Author :
Ms. Shao Wei
Last Updated : 18 May 2007
Image
Registration for integration of MRI/MRSI information in TRUS-guided prostate
biopsy
1. Background
Prostate cancer is ranked as the second leading cause of cancer death in men,
and among the six most common cancer diseases for Singapore male. Early
detection of prostate cancer can gain much more chances of successful treatment.
However, there are no clear symptoms of prostate cancer until it is quite
advanced, which makes it different from breast cancer or testicular cancer in
which regular self examination can be important in finding early signs of the
disease.
Figure 1. Anatomy of the prostate
Digital rectal examination (DRE) and prostate-specific antigen (PSA) blood
test are routine screening methods for the early detection of prostate cancer.
When a patient has an abnormal DRE result (enlarged or irregular shape) and/or
an elevated PSA level (>4ng/ml), he is suspected to have a cancerous prostate
and will be recommended to undertake a needle biopsy, mostly often under
transrectal ultrasound (TRUS) guidance. In order to detect the cancer, the
biopsy needle is required to sample at the cancer site (if one exists). So an
ideal biopsy protocol should be capable to yield high cancer detection rate with
minimal biopsy points. TRUS guided sextant biopsy for the prostate has been the
standard protocol in this field since the early 90s. With the obtained 2D
ultrasound images of the prostate, the urologist decides the area of interest
and inserts the needle transrectally into the gland to take a tissue sample.
However, the biopsy may still yield a false negative result because of missing
the cancerous targets among the limited number of biopsy sites. Studies have
shown that the TRUS-guided transrectal biopsy can only achieve a positive
detection rate no higher than 30%. Another reason partially attributed to the
low detection rate is the inaccuracy of the needle placement of the current
biopsy procedure.
From 2002, our group have been dedicated to developing a TRUS-guided robotic
system for transperineal prostate biopsy,named URobot. This project is under the
collaboration between a group of researchers in Computer Integrated Medical
Intervention Lab, Nanyang Technological University, and medical consultants in
Department of Urology, Singapore General Hospital. This system aims at a
percutaneous biopsy, and performs multiple-core biopsy through just a
singlepoint, thus overcoming the drawbacks of the conventional prostate biopsy
whose trajectories pass through the fragile rectal wall. Our system allows the
urologist to define the needle¡¯s entry point at the perineal wall, and make
the biopsy plan on the fly, directly on the patient's 3D prostate model which is
constructed from the TRUS images. With this robot, a uniform standard can be
established regardless of the urologist's skills and experience. And the use of
TRUS as the guidance keeps the system radiation-free and low-cost for clinical
application. However, the image quality of ultrasound is just good enough to
determine the gland location and shape, thus to construct the 3D organ model.
Even though the urologist can decide the biopsy cores based on the established
biopsy protocols, such as the sextant or 10-core, it is quite possible that the
biopsy could miss the malignant target, leading to inaccurate or false negative
results, i.e., informing a patient free of cancer but in fact not. In this case,
the robotic implementation could not gain much higher cancer detection rate than
that of the manual approach, since they follow the same standard to choose the
biopsy cores.
A second generation of the robotic system, BioXBot, was designed in 2005,
which inherited the advantages of the previous robotic system and will be
enhanced with new features. The first improvement is to increase the number of
puncture point from one to two, to guarantee an overall coverage of the prostate
for biopsy, which may not due to the space restriction between the biopsy gun
and the TRUS probe, So one of the new features added to BioXBot. Another
important feature is to integrate a suspected cancer map obtained from Magnetic
Resonance Imaging (MRI) and Magnetic Resonance Spectroscopy Image (MRSI) onto
the TRUS image, so that the biopsy can be guided with potential targets. MRI is
a well-established technique that produces high-contrast images regarding to
tissue components. When the cancer carcinoma grow to some extent, it is possible
for experienced radiologists to distinguish them from the normal regions, as the
cancer is usually identified as an area of low signal intensity within the
peripheral zone on T2-weighted images. As an extension of MRI, MRSI is a method
of obtaining biochemical information from a series of local spectrum analysis
over the prostate gland. Significantly higher choline/creatinine levels and
lower citrate levels are usually obtained in regions of cancer compared with
benign and normal prostatic tissue, so the ratio of these metabolites (choline/creatinine
to citrate) in a local region could indicate a positive suspect. There are
already plenty of studies on MRI/MRSI ability in predicting the cancer existence
in the prostate.
