[ieee 2009 ieee international conference on signal and image processing applications - kuala lumpur,...
TRANSCRIPT
Automatic Removal of Extracardiac Hotspots in
Technetium-99m Myocardial Perfusion SPECT W. H. Tan
#1, G. Coatrieux
*2, R. Besar
#3, B. Solaiman
*4
#Center for Multimedia Security and Signal Processing, Multimedia University
Persiaran Multimedia, 63100 Cyberjaya, Selangor, [email protected]@mmu.edu.my
*LaTIM, INSERM U650, Departement ITI, Institut Telecom Bretagne
Technopôle Brest-Iroise, CS 83818, 29238 Brest Cedex 3, France2 [email protected]
4
Abstract— In Technitium-99m myocardial perfusion SPECT
(MPS) tomograms, there is usually a substantial radioactive
tracer uptake in the abdominal organs, especially the liver, bowel
and stomach. This extracardiac activity frequently emerges as
areas of intense brightness or hotspots, which hamper efforts in
automatic MPS quantification. Though it would be favourable to
remove the hotspots, their multitudinous appearance and
proximity to the heart have made them difficult to be removed.
In this paper, we propose an image processing technique to
automatically remove the hotspots. Our technique uses the
morphological watershed segmentation to delineate the hotspots
before they are iteratively removed. The proposed technique has
been applied on clinical MPS tomograms in which it has
completely removed the hotspots in 90% of the test data. In
addition, it has also shown to increase the success rate of an
automatic left ventricle detection scheme to 100%.
Keywords— Myocardial perfusion SPECT, extracardiac
hotspots, artifact removal, image processing, watershed
segmentation
I. INTRODUCTION
Single photon emission computed tomography (SPECT) is
an imaging technique based on nuclear medicine. It detects
gamma rays emitted singly by the radiopharmaceutical, which
is injected into the patient and localizes within one or more
organs based on its biochemical properties [1]. The gamma
rays are then detected using either a rotating camera or special
purpose multi-detector device.
Myocardial perfusion SPECT (MPS) is a form of functional
cardiac imaging, employed for the diagnosis of ischemic heart
disease [2]. MPS reveals the distribution of the
radiopharmaceutical, and therefore the relative blood flow to
the different regions of the myocardium. Diagnosis is then
made by comparing a set of stress images to a set of images
taken at rest, based on the principle that there is less blood
flow in diseased myocardium under the condition of stress.
The quantification of MPS images begins with the
generation of transaxial images through the body by using
image reconstruction techniques. The reconstructed images
will then undergo a series of image processing routines to
suppress the noise and enhance the desired features within the
images. As the transaxial images are perpendicular to the long
axis of the patient which varies from patient to patient, it is
customary to reorient them into short-axis images that are
perpendicular to the long axis of the left ventricle (LV) [3]. In
order to represent all areas of the myocardium in a single
image, the cardiac polar map is used. It is generated by
plotting the myocardial circumferential profiles as histogram
values against angular location through a polar coordinate
transformation with increasing radii [4].
Since the quantification of MPS images involves a series of
processes, automation is desired as it will drastically reduce
the labour-intensiveness of the underlying tasks. One of the
essential steps towards the automated approach is the
recognition of the LV in the MPS tomograms. Once the LV is
detected, reorientation and selection of ROI for polar map
generation can be carried out automatically.
In Technitium-99m (99m
Tc) MPS tomograms, automatic LV
recognition is hampered by the presence of the intensely
bright regions or hotspots. Due to the prominent hepatobiliary
excretion of 99m
Hotspots Myocardium
Tc-labelled perfusion agents used in MPS,
there is usually a substantial radioactive tracer uptake in the
liver [5]. Besides, there is also noticeable tracer uptake in the
stomach, bowel and occasionally in the thyroid gland. This
intense extracardiac activity frequently emerges as hotspots
that introduce streaking artefacts in the reconstructed
tomograms, as shown in Fig. 1. The hotspots often appear in
various shapes, sizes as well as intensities, and they are
usually located in close proximity to the heart. Thus, it is
difficult to exclude them from the tomograms and their
presence obscures the recognition of LV and complicates the
subsequent processes in MPS quantification.
Fig. 1. The hotspots and the myocardium in the mid-ventricular slice of a MPS tomogram.
