Page 1 IMAGE FUSION FOR CONFORMAL RADIATION THERAPY

Marc L. Kessler, Ph.D. and Kelvin Li, M.S.
Department of Radiation Oncology, The University of Michigan

1. INTRODUCTION
2. IMAGING MODALITIES IN TREATMENT PLANNING
A. X-ray CT
B. Magnetic Resonance Imaging
C. Nuclear Medicine Imaging
D. Ultrasound Imaging
3. TECHNIQUES
A. Dataset Registration
1) Surface-based Registration
2) Image-based Registration
3) Interactive Registration
B.Data Mapping and Image Fusion
1) Structure Mapping
2) Image Mapping and Fusion
4. CLINICAL EXAMPLE
5. SUMMARY
6. REFERENCES
Acknowledgements
This paper is reprinted with permission from 3D Conformal Radiation Therapy and Intensity
Modulated Radiation Therapy: Physics and Clinical Considerations. Ed. J. A. Purdy, W. Grant III,
J. R. Palta, E. B. Butler, C. A. Perez, Madison, WI, Advanced Medical Publishing; 2001.
Page 2
Marc L Kessler, PhD and Kelvin Li, MS Image Fusion for Conformal Radiotherapy AAPM 2001
2
1. INTRODUCTION
Medical imaging is now a fundamental tool in conformal radiation therapy. Almost every aspect of
patient management involves some form of two or three dimensional image data acquired using
one or more modality. Image data are now used for diagnosis and staging, for treatment planning
and delivery and for monitoring patients after therapy. While x-ray computed tomography (CT) is
the primary modality for most image-based treatment planning, other modalities such as magnetic
resonance imaging (MR), positron and single photon emission computed tomography (PET and
SPECT) and ultrasound imaging (US) can provide important data that may improve overall
patient management. For example, MR provides excellent soft tissue contrast, allowing superior
delineation of normal tissues and tumor volumes in many sites. SPECT and PET provide unique
metabolic information capable of resolving ambiguities in anatomic image data and can quantify
partial organ function. Ultrasound imaging can now provide real-time volumetric information for
delineating organ boundaries for both treatment planning and treatment delivery. The data from
these modalities help the clinician to develop a more complete description of the patient. Each of
these modalities alone, however, does not provide all of the physical and geometric information
required for conformal radiotherapy treatment planning. Therefore, in order to take full
advantage of the information from each modality, it must be integrated with the primary CT data
before being used for treatment planning and delivery (Kessler, Rosenman).
This presentation will describe the general techniques available to carry out this integration
process in a quantitative fashion. First, an overview of the advantages and limitations of each
image modality is presented. Next, the basic procedures implemented in academic and
commercial treatment planning systems for supporting the use of multimodality data are
described. Finally, a clinical example is presented to illustrate the steps involved in using
multiple image datasets for conformal radiotherapy treatment planning.
2. IMAGING MODALITIES IN TREATMENT PLANNING
The radiation oncologist now has access to image data from several imaging modalities. In
addition to x-ray CT, data from MR, PET and SPECT and ultrasound imaging may be acquired
during the course of patient management. Each of these modalities has both advantages and
limitations for use in conformal radiation therapy. By registering and fusing the data from the
different modalities, the advantages of each modality can be combined and a more accurate
representation of the patient obtained (Figure 1).
Figure 1. Data from multiple imaging modalities are registered to a common patient reference system (usually
the planning CT) and then combined to construct a more complete representation of the patient.
Page 3
Marc L Kessler, PhD and Kelvin Li, MS Image Fusion for Conformal Radiotherapy AAPM 2001
3
A. X-ray Computed Tomography
X-ray computed tomography (CT) provides the primary dataset for most aspects of conformal
therapy treatment planning. The data from CT are used to construct geometric and physical
models of the patient. The geometric models are used to define anatomic structures, target
volumes and to aid in radiation beam placement and shaping. The physical models provide
density information required by most dose calculation algorithms. The planning CT dataset is
also used to generate graphical aids such as beams-eye-view (BEV) displays and digitally
reconstructed radiographs (DRRs) for planning and treatment verification. The major drawback of
CT data is the limited soft tissue contrast, which can hinder accurate tissue discrimination.