A successful image fusion can superimpose the suspected cancer distribution
from MRI/MRSI information on the real-time prostate images, thus provide a
predictive cancer map to the scenario, from which are used to plan the biopsy
cores. This technique should be able to improve the detection rate of the
prostate caner and reduce the likelihood of false negative result in biopsy
findings.
2. Image registration
Image registration is a procedure to determine a transformation between two
image spaces or between an image space and a physical space so that
correspondent features can be matched. Technically, it is an algorithm
concerning the following points:
- Nature of transformation ¨C this must be determined before the
registration is applied: whether the transformation should be a rigid one, with
6 degrees of freedom (DOFs) for three-dimensional space, or a non-rigid one,
allowing much more complex deformation (involving more DOFs).
- Representation of transformation ¨C this is determined by the problem
solver. A unwise selection of the representation would make your problem even
more complex to solve. And it is quite related to the complexity of the
transformation because the DOFs is proportional to the number of unknown
parameters in the transformation to be solved. The usual expression of the
transformation can be polynomial (for either rigid or nonrigid transformation,
mostly for the former), radial basis function like "thin-plate spline"
or "B-spline" (mostly for nonrigid transformation), or deformable
model like ¡°physical model¡± or ¡°diffusion model¡± (for nonrigid
transformation).
- Solve the transformation ¨C The transformation can be solved as a pure
interpolation problem when correspondences at some sparse points are known. For
example, some feature points can be easily identified in both images by human or
algorithms. In that case, any displacement vector in this field can be
calculated by interpolating among those displacement vectors available at
neighborhood correspondences. If the correspondences could not be established
beforehand, searching for the transformation falls into an iterative
optimization procedure: the correspondences have to be searched and evaluated
with certain similarity measurements.
A notable application of the image registration technique in medical field,
is to fuse two image information, that is, to bring the detailed structural or
metabolic information from a high-contrast image source, which is usually
collected pre-operatively regardless of time consumption concern and may even
contain enriched expert diagnosis information, to the another image resource,
which is real-time, easily applicable in intra-operative situation but of poor
quality, like the ultrasound.
However, we know that the prostate is made of soft-tissue, and so does most
of its surroundings tissue. The difference between them is their component and
stiffness. This explains why the prostate would deform comparing the images
captured separately with the two image modalities for the same patient.
Researchers have concluded two possible reasons. One is the different amount of
rectal filling caused by the imaging probes. This cause account for most of the
deformation occurred in the prostate. Since the endorectal MRS probe is much
larger than the TRUS probe, the prostate would be more pushed against the pubic
arch by the MRS probe. The description of the prostate deformation in MRSI
compared to TRUS image has been summarized, that the whole gland increases at
transverse dimension and decreases at anterior-posterior (AP) direction, with
relatively greater decrease for peripheral zone (PZ) than for central gland (CG,
including central zone, CZ, and transition zone, TZ), while no statistical
change at superior-inferior (SI) direction. Another nonnegligible cause comes
from the change of the patient postures where the patient would lie supine in
MRI/MRSI exams, while keeps a lithotomy position in TRUS exam. Different posture
of the feet would also bring different constrains to the prostate. So in our
application, the rigid registration which just solves six degrees of freedom is
not enough to explain the transformation that is far from no or little
deformation. For the sake of accuracy, which is crucial for medical purpose, a
deformable registration procedure is highly preferred. Nevertheless, the rigid
registration still plays an important role in the deformable strategy, as it can
provide an initial transformation ahead of the nonrigid one. This initial
transformation can compensate the global rotation and translation, so that
reduce the amount of displacement has to be solved in the nonrigid step. This
consideration can ensure a stable solution through replacing a transformation
concerning a large number of DOFs by a rigid transformation plus a residue
transformation concerning less deformation.