2009 IEEE International Conference on Signal and Image Processing Applications
978-1-4244-5561-4/09/$26.00 ©2009 312
One way to mitigate the effect of these hotspots is by
having attenuation correction using simultaneous
transmission-emission tomography [6]. Nevertheless, this
approach requires dedicated hardware and thus increases the
cost of the equipment. An alternative solution is to remove the
hotspots via image processing techniques. It can be observed
that the high radiotracer uptake gives rise to intensely bright
regions in the reconstructed tomograms. Thus, the hotspots
can be removed by detecting and masking out these high
intensity regions from the three-dimensional MPS tomograms.
In this paper, we propose an image processing technique
based on the morphological watershed segmentation to
automatically detect and remove the hotspots from the MPS
tomograms. This paper is organized as follows. Section II
describes the underpinning theories and methods of the
proposed technique. Section III gives the details of the
experiments and discusses the results. Finally, conclusion is
drawn in Section IV.
II. METHODOLOGY
The solution adopted in our work consists of delineating the
hotspots by making use of three-dimensional morphological
watershed segmentation. The regions corresponding to the
hotspots are then iteratively removed depending on their
spatial location within the MPS tomograms.
A. Morphological Watershed Segmentation
Segmentation by morphological watersheds was first
proposed by Beucher and Lentuéjoul [7]. Since its debut,
morphological watershed segmentation has gained much
popularity due to its attractive attributes: natural borders and
elegant combination of edge- and region-based segmentation.
Over the years, several implementations have been proposed
[8], [9]. In this paper, we adopt and extend the approach
proposed by Tan et al. [10], which has been derived from the
rainfalling simulation [9].
In this approach, a linking scheme is defined to recursively
link each pixel in the opposite direction of the strongest edge
in its neighbourhood. By using this scheme, every pixel in the
image is gradually linked to a regional minimum. The pixels
that have been linked to the same regional minimum are in the
same catchment basin. Adjacent catchment basins are
separated by the watershed lines, which coincide with the
strongest edges between them. Thus, the catchment basins
represent dissimilar regions in the image, with the watershed
lines acting as the borders between neighbouring regions.
Consider the synthetic test image with two bright blobs,
one at the upper half and one at the lower half of the image, as
shown in Fig. 2(a). When morphological watershed
segmentation is applied, the image is divided into two regions,
the left and the right halves, corresponding to the left and right
regional minima in the image. Fig. 2(b) depicts the image’s
topographical profile, with the watershed line (hashed)
dividing the image into the left and right regions. This
example illustrates that morphological watershed
segmentation does not segment the image into the upper and
lower halves according to the position of the bright blobs.
Next, let us invert the image by using the following function:
'( , ) max ( ) ( , )P x y P P x y! " (1)
where P(x, y) is the pixel at location (x, y) of the input image,
P’(x, y) is the pixel at location (x, y) of the inverted image and
max(.) is a function which returns the maximum intensity
value of the image. Fig. 2(c) shows the inverted image of Fig.
2(a). When morphological watershed segmentation is applied
on this image, it is divided into two regions as well. The upper
region corresponds to the regional minimum at the bottom of
the inverted upper blob, while the lower region corresponds to
the regional minimum at the bottom of the inverted lower blob.
The topographical profile of the inverted image is shown in
Fig. 2(d) with the watershed line (hashed) dividing the image
into the upper and lower regions. Thus, to delineate the image
according to the location of the bright regions, the image
should be inverted before morphological watershed
segmentation is applied.
(a) Test image.
(b) Topographical profile of the test image.
(c) Inverted test image.
(d) Topographical profile of the inverted test image. Fig. 2. Test images and their topographical profile.
313
The approach proposed in [10] is only applicable to two-
dimensional images. In this work, we extend it to three-
dimensional image volumes such as MPS tomograms. The
pseudocode of the three-dimensional morphological
watershed segmentation is outlined below:
v D# $!
% &
% &% &arg min
w n vL v w
$!
! ! !
! ! !
While % & % &% &L v L L v'! ! !! !
% & % &% &L v L L v!! ! !! !
v D# $!
% & % &S v L v!
! !! !
where
% &, ,v x y z!! : running voxel;
% &, ,w x y z!! : neighbour of running voxel;
% &( ), , | 0 ,0 ,0x y zD x y z x N y N z N! * + * + * + : input domain;
% & % &(, , , , | 1 1, 1 1,n x y z i j k D x i x y j y! $ " * * , " * * ,!
)1 1z k z" * * , : neighbourhood of v!
;
% &, ,L x y z!
: linked volume;
% &, ,S x y z!