B. Magnetic Resonance Imaging
Magnetic resonance imaging (MR) now plays an important role in treatment planning for several
tumor sites, offering several advantages over CT. The excellent soft tissue contrast provided by
MR, permits better discrimination between normal tissues and many tumors (Khoo). A wide
variety of MR imaging pulse sequences are available that can improve image contrast by
enhancing or suppressing specific tissues such as fat and conditions such as edema. Also, MR
images can be directly acquired along sagittal and coronal planes, offering better visualization of
certain tissues. While these features make MR an excellent choice as a primary dataset for
treatment planning, several limitations have prevented the use of MR data alone. These
drawbacks include the greater susceptibility of MR to spatial distortions and intensity artifacts,
the lack of signal from cortical bone, and image intensity values that have no relationship to
electron or physical density (Fraass, Schad).
C. Emission Computed Tomography (Nuclear Medicine Imaging)
SPECT and PET permit imaging of the transport and accumulation of biologically active tracers.
Examples of tracers for PET include
18
F-deoxyglucose and
82
Rb. and for SPECT include
131
I and
99m
Tc. SPECT and PET studies provide information about physiology rather than anatomy. The
data from these modalities can provide important information about tumor metabolism and
partial-volume tissue function. For patient monitoring, PET imaging has proven useful for
discriminating between tumor recurrence and radiation necrosis (Doyle). For treatment planning
SPECT data has been used to demonstrate regional lung function to help determine beam
directions (Marks) and to quantify the distribution of radiolabeled mono-clonal antibodies (Koral).
D. Ultrasound
Ultrasound imaging uses mechanical (sound) energy rather than electromagnetic energy to image
tissue. Basically, ultrasound images tissue boundaries rather than internal structure. It can also
be used to image blood flow. The major advantages of ultrasound are that images are produced in
real-time, the required apparatus is relatively small, and it does not involve ionizing radiation.
Three-dimensional data acquisition with ultrasound is now also possible. Together, these features
have been exploited for radioactive source placement and boundary definition for prostate
treatments (Narayana, Pathak). A disadvantage of ultrasound is that not all regions of the body,
such as the brain and lung can be imaged effectively due to signal loss at large tissue density
changes such as at tissue-bone and tissue-air interfaces.
While the imaging strengths of CT, MR, PET, SPECT, and ultrasound imaging techniques make
them important tools in patient management, each method has its drawbacks. The goal of image
fusion is to overlap the strengths of each modality. In the next section, the techniques to carry this
out in a quantitative fashion are described.
Page 4
Marc L Kessler, PhD and Kelvin Li, MS Image Fusion for Conformal Radiotherapy AAPM 2001
4
3. TECHNIQUES
The process of incorporating data from multiple imaging studies into the treatment planning
process involves two main tasks. The first task is to estimate the parameters of the coordinate
transformation that relates homologous points in the two studies (data registration). The second
task is to apply the resultant transformation to map structures or features of interest from one
study to another (structure mapping) or to directly combine grayscale data from the two studies
(data fusion). For the discussion that follows the two datasets will be labeled Study A and Study B.
Study A is the base or reference dataset held fixed and Study B the homologous dataset that is
manipulated to be brought into geometric alignment with Study A.
A. Dataset Registration
Numerous techniques for estimating the coordinate transformation that relates homologous points
in two imaging studies have been reported and reviewed in the literature (van den Elsen). The
general approach of these methods is to devise a metric that measures the degree of mismatch (or
similarity) between homologous features in two datasets and to use standard numerical methods
to determine the transformation parameters that minimize (maximize) the metric (Press).
Features used to compute this metric are typically geometric structures (homologous points, lines
or surfaces or combinations of these) extracted from the datasets or the native grayscale data.
Geometric features include manually placed point fiducials and stereotactic frames or internal
landmarks such as anatomic points or extracted surfaces.
The parameters used to model the coordinate transformation between two datasets depend on the
modalities involved and the clinical site. In the simplest situation, it is only necessary to account
for differences in patient orientation (pose) at the time of imaging. For rigid anatomy such as the
skull and pelvis, only 3 rotations (θ
x
, θ
y
, θ
z
) and 3 translations (t
x
, t
y
, t
z
) are required. The presence
of image distortions or mis-calibrated imaging devices would require more degrees of freedom
(DOF) such as anisotropic scaling (s
x
, s
y
, s
z
). While this really a quality control issue and not a
registration problem, one rarely has control over the all the imaging equipment to correct for these
before registration. The situation is more difficult when the anatomy involved is not rigid. In
these cases, a more complicated (spatially variant) transformation involving a larger number of
degrees of freedom is required to register the data properly. While (ad hoc) approaches have been
described to address such situations, it is still an active area of investigation. As such, the major
uses of multimodality image data in conformal therapy are regions of the brain and the pelvis.