Manual alignment is a basic and reliable way in image registration since it
often provides a subjective criterion for other techniques claiming free of
human intervention. So it still serves as a popular method in the prostate
registration problem. And usually a good navigation tool is necessary to allow
the user to choose the correspondences between the two sets of images to be
matched. The drawback of this kind of system is that it is not a relaxing work
for the user when the number of correspondence to be defined is large. So the
trend is that the registration algorithms were developed to determine the
transformation (direct way) or correspondences intuitively (indirect way), thus
substitute or decrease the manual work. One common solution to the prostate
registration is surface-based. The prostate surface of both images (regardless
image modalities) is extracted, then registered together either by a geometric
measurement (such as distance) or as a biomechanical model. These methods were
prevalent for registration whose imaging modality is poor in interpreting
intensity contents, e.g., ultrasonography. The surface extraction could be done
by human, or by some processing algorithms, depending on the image quality and
applicability of the processing on the image. Matching of two surface can be
driven by a distance measure like what Zaider et al did for MRI/CT prostate
registration, or by a deformable model taking the tissue stiffness and stress
into account. In theory, the physical model simulates the prostate deformation
best, but it requires a good guess of the tissue parameters (stiffness and
stress) and the boundary conditions of the prostate surface before and after
deformation. Another big family of registration techniques is based on
intensity, which relies on the original image information, usually with no or
quite few pre-processing. This kind of method generally gives more freedom to
the user, but it has imposed a high requirement on the image quality inherently.
Most of the prostate registration using intensity-based method were under the
fact that both the source and target images were of high SNR, and much often the
peripheral structural information other than the prostate in the pelvic images
played an important role, as the monomodal prostate registration presented by
Court et al, Fei et al, Wang et al, and Wu et al, and the multimodal
registration by Lee et al and Schreibmann et al.
In our BioXBot system where MRI/MRSI and TRUS are to be matched, the
intensity-based approach is not reliable, due to the vast lack of anatomic
similarity from the former to the latter. Physical model is a good choic,e but
it requires knowledge of the tissue properties and boundary conditions which are
actually unavailable in reality. Last but not least, FEM is a complex
computation method that consumes time as well as memory. So, in order to reach a
good balance between accuracy and computation load, we utilize a framework
including a global rigid alignment followed by a nonrigid transformation using
thin-plate spline, to match the cross-model prostate surfaces and thereafter
their image volumes.
3. Methods
We employed the following two steps as our strategy to solve the problem:
firstly a rigid registration is applied to search for a global transformation
including three translation vectors and three rotation angles around the x, y
and z axes; a deformable registration is executed to calculate the residual
displacements over the prostate volume. ¡¡
3.1 Rigid registration
Utilization of the global registration assumes a rigid-body between the two
image spaces. Although within the pelvic cavity, only the bone structures, like
the pubic arch, satisfies with this assumption, a rigid registration is still
applicable to some other organs like the prostate, the rectum and etc. The
dispensable condition is that the object chosen for registration must be
identifiable in both image modalities. Otherwise, no correspondence can be found
between the two image spaces therefore no similarity can be evaluated.Two
choices of rigid registration are offered in xRegLib, based on the data
representation of the object selected for registration. The first option is the
surface-to-surface registration technique, given that the organ¡¯s surface in
both images is available for registration. The classic Iterative Closest Point (ICP)
algorithm, proposed by Besl and McKay, is a typical representative for this kind
of technique. ICP algorithm is fast and simple, but requires extra effort to
delineate surface from both images and somehow sensitive to the outlier problem.
BioXBot operating system provides a convenient interface that allows the user to
outline the organ boundaries in the parallel slices and construct a NURBS
(Non-Uniform Rational B-Spline) surface to describe the organ shape. Inputs to
the ICP algorithm are two sets of point clouds sampled on the two surfaces from
different image modalities. The ICP algorithm searches for the transformation Tg
by minimizing a dissimilarity measure fICP, i.e., the mean squared distance
between the two sets of points (source and target), as shown in Equation
1:
(1)
where N denotes the number of points used for correspondence estimation,
which is also the smaller number between the two. (x,y,z) is the point
coordinate in target space, and (x',y',z') is the coordinate in source space.