: segmented volume
B. Proposed Hotpot Removal Technique
To segment the hotspots due to extracardiac activity, the
MPS tomogram is first inverted. Gaussian filtering with a
scale of - = 1 is then applied to smoothen the volume as well
as to remove regions with homogenous intensity [11]. Next,
the aforementioned three-dimensional morphological
watershed segmentation is applied on the inverted MPS
tomograms. This demarcates the bright structures within the
tomogram, including the hotspots. To detect the region of the
most prominent hotspot, the position of its epicentre which
corresponds to brightest voxel is determined and the
watershed region in which it resides is found. Once the region
is known, the hotspot is removed from the MPS tomogram by
setting the intensity of all voxels in that region to zero. As
there are usually more than one hotspots, these processes are
repeated until all the hotspots are removed. However, doing so
may also accidentally remove the myocardium, which is also a
bright structure. To overcome this, we consider the spatial
distribution of the human organs within the MPS volume. In
this case, the transaxial slices in a MPS tomogram are
separated into three parts, as shown in Fig. 3. The slices in the
first and third tierces are categorized as the extracardiac slices,
while those in the second tierce are categorized as the
myocardial slices since the myocardium can be found in these
slices. As a result, a bright region will be removed only if its
epicentre is currently the brightest voxel and it is located in
the extracardiac slices. This process is iteratively conducted
until the brightest voxel is no longer found in the extracardiac
slices. The whole flow of our approach is illustrated in Fig. 4.
Tierce 1
Tierce 2
Tierce 3
Fig. 3. Division of transaxial MPS slices into three tierces.
Gaussian
filtering
Invert MPS
tomogram
Start
End
Morphological
watershed
segmentation
Locate highest
intensity voxel
in MPS
tomogram
Is it in the
extracardiac
slices?
Set all voxels
in its region
to 0
Read MPS
tomogram
Update
MPS
tomogram
Yes
No
Fig. 4. Flowchart of the processes in the proposed technique.
III. EXPERIMENT AND DISCUSSION
In order to evaluate its effectiveness, the proposed solution
was applied for hotspot removal on 50 sets of MPS
tomograms. These tomograms were smoothed with a two-
dimensional Butterworth filter (order=5, cutoff=0.25Nyquist),
as routinely done for MPS studies and reconstructed over 180.
(45. RAO to LPO) with ramp filter and filtered
backprojection. Our procedure (see Fig. 4) was applied on
each set independently.
From these experiments, it can be observed that the
proposed technique has successfully removed all hotspots for
45 over 50 sets of tomograms used in the experiment. Hence,
it has achieved a success rate of 90% in complete hotspot
removal. To illustrate the obtained results, a sample of the
tomogram before and after hotspot removal is shown in Fig. 5.
In all five cases for which our technique failed, the remaining
hotspots have lower intensity than the myocardium. Hence,
they were not removed. Nevertheless, their presence would
not considerably hamper the subsequent processes in MPS
quantification due to their lower intensity.
314
(a) Before hotspot removal.
(b) After hotspot removal.
Fig. 5. A sample result illustrating the effectiveness of the proposed technique.
To exemplify its effectiveness, the proposed technique was
compared to the conventional gray-level thresholding
technique. In this case, the gray-level threshold was set to the
95th percentile of the cumulative frequency distribution of the
gray-level values in the reconstructed MPS tomogram. For all
voxels with gray-level values above the threshold, their gray-
level values were set to zero. To avoid accidental removal of
the myocardium, the thresholding scheme was only applied in
the extracardiac slices, similar to the approach discussed in
Section II. Fig. 6 depicts a sample of the original
reconstructed MPS tomogram, the result after hotspot removal
via thresholding and the result after hotspot removal with the
proposed technique. With the thresholding based approach,
residues of the hotspots can still be seen in Slice 15 to 28 of
the resulting tomogram, as shown in Fig. 6(b). These residues
were not removed via thresholding as they were located in the
myocardial slices. Conversely, the hotspots were completely
removed using the proposed technique, as illustrated in Fig.
6(c). Even though some parts of the hotspots span into the
myocardial slices, they were removed by the proposed
technique as they were located in the same three-dimensional
watershed segments. Thus, the hotspots will be removed as
long as their epicentres (brightest voxels) are located in the
extracardiac slices.
(a) Original reconstructed MPS tomogram
(b) After hotspot removal via thresholding
315
(c) After hotspot removal with the proposed technique
Fig. 6. Results of hotspot removal via thresolding and the proposed technique.