To illustrate in more detail the process of dataset registration, two popular algorithms in clinical
use will be described. One algorithm makes use of extracted anatomic structures (surface
matching), and the other uses the grayscale information directly (mutual information).
1) Surfaced-based Registration
In surface-based registration, the surfaces of one or more anatomic structures are extracted from
the image data and used for computing and minimizing the mismatch between the datasets. The
most common anatomic structures used are the skull for brain studies and the pelvis for prostate
studies. These structures can be easily extracted using automated techniques and minor hand
editing. The surfaces from Study A are represented as a binary volume or as an explicit polygon
surface and the surfaces from Study B are represented as a set of points sampled from the surface
(Figure 2). The metric, which represents the degree of mismatch between the two datasets, is
computed as the sum or average of the distances from the points from Study B to the surfaces from
Study A (van Herk).
Page 5
Marc L Kessler, PhD and Kelvin Li, MS Image Fusion for Conformal Radiotherapy AAPM 2001
5
Figure 2. Illustration of surface-based dataset registration. a) Extracted surfaces from the different datasets
(surface from Study B is represented by points sampled from surface, surface from Study A is represented as
an explicit polygon surface. b) Distance of closest approach from each point to the surface is computed. c)
Distances between points and surface minimized.
2) Image-based Registration
In image-based registration, the grayscale data is used directly to compute a measure of mismatch
or similarity between two datasets. One advantage of this approach over surface-based algorithms
is that much less pre-processing of the data is required. Several metrics which measure the
similarity between the grayscale distributions of the two datasets have been described (Meyer,
Studholme, Wells). One method that is being widely developed and used is registration using a
mutual information (MI) based metric.
Mutual information, a concept from the field of information theory, is a measure of how much
redundant information is present between the pair of datasets being analyzed. The amount of
information content in a dataset, in this case images, is determined by integration of the joint and
individual probability densities. The joint probability distribution is computed from the 2-D
histogram of grayscale pairs from the datasets at each iteration of the coordinate transformation
estimate (Figure 3). When two images are not properly aligned, the probability of one voxel value
predicting the content of the homologous voxel value is small, resulting in a reduces MI (less
clustered histogram). Both increasing information redundancy and a higher measure of mutual
information MI (more clustered histogram) is achieved by maximizing these joint probabilities
when the images are aligned.
3) Interactive Techniques
In addition to numerical approaches, interactive dataset registration techniques are also
available. These techniques replace the numerical metric with some form of visual feedback and
the appropriate widgets to manipulate and register two datasets. These techniques are effective in
cases when the transformation can be represented by a small number of degrees of freedom such
as only rotations (3 DOF), translations (3 DOF) and possibly isotropic scaling (1 DOF).
Unfortunately, interactive alignment with more degrees of freedom, such as anisotropic scaling (3
DOF) becomes increasingly difficult and error prone.
Even when using numerical methods, interactive tools can help to determine a good first
approximation for the parameters of the coordinate transformation. As with all search methods, a
good starting point can greatly reduce the time required for the algorithm to converge.
Page 6
Marc L Kessler, PhD and Kelvin Li, MS Image Fusion for Conformal Radiotherapy AAPM 2001
6
Figure 3. Illustration of image-based registration using mutual information. a) T1-weighted and T2-weighted
images from the same anatomic plane but displaced by 6mm. b) The 2-D histogram of joint grayscale values
from the mis-registered images. c) The same images but with no displacement. d) The 2-D histogram for these
images shows that the data is now more clustered (increased MI). (Data courtesy of Charles Meyer, University
of Michigan).
B. Structure Mapping and Image Fusion
Once the transformation relating imaging studies is calculated, part or all of the data relating to
one study may be integrated or fused with that of another. Depending on the goal of the clinician,
one of two types of fusion is generally employed. One approach, called "structure mapping" maps
the outlines of anatomic structures or treatment volumes defined from one imaging study to the
other. Another approach called “image mapping” or “image fusion” involves transforming and
reformatting image data from one study to match the orientation and scale of the images of the
other. This makes it possible to simultaneously visualize grayscale information from
corresponding anatomic planes.