Usually it is required that the number of points in target space should be no
less than that of the source.
The other option provided is a surface-to-image registration technique, where
only organ surface in MRI/MRSI need to be extracted; the corresponding organ in
TRUS image can be automatically located by maximizing an intensity-based
similarity measurement. Compared to the surface-to-surface based registration,
this approach is superior when less human intervention is preferred in
intra-operative case. It is designed to be a procedure driven by a genetic
optimization to maximize an intensity-based fitness function with respect to the
surface points. Therefore, only surface construct in the pre-operative MRI image
is needed. The intensity-based fitness function, i.e., the similarity measure
that evaluates the registration quality, is formulated for the registration
approach using the prostate surface or the pubic arch surface.
Genetic algorithm (GA) is well-known for its strength in global optimization
and free of initial guess. This is also the reason why we choose GA as the
optimization technique for this approach. GA explores the solution space of a
function through the use of simulated evolution, i.e., the survival of the
fittest strategy. It is a massively parallel (global) search method: rather than
work on one species at a time, it can test and change millions of species in
parallel. Species are chromosomes that encode solutions to the problem at hand.
A fitness function then judges how well a chromosome solves the problem by
assigning a fitness score to each chromosome accordingly. Species evolve by
means of random variation (via mutation, recombination, and other operators),
followed by natural selection in which the fittest tend to survive and
reproduce, thus propagating their genetic materials encoded in chromosome to
future generations.
The formulation of the similarity measure for the prostate surface, is
derived from the fact that there is high image gradient at the boundary of the
organ which is usually of different acoustic properties with its neighbors. is
its intensity is much higher than its neighborhood. So the metric is formulated
as the averaged image gradients along the normals of the surface. We refer it as
¡°Projective Gradient¡± (PG).
(2)
where is the intensity gradient calculated from the original ultrasound image
I. is the normal vector at the i-th surface point . <> is the inner
product operator. Similarly, the nearest-neighbor operator [] is used to
calculate the intensity value at any surface point in ultrasound image.
The similarity measurement for the pubic arch is formulated from the fact
that the ultrasound can hardly penetrate the bone structure. Therefore the only
bon structure in pelvic cavity, the pubic arch, will appear as a high-intensity
echogenic surface with homogenous dark in anterior shadow. So it was evaluated
as averaged radial ¡°gradient¡± between the voxels fallen at the surface
against those at the rear, referred as ¡°Intensity Shadow¡± (IS) measure.
(3)
where M is the user-defined depth of the posterior region (counted in voxels).
denotes the unit vector of the radial echo emitted from the ultrasound probe to
the point on surface at depth of zi. Since the 3D TRUS image was constructed by
a series of transrectal images (in x-y plane) scanned at even intervals (in z
direction), the gray value at any voxel was only determined by the reflected
echo energy emit and received at the same z depth of the voxel. According to
this rule, the intensity ¡°gradient¡± is calculated radially, along the
ultrasound propagation direction, from the transducer center to the surface
point, and within the 2D image slice at this depth. I
3.2 Deformable registration
Once the overall translation and rotation has been solved, the deformable
registration can be applied, to refine the transformation locally, i.e., by
allowing for more DOFs in the transformation. The transformation at this stage
can be described as a partial differential equation of a physical model, or a
displacement field represented in form of polynomial or spline. In xRegLib, the
deformable transformation is represented as a combination of radial basis
function, i.e., thin-plate spline.
An assumption of the non-rigid registration was made in the coordinate system
of the TRUS images, that once the prostate has been globally rotated and
translated by the previous rigid registration, the cross-plane deformation along
SI direction is only scaling. Therefore, the 3D non-rigid registration is done
in slice-by-slice manner. When TPS was employed to describe the deformation
field as an interpolant that minimizes the bending energy through control
points. To determine the control points, say point sets P and Q for source and
target space respectively, manual selection of landmark based on intensity
information is infeasible here because of the poor image quality. Hence, the
geometric information, such as the curvature information, as well as the
arc-length-based parameterization of the pre-deformed and deformed surface, was
utilized to establish the correspondence.