To illustrate how the proposed technique may assist in
automated MPS processes, we applied an automatic LV
detection scheme, described in [12], on the same sets of MPS
tomograms. The LV detection rate before and after hotspot
removal was then compared. Prior to hotspot removal, the LV
detection scheme had successfully detected LV in 46 sets of
the tomograms. It failed in four sets of tomograms due to
hotspots’ obscuration. Once our procedure had been applied,
LV detection was successfully conducted in all 50 sets of the
tomograms. A sample of the results in which the LV was
successfully detected after the hotspot removal is shown in
Fig. 7. By improving the LV detection rate, it would facilitate
automatic LV reorientation and ROI selection for polar map
generation, which would lead to better MPS quantification.
(a) Unsuccessful LV detection before hotspot removal.
(b) Successful LV detection after hotspot removal.
Fig. 7. A sample result depicting the improvement in LV detection with the
proposed technique.
IV. CONCLUSION
Extracardiac activity is a common occurrence in 99m
ACKNOWLEDGMENT
Tc
MPS. It is induced by excessive radioactive tracer uptake in
the abdominal organs, especially the liver, bowel and stomach.
This activity frequently emerges as hotspots in the MPS
tomograms. Due to their intense brightness, the hotspots often
severely hinder automatic MPS quantification. In this paper,
we have proposed an image processing technique to
automatically remove the hotspots from MPS tomograms
based on the morphological watershed segmentation and MPS
slice categorization. The efficiency of the proposed approach
has been underlined on 50 sets of clinical MPS tomograms
with a successful rate of 90%. Furthermore, we have also
illustrated how this approach can be used to improve the
detection rate of an automatic LV detection scheme, which is
a critical step towards successful MPS quantification.
The authors would like to thank Dr. Lee Boon Nang from
Department of Nuclear Medicine, Hospital Kuala Lumpur,
Malaysia for providing the cardiac SPECT tomograms for use
in this work.
REFERENCES
[1] R. J. Jasczak, “Tomographic radiopharmaceutical imaging,” Proceedings of the IEEE, vol. 76, pp. 1079 – 1094 , Sep. 1988.
[2] (2009) Wikipedia webpage on SPECT. [Online]. Available:
http://en.wikipedia.org/wiki/SPECT [3] G. Germano and D. S. Berman, Clinical Gated Cardiac SPECT, 2nd
ed., Massachusett, United States of America: Blackwell Futura, 2006.
[4] E. G. DePuey, E. V. Garcia and D. S. Berman, Cardiac SPECT Imaging, 2nd ed., Philadelphia, United States of America: Lippincott
Williams & Wilkins, 2001.
[5] G. Germano, T. Chua, H. Kiat, J. S. Areeda and D. S. Berman, “A Quantitative Phantom Analysis of Artifacts Due to Hepatic Activity in
316
Technetium-99m Myocardial Perfusion SPECT Studies,” Journal of Nuclear Medicine, vol. 35, pp. 356 – 359, Feb. 1994.
[6] E. N. Heller, P. DeMan, Y. H. Liu, D. P. Dione, I. G. Zubal, F. J.T.
Wackers and A. J. Sinusas, “Extracardiac Activity Complicates Quantitative Cardiac SPECT Imaging Using a Simultaneous
Transmission-Emission Approach,” Journal of Nuclear Medicine, vol.
38, pp. 1882 – 1890, Dec. 1997. [7] S. Beucher and C. Lentuéjoul, "Use of watersheds in contour
detection," Proc. Int. Workshop Image Processing, Real-Time Edge
and Motion Detection/Estimation, 1979, pp. 17 – 21. [8] L. Vincent and P. Soille, "Watersheds in digital spaces: an efficient
algorithm based on immersion simulations," IEEE Trans. Pattern Anal.
Machine Intell., vol. 13, pp. 583 – 596, Jun. 1991.
[9] A. N. Moga, B. Cramariuc, and M. Gabbouj, “A parallel watershed algorithm based on rainfalling simulation,” In European Conf. on
Circuit Theory and Design., 1995, vol. 1, pp. 339–342.
[10] W.H. Tan, M. Bister and G. Coatrieux, “Uncommitted Morphological Merging of Watershed Segments”, Proc. 2nd IEEE Int. Conf.
Information & Communication Technologies: from Theory to
Applications, 2006, vol. 1, pp. 1573 - 1577. [11] J. M. Gauch, “Image Segmentation and Analysis via Multiscale
Gradient Watershed Hierarchies,” IEEE Trans Image Processing, vol.
8, pp. 69 – 79, Jan. 1999. [12] W. H. Tan, R. Besar, G. Coatrieux and B. Solaiman, “PSO based
parametric object recognition,” IEICE Electron. Express, vol. 5, pp.
1080 – 1087, Sep. 2008.
317