1) Structure Mapping
Structure mapping produces outlines of structures defined from one study on the images of
another. Figure 4 illustrates this process for a patient with an astrocytoma, imaged well by MR
but difficult to define on the CT. The first step is to construct a 3-D approximation of the structure
from the collection of 2-D outlines in Study B. This is accomplished by tiling outlines in adjacent
images and "capping" the top and bottom outlines to create a closed surface (Fuchs). Most
treatment planning systems incorporate such algorithms. The 3-D surface model defined in Study
B is then mapped to Study B using the computed transformation. The desired 2-D outlines for
Study A are derived by "slicing" the transformed surface model along the image planes of Study A.
The resultant outlines are used as input for the treatment planning process.
2) Image Mapping and Fusion
Another approach to integrating information from Study A and Study B is to simultaneously
display images of corresponding anatomic planes from these two imaging studies, regardless of
differences in the imaging parameters under which they were acquired. To accomplish this, the
images from one study are re-sampled to match the scale and orientation of the other (Figure 5).
Various display methods can then be used to see the relationship between the data contained in
the two studies (Figure 6).
Page 7
Marc L Kessler, PhD and Kelvin Li, MS Image Fusion for Conformal Radiotherapy AAPM 2001
7
Figure 4. Process of mapping a structure from one set of images to another. a) The structure is outlined on
the images of Study B (the MR). b) The outlines are stacked and c) used to construct a surface model of the
structure. d) The computed transformation is applied to the model. e) The transformed model is intersected
along the image planes of Study A to produce f) the corresponding outlines of the structure on Study A (the
CT).
Figure 5. Image Mapping. The goal is to create a version of Study B with images that match the scale,
location and orientation of those in Study A. The voxel values for Study B' are determined by back
transforming the coordinates of the voxel into Study B using the inverse of the computed transformation and
interpolating between the surrounding voxels.
Page 8
Marc L Kessler, PhD and Kelvin Li, MS Image Fusion for Conformal Radiotherapy AAPM 2001
8
Figure 6. Various image fusion displays possible using the Study B’and Study A. a) An “electronic
pantograph.” Points drawn on Study A are displayed Study B’and vice versa. b) Split screen display of Study
B’and Study A. c) Movable sub-window over Study B’that displays the grayscale data from Study A.
One method is to display identical anatomic planes in a side-by-side fashion. Such a presentation
allows the definition of outlines of structures using both of the imaging studies simultaneously.
This is achieved by tracking the movement of the cursor over an image from one study with
corresponding movement over the analogous image of the other and vice versa. This technique can
be thought of as an electronic version of a mechanical pantograph. Other presentations made
possible using registered datasets are synthetic images that consist of selected information from
each study. The different information can be composed using overlays, pseudo-coloring, or
grayscale information directly. For example the hard bone features of a CT imaging study can be
combined with the soft tissue features of an MR imaging study by adding the bone extracted from
the CT to the MR dataset. Similarly, a rectangular region from an image of one study can be
removed and replaced with the corresponding region from the other study (Figure 6)
4.
CLINICAL EXAMPLE
An example is presented to illustrate the general process of multimodality data fusion. The case
involves a patient with a large left skull base lesion that was determined to be a benign
menigioma Although there was nothing extraordinary about this particular case, it highlights
many of the steps and requirements involved when using multiple image datasets for conformal
therapy treatment planning.
Data Acquisition
A treatment planning CT study was performed with the patient in a customized immobilization
mask. An MR study was also performed, but without the mask because of space limitations within
the head coil. No attempt was made to adjust the orientation of the image plane (pitch) on the
MRI study to match that used for the CT study, although having the patient in a similar position
during the different imaging studies can greatly simplify the registration process. The CT study
was acquired without contrast (80 images, 3.0mm thickness, 3.0mm separation, 0.67mm pixel
size, 512 x 512 image size). The MR study involved three series using a spin echo sequence;
proton- and T2-weighted axial images (21 images, TE = 15 & 90msec, TR = 4918msec), T1-
weighted post-gadolinium axial images (22 images, TE = 550msec, TR = 1400msec) and T1-
weighted post-gadolinium coronal images (22 images, TE = 16, TR = 500). All MR images were
256 x 256 pixels and had a pixel size of 0.78mm, a slice thickness of 5.0mm, and were separated
by 7.0mm. All scans included the entire head to help facilitate accurate dataset registration and
other aspects of treatment planning (such as DRR generation).