The prostate surface in TRUS images can either segmented manually, or
predicted based on the initialization from registered result. Its
parameterization should follow the same way how SMR was parameterized, thus
described as SUS: xB(u,v) = [x(u,v), y(u,v), z(u,v)]T. And the
rigidly-transformed surface Tg(SMR) will be re-parameterized to accommodate with
the US image coordinate system. Once the two surfaces were parameterized in a
same space, a rough one-to-one mapping T could be established on (u,v) domain,
i.e., correspondence on parametric coordinate:
(4)
where WMR and WUS represent the respective space domain, and D is the
parametric space where 0?u?1 and 0?v?1. This matching is actually relevant to
arc length in u and v directions. Let Q denotes the conversion from parametric
coordinate (u,v) to Cartesian coordinate (x,y,z) in one NURBS surface
formulation, , and Q-1 its reverse conversion, , the transformation T in
Cartesian coordinate system could be worked out by the two point sets P and Q of
same coordinate in parametric coordinate system, using TPS as the radial basis
function.
However, this matching might fail for irregular shapes. Solution to this
problem is to take additional geometric information, such as the curvature
information, into account to determine the correspondence. Once the deformation
was decoupled into u- and v-space according to our assumption, some feature
points were to be identified within each image plane. These prominent features
can be chosen at those corner points which are mostly representative to the
closed contour of the prostate shape in MRI/MRSI scanning. They are the anterior
corner (AC), the posterior corner (PC), the left corner (LC) and right corner
(RC) along the closed 2D contour. Not only for automatic detection, selection of
control points on these corner features can also ensure maximum coverage of the
prostate region. Firstly, slice correspondence was set up by proportional
scaling with respect to the centroid along the z axis (v-space). Secondly, the
prominent corner points (PAC, PPC, PLC and PRC) were identified automatically in
MRI data based on the Gaussian curvature calculated along the contour in each
slice. By connecting the same feature points cross sections, four ridges were
constructed, representing the global frame of the prostate shape. The point
coordinate (x,y) on each ridge will be a function with respect to the section
depth z: . A low-pass filtering (Gaussian filter) is thereafter applied to
reduce the interslice fluctuation in planar point identification, . The smoothed
ridge was then projected back to each section contour and the projection will be
the updated feature points. This procedure can also be repeated until the
projection distance after smoothing is within a threshold. Thirdly, the
correspondent QAC and QPC points in TRUS data were still identified based on
curvature and arc length information. But the correspondence of the other two
features, QLC and QRC, will be determined based on arc length information, that
is, proportional arc length in left (or right) half segment.The number of the
sampling points within each segment was usually selected empirically,
considering the balance between accuracy and computation expense.
Thin-plate spline (TPS) was originally proposed by Harder et al in aircraft
wing designs and later employed to describe deformations that minimizes the
bending energy through control points. The TPS based transformation solves the
transformation in an interpolant manner with respect to displacement between
surface correspondences P and Q, where the displacement vectors are propagated
from the sparse control points to their neighborhood. Its expression for 3D is
as follows,
(5)
where (x, y, z) and (x', y', z') are coordinate in source and target space,
respectively. ci,j and wj,k are the unknown coefficients and weights determined
by Q=Ti(P). Because of the planar assumption on the transformation, the
volumetric deformation would be decoupled to component Tx,y in x-y plane and Tz
along z-axis ,
(6)
with the basis function . Coefficient cz was determined by the proportional
scaling in z axis, with respect to the center of mass. ¡¡
4. Experimental Results
Experimental data, the MRI/MRSI and TRUS images of the patient or the phantom
were acquired using the same systems, but may varying in the acquisition
parameters. We used a 1.5 Tesla whole body GE MR systems (GE Inc, USA) for all
the MRI scans. The patient would be kept in a supine position during the scan.
