data independent measure of scatter-plot
dispersion. The main problem with the entropy matching criterion, how-
ever, is its sensitivity to the problem of partial overlap, and, even worse,
also to the problem that the information content of the separate overlap-
ping parts of the images is not constant, causing it to have multiple local
272 A. COLLIGNON ET AL.
minima all over -space. In this paper we have solved the latter problem
also using relative entropy.

The results from sections 3.1 and 3.2 are all subvoxel accurate if the largest
voxel size (mostly equals the slice distance) is taken as a reference, and if
the stereotactic registration solutions are taken to be correct. The latter
assumption was clearly not true for the pair of CT/MR volumes in section
3.2. In this case visual inspection showed that our mutual information based
registration solution is better than the stereotactic one, which was particu-
larly noticable in the caudal-cranial direction. This can be explained by the
fact that the stereotactic localiser was xed to the head by means of a mask
and not by screws. Taking into consideration the nature of the data in these
experiments we may also conclude that the mutual information is inherently
robust against noise and local di erences (e.g. pathological). Comparison
of the results for the di erent interpolation methods tells us that nearest
neighbour interpolation o ers a very rapid solution with an acceptable,
near subvoxel accuracy, even when it is used with a subsample of the oat-
ing volume, f2 . All experiments with nearest neighbour interpolation and
sub-sampling were performed in less than 7 minutes. The trilinear partial
volume distribution method was especially designed to overcome some edge
e ects, and e ects of image grid interference, that are both inevitable when
calculating the mutual information of a nite sample. Therefore, it does not
go well with sub-sampling, which explains why no sub-sampling was used in
tables 1 and 2. When the same tests were performed using the sampling fac-
tors that were used with nearest neighbour and trilinear interpolation the
search algorithm got stuck in a local minimum far from the registration so-
lution. Our implementation of the trilinear partial volume distribution has
approximately the same complexity as that of trilinear interpolation which
was about twice as slow as nearest neighbour interpolation. This factor two
together with the di erences in the in-plane sampling factor explains why
the partial volume distribution based experiments in tables 1 and 2 took 90
minutes. The results from section 3.3 show us that it is possible to register
images of di erent modalities without using any prior knowledge about the
grey-value mapping between parts of the image, and yet obtain essentially
the same registration results. E.g. both methods proposed by van den Elsen
require the application of feature extraction algorithms to make simple cor-
relation work. This approach makes her algorithms data dependent. The
results obtained in section 3.4 have not yet been compared to that of other
algorithms. However, our algorithm is more user-friendly than Woods' AIR
software [16] because it does not require the prior segmentation of the brain
surface from the MR images. Our algorithm does not require the selection
of a part of the histogram, as does Hill's more robust version of Woods'
 AUTOMATED MULTI-MODALITY IMAGE REGISTRATION 273
algorithm.
In all our experiments so far, using the search procedure described in
section 2.2.4 we have observed translations up to 2:5 cm and rotations up
to 11o that were all recovered fully automatically. Thus, for our practical
purposes the mutual information criterion o ers a well behaved optimi-
sation function in a reasonably large region around the initial estimate
of the registration parameter values. In theory we need to prove that it is
strictly quasi-convex to guarantee its unimodality. Further experiments will
be needed in the future to obtain an estimate of the range of quasi-convexity
around the global optimum.

5. Conclusion
The new voxel similarity based (VB) registration algorithm has been ap-
plied successfully to real world images of the human head (CT, MR, and
PET volumes) which satisfy the rigid body assumption well. Comparison
of the registration outcome with that of stereotactic and 2 other VB regis-
tration solutions indicates that mutual information is a subvoxel accurate
matching criterion and that it is robust in multi-modality situations.
The experimental results indicate that the mutual information matching
criterion is a well-behaved multi-modality 3-D registration criterion, and
that we have found satisfactory solutions for all implementation problems.
Most importantly, we have found that due to the niteness of the sample
of grey-value pairs involved in the calculation of the mutual information
matching criterion the use of partial volume distribution of grey-values is
required to obtain the most accurate results.