Page 9
Marc L Kessler, PhD and Kelvin Li, MS Image Fusion for Conformal Radiotherapy AAPM 2001
9
Figure 7. Visual validation of estimated coordinate transformation. a) CT were reformatted to match the MR
images and displayed together as an alternating "checkerboard" or "Venetian blind" of two images at the
anatomic same level b) Threshold extracted outlines from a reformatted CT image overlaid on the
corresponding MR image.
Dataset Registration
Registration of the CT and MR datasets was performed using a mutual information-based image
registration program originally developed in the U-M Digital Image Processing Laboratory
(www.med.umich.edu/dipl/). The CT images were downsized to 256 x 256 and the data outside the
head cropped (e.g. air and head holder) to reduce memory and computational requirements. Four
points were placed in approximately the same anatomic locations on both data sets (centers of the
orbits and the top and back of brain) to initialize the algorithm. Convergence to a maximal MI
took less than five minutes. The accuracy of the registration was accessed visually using the
displays shown in Figure 7.
Tumor and Target Volume Definition
The gross tumor volume (GTV) was defined as the region of enhancement in the post Gd-DPTA
contrast MR studies. The physician outlined this volume on both the axial and coronal set of
images (Figure 8). Surface representations of the GTV on both studies were created from the
manually drawn outlines. These surfaces were then transformed in the coordinate system of the
planning CT by applying the computed transformation to the vertices of the surfaces. Once
transformed, the surfaces were re-sliced along the planes defined by the CT images yielding a set
of contours for the coronal-MR GTV and the axial-MR GTV. These contours were used to define a
"composite" GTV from the union of the two sets of MR-based GTV outlines. For this case, the CT
grayscale data did not contribute any information for the definition of the GTV. Had the physician
outlined a CT-based GTV it could have been incorporated directly the composite GTV or compared
to the MR based outlines to reconcile potential ambiguities. The final planning target volume was
created by expanding the composite GTV surface isotropically using a 5 mm margin for setup
uncertainty (in this case the clinical target volume (CTV) was the same as the GTV). The
expanded surface was then re-sliced back onto the CT images and checked by the physician for
final approval.
Treatment Planning and Dose Visualization
At this point, with the necessary coordinate transformations in hand; most of the steps of
treatment planning can be accomplished using any dataset. Because of the confidence in the
geometric accuracy and extent of the external patient surface and the density information from the
planning CT data, this dataset is chosen. Except for these criteria, the MR and CT data are
interchangeable; all options in the planning system can operate with either. For example, it is
just as simple to display dose over the MR data as the CT (Figure 9).
Page 10
Marc L Kessler, PhD and Kelvin Li, MS Image Fusion for Conformal Radiotherapy AAPM 2001
10
Figure 8. Left side a) Post contrast axial with GTV. b) Reconstructed sagittal CT with composite MR derived
GTV. c) Axial CT with coronal and axial MR derived GTVs and composite GTV. d) Post contrast coronal MR
with GTV. The thin white lines show the intersection of the plane of the other image in the same row. Right
side a) MR axial derived GTV. b) MR coronal derived GTV volume, optic nerves, and optic chiasm. c)
Composite GTV from MR axial and MR coronal tumor volumes. Orbits are derived directly from the CT
images.
Figure 9. Dose display using a CT registered MR dataset. a) Isodose lines. a) Dose "Colorwash."
Page 11
Marc L Kessler, PhD and Kelvin Li, MS Image Fusion for Conformal Radiotherapy AAPM 2001
11
5.
SUMMARY
In this presentation the benefits and process of combining multiple imaging modalities together to
enhance the accuracy of conformal radiation therapy have been described. While the imaging
strengths existing in CT, MR, PET, SPECT, and ultrasound imaging techniques make them
important tools in patient management, each method has its drawbacks. The goal of image fusion
is to effectively overlap the strengths of one imaging modality over the weaknesses of another.
Various types of image registration techniques are now being used in the clinic. The objective of
each is to derive a coordinate transformation that maps homologous anatomic points from one
dataset to the other. Each method uses a different measure of success. Surface-based registration
measures success by minimizing the distances between points and surface representations of
anatomic structures extracted from each dataset. Image-based registration measures success
directly from voxel grayscale data in the two datasets. One robust image-based algorithm operates
by maximizing the mutual information content between the two datasets. Interactive techniques
are also in wide use and involve manual manipulations of structure or image orientation and
allow immediate visual feedback and verification of success.