T2-weighted fast spine-echo scanning was performed under TR/TE of ~6400.0/85.6ms
The slice thickness may range 3~4mm, and slice spacing 0-1.0 mm. Imaging
resolution could be 256x256, or 512x512 pixels in field of view (FOV) of
12.0x12.0 mm2. The ultrasound system integrated in BioXBot system was used to
collect the TRUS image, with a stepping driving the endorectal probe at 1.0mm
interval. During the intra-operative ultrasound imaging, patients were kept in a
lithotomy position for the convenience of delivering biopsy or brachytherapy.
The typical resolution of the transrectal image was 0.18x0.18 mm2.
4.1 Experiment results on the surface-to-image registration algorithm
Experiments on the surface-to-image registration algorithm were only applied
to the prostate and the pubic arch. Due to the space limitation, here we just
illustrated some of the results reported in our previous paper.
We have tested the surface-to-image algorithm applied to the prostate using
projective gradient measure in the self-registration. In the TRUS image self
test, the accuracy evaluated among 5 patients was found to be around
0.59¡À0.20mm in translation and 1.45¡À0.53¡ã in axis-angle rotation.
A sysmetic validation test was performed on the surface-to-image registration
technique applied to the pubic arch, which is used to establish the global
transformation between MRI/MRSI and TRUS. The three fitness function, AI, PG and
IS have been evaluated over fourteen patient data with self-test mode. The
averaged translation error was found to be 2.04¡À0.68mm, 3.31¡À1.15mm and
1.87¡À0.54mm for the AI, PG and IS measure, respectively. Correspondingly, the
averaged rotation error was 3.22¡À1.63¡ã, 4.85¡À1.92¡ã and 2.55¡À
1.13¡ã.
Based on the experimental results, and considering the transformation nature
for the surface-to-image technique, we recommended using the pubic arch surface
as the object to be registered, and use the intensity shadow as its similarity
measure.
4.2 Experimental results on the deformable registration of the prostate
Because of the difficulty to establish the ground truth for the deformable
registration, we use the phantom data, instead of the patient data, to evaluate
the deformable algorithm. A specially-designed elastic phantom, which includes a
simulated prostate, a rectum, a set of pubic bone and surrounding tissues, was
set up for validation purpose (Fig. 2). To quantitatively evaluate the
deformation of the prostate, 45 fiducial markers were seamlessly implanted into
the prostate in a regular distribution. These dummy markers were designed as
small-sized cylinders of 1mm in diameter and 2~3mm in length. They were visible
in MRI images, so their positions could act as the "ground truth" for
error estimation, and would not used for registration itself. Another
superiority of this phantom over other commercially available ones is that its
¡°rectum¡± was designed to be elastic; it is able to expand when the MRS
probe is inserted and shrink back when probe is taken out. The diameter of the
resting rectum is around 10mm. The ¡°pubic bone¡± was mounted beyond the
prostate, to simulate the real patient condition that it restrains the prostate
from displacement in the anterior direction.
In our phantom study with TRUS imaging, we found that the image quality is
not good enough to differentiate the markers from the background speckle noise.
So we make a MRI to MRI pre- and post-deformation registration to simulate the
MRI to TRUS registration problem, as our deformable algorithm have nothing to
doing with the ultrasound image intensity, but only the prostate surface, which
could be replaced by the same organ surface shown in the pre-deformation MRI
images.
The experimental results of the marker¡¯s displacement error (DEi) is
calculated against the markers' displacement error:
(7)
where xA and xB represent the point coordinate in the source and target
space. T is the transformation between the two spaces. Fig.3 showed the MRI-scanned
phantom image (a) with relaxed rectum, and (b) with the inflated endorectal
balloon in rectum, where the scanning condition without probe inserted is used
to simulate the similar TRUS imaging. The statistical analysis of the
registration error over 45 identifiable markers, showed that our method could
achieve an accuracy of about 1.71¡À0.55 mm.
Figure 2. The prostate phantom under the TRUS scanning (the TRUS probe is
driven by a step motor.)