6. Acknowledgments
This work is part of COVIRA (Computer Vision in Radiology), project
A2003 of the AIM (Advanced Informatics in Medicine) programme of the
European Commission. Special thanks go to prof. C. Schiepers, and dr. P.
Dupont from the PET Center of the Department of Nuclear Medicine, at
the University Hospital Gasthuisberg of Leuven for supplying us with the
PET/MRI data and for visually inspecting our results. Special thanks also
go to Petra van den Elsen for our informal cooperation. Her volumetric
MRI was acquired in the Hospital \Medisch Spectrum" in Enschede, with
special thanks to dr. G. Wilts. The high resolution CT was acquired in
Utrecht University Hospital by prof. dr. F.W. Zonneveld, Drs. D.N. Velis,
and P.A. van den Elsen coordinated the multimodal acquisition.
274 A. COLLIGNON ET AL.
References
1. L.G. Brown, "A Survey of Image Registration Techniques", ACM Computing Sur-
veys, Vol. 24, No. 4, pp. 325-376, December 1992.
2. A. Collignon, D. Vandermeulen, P. Suetens, G. Marchal, "3D Multi-Modality Med-
ical Image Registration Using Feature Space Clustering", accepted for presentation
at the CVRMed'95 conference on april 3-5, 1995, and publication in the Lecture
Notes in Computer Science series, Springer-Verlag.
3. T.M. Cover, J.A. Thomas, Elements of Information Theory, John Wiley & Sons,
Inc., New York, 1991.
4. AIM Programme, Project A2003, COVIRA Computer Vision in Radiology, Deliver-
able (2) D1/1.1 Set of raw images and clinical description, 26 June 1992, distributed
by the Commission of the European Communities DG XIII, Brussels.
5. R.O. Duda, P.E. Hart, Pattern Classi cation and Scene Analysis, Stanford Research
Institute, Menlo Park, CA, USA, A Wiley-Interscience Publication, 1973.
6. P. Dupont, G.A. Orban, R. Vogels, G. Bormans, J. Nuyts, C. Schiepers, M. De
Roo, L. Mortelmans, \Di erent Perceptual Tasks Performed with the Same Visual
Stimulus Attribute Activate Di erent Regions of the Human Brain: A Positron
Emission Tomography Study", Proc. Natl. Acad. Sci., USA, Vol. 90, pp. 10927-
10931, December 1993, Neurobiology.
7. D.L.G. Hill, D.J. Hawkes, N.A. Harrison, C.F. Ru , \A Strategy for Automated
Multimodality Image Registration Incorporating Anatomical Knowledge and Im-
ager Characteristics", 13th Int. Conf., Information Processing in Medical Imaging,
IPMI'93, Flagsta , Arizona, USA, June 1993, Lecture Notes in Computer Science,
Vol. 687,pp. 182-196, Springer-Verlag, Berlin Heidelberg.
8. D.L.G. Hill, C. Studholme, D.J. Hawkes, "Voxel Similarity Measures for Automated
Image Registration", Proc. SPIE Vol. 2359, Visualization in Biomedical Computing,
SPIE Press, Bellingham, WA, Rochester, MN, October 4-7, 1994.
9. C.R. Maurer, J.M. Fitzpatrick, "A Review of Medical Image Registration", Inter-
active Image-Guided Neurosurgery, R.J. Maciunas (Ed), Park Ridge, IL, American
Association of Neurological Surgeons, 1993, pp. 17-44.
10. W.H. Press, B.P. Flannery, S.A. Teukolsky, W.T. Vetterling, Numerical Recipes in
C, The Art of Scienti c Computing, Cambridge University Press, 1988.
11. I. Vajda, Theory of Statistical Inference and Information, Kluwer Academic Pub-
lisher, Dordrecht, 1989.
12. P.A. Van den Elsen, Multimodality matching of brain images, Ph.D. Thesis, Utrecht
University Thesis, June 1993.
13. P.A. van den Elsen, E-J.D. Pol, M.A. Viergever, "Medical Image Matching - A
Review with Classi cation", IEEE Eng. in Medicine and Biol., pp. 26-38, March
1993.
14. P.A. Van den Elsen, E.J.D. Pol, T.S. Sumanaweera, P.F. Hemler, S. Napel, J. Adler,
\Grey value correlation techniques used for automatic matching of CT and MR brain
and spine images", Proc. SPIE Vol. 2359, Visualization in Biomedical Computing,
pp. 227-237, SPIE Press, Bellingham, WA, Rochester, MN, October 4-7, 1994.
15. R.P. Woods, J.C. Mazziotta, S.R. Cherry, "MRI-PET Registration with Automated
Algorithm", JCAT, 17(4):536-546, July/August, 1993.
16. R.P. Woods, Automated Image Registration (AIR): User's Manual and Program
Documentation, Division of Brain Mapping Neuropsychiatric Institute, UCLA
School of Medicine, 8/7/1993.