Once a coordinate transformation has been computed between two datasets, structure mapping
and image fusion can be used to visualize the data from one dataset in relation to another.
Structure mapping permits the display of outlines drawn on one dataset over the images of the
other. Image mapping and fusion allows 2-D images from the two datasets to be visualized
adjacent to or superimposed over the other.
The application of these techniques has been demonstrated using a typical clinical example using
CT and MR studies for a patient with a menigioma. While only two imaging modalities were used
in this example, the general steps required to exploit data from any imaging modality for
treatment planning in conformal therapy are the same.
6.
REFERENCES
Doyle WK, et al, Differentiation of cerebral radiation necrosis from tumor recurrence by [18F]FDG
and 82Rb positron emission tomography. J Comput Assist Tomogr, 11(4):563-70, 1987.
Fraass BA, et al. Integration of magnetic resonance imaging into radiation therapy treatment
planning: I. Technical considerations. Int J Radiat Oncol Biol Phys, 13:1897-908, 1987.
Fuchs H, Kedem ZM, Uselton SP. Optimal Surface Reconstruction from Planar Contours. Comm
ACM 1977; 10:693-703.
Kessler ML, Pitluck S, Petti P, Castro JR. Integration of multimodality imaging data for
radiotherapy treatment planning. Int J Radiat Oncol Biol Phys, 21:1653-67, 1991.
Khoo V.S., et al., Magnetic resonance imaging (MRI): considerations and applications in
radiotherapy treatment planning, Radiother Oncol,42(1): 1-15; 1997.
Koral KF, Lin S, Fessler JA, Kaminski MS, Wahl RL, Preliminary results from intensity-based
CT-SPECT fusion in I-131 anti-B1 monoclonal-antibody therapy of lymphoma. Cancer, 80(12
Suppl):2538-44, 1997.
Page 12
Marc L Kessler, PhD and Kelvin Li, MS Image Fusion for Conformal Radiotherapy AAPM 2001
12
Marks LB, et al, The role of three dimensional functional lung imaging in radiation treatment
planning: the functional dose-volume histogram. Int J Radiat Oncol Biol Phys, 30;33(1):65-75,
1995
Meyer CR, et al, Demonstration of accuracy and clinical versatility of mutual information for
automatic multimodality image fusion using affine and thin-plate spline warped geometric
deformations, Med Image Anal, 1(3):195-206, 1997
Narayana V, Roberson PL, Winfield RJ, McLaughlin PW, Impact of ultrasound and computed
tomography prostate volume registration on evaluation of permanent prostate implants, Int J
Radiat Oncol Biol Phys, 39(2):341-6, 1997
Pathak SD, Grimm PD, Chalana V, Kim Y, Pubic arch detection in transrectal ultrasound guided
prostate cancer therapy, IEEE Trans Med Imaging 17(5):762-71, 1998.
Press WH, Teukolsky SA, Vetterling WT, Flannery BP, Numerical Recipes in C: the art of
scientific computing. 2nd ed, Port Chester, NY: Cambridge University Press, 1992.
Rosenman, J.G., et al., Image registration: an essential part of radiation therapy treatment
planning, Int J Radiat Oncol Biol Phys, 40(1): 197-205; 1998.
Schad LR, Loot S, Schmidt F, Sturm V, Lorenz WJ. Correction of spatial distortion in MR imaging:
a prerequisite for accurate stereotaxy. J Comput Assist Tomogr, 11:499-505. 1987.
Studholme C, Hill DL, Hawkes DJ, Automated 3-D registration of MR and CT images of the head, Med Image Ana1,
1(2):163-75, 1996.
Ten Haken R.K., et al., A quantitative assessment of the addition of MRI to CT-based, 3-D
treatment planning of brain tumors, Radiother Oncol 25(2):121-133, 1992.
van den Elsen PA, Pol MA, Viergever MA. Medical Image Matching - A review with classification.
IEEE Engineering in Med and Biol, 12:26-39, 1993.
van Herk M., et al., Automatic three-dimensional correlation of CT-CT, CT-MRI, and CT-SPECT
using chamfer matching, Med Phys, 21(7): 1163-1178; 1994.
Wells WM 3rd, Viola P, Atsumi H, Nakajima S, Kikinis R, Multi-modal volume registration by
maximization of mutual information. Med Image Anal, (1):35-51, 1996.