Figure 3. Transverse, sagittal and coronal views of the phantom in MRI images
scanned (a) with the empty ¡°rectum¡±. (b) with MRS endorectal coil filled
in the ¡°rectum¡±. The endorectal balloon was injected by 40ml air. In both
set of volumes, the 3D prostate model
The registration experiments on patient data were demonstrated in Fig 4-7.
Fig. 4 showed the procedure of the ICP algorithm applied to the prostate. Fig. 5
is the application of the surface-to-image registration applied to the prostate
and Fig 6 is the result of the surface-to-image algorithms applied to the pubic
arch for two patients. Fig. 7 is the deformable registration applied to the
prostate.
Figure 4. Registration of the prostate using ICP algorithm (a) segmenting the
prostate from MRI image stack (b) segmenting the prostate from TRUS image stack
(c) registering the MRI surface to TRUS image space.
Figure 5. Registration of the prostate by Surface-to-image algorithm using
projective gradient similarity measurement (a) segmenting the prostate from MRI
image stack (b) import the prostate surface into TRUS image space (c)
registering the MRI surface
Figure 6. Registration experiment by the surface-to-image method applied to
the pubic arch "intensity shadow" similarity measure. (a) Before
registration for patient 1 (b) Before registration for patient 2. (c) After
registration for patient 1. (d) After registration for patient 2
Figure 7. Deformable registration applied to the prostate between MRI and
TRUS. (a) The prostate surface from MRI image (semi-transparent pink) and the
prostate in TRUS image (semi-transparent green). (b) Sagittal view of (a). (c)
The deformed MRI prostate surface (opaque pink), the original surface
(semi-transparent pink) and the TRUS prostate surface (semi-transparent green).
(d) Sagittal view of (c). (e) The MRI prostate surface mesh before registration.
(f) The target TRUS prostate surface mesh. (g) The deformed mesh of (e) after
registration.
5. Conclusion
As we have mentioned, the objective of the image registration is to provide a
suspected cancer map to the intra-operative ultrasound guided biopsy for BioXBot
system. In this means, the cancer information provided by the pre-operative MRI
and MRSI can raise a new protocol for the biopsy planning, which will reduce the
blindness and put more focus those sites with high possibility. As shown in Fig.
8, where the tumor diagnosed from MRI/MRSI are transformed and superimposed onto
the ultrasound image. With the knowledge of the visualized tumor and maybe some
other anatomic structures brought from MRI images, the biopsy could be planned
by clear indication of targets and the biopsy needle could be guided to avoid
some sensitive structures such as the urethra (if its boundary can be extracted
from MRI images).
Figure 8. Deformable registration applied to the prostate between MRI and
TRUS, with simulated tumor inside. (a) The prostate surface from MRI image
(semi-transparent pink), the tumor (opaque pink), and the prostate in TRUS image
(semi-transparent green). (b) Sagittal view of (a). (c) The deformed MRI
prostate surface (semi-transparent pink), and the transformed tumor (opaque
pink). (d) Sagittal view of (c). (e) The surface meshes of the MRI prostate and
tumor before registration. (f) The target TRUS prostate surface mesh (g) The
deformed mesh of (e) after registration.
Acknowledgements
The research group wishes to acknowledge the support of National Medical
Research Council (NMRC) Grant 0859/2004, Urology Center of Singapore General
Hospital and National Cancer Centre Singapore. The author owned great
thankfulness to our collaborators Dr. Choon Hua Thng from National Cancer Centre,
Dr.Christopher Cheng and Dr.Henry Sun from Urology Center.
Publications related to
Registration.
We would be glad if you could sign our guest
book.
For more information, please contact the principal investigator:
A/P Ng Wan Sing
School of Mechanical & Aerospace Engineering
Nanyang Technological University
Nanyang Avenue, Singapore 639798
Fax:(65) 6791 1859
Hospital Partner:
Dr. Christopher Cheng
Head and Senior Consultant
Urology Department, SGH
Outram Road, Singapore 169608
Fax: (65) 6227 3787
Dr. Peter Lu
Head of Div. of Otolaryngology
Changi General Hospital
Singapore
Fax: (65) 6781 6435
