Upload
vanxuyen
View
236
Download
1
Embed Size (px)
Citation preview
Application of Pattern Recognition Framework for Quantification of Parkinson’s Disease in DAT
SPECT Imaging
Saurabh Jain, Yousef Salimpour, Laurent Younes, Gwenn Smith, Zoltan Mari, Vesna Sossi, Arman Rahmim
Abstract–Dopamine transporter (DAT) SPECT imaging is
increasingly utilized for diagnostic purposes in suspected
parkinsonian syndromes. Visual classification or
quantitative analysis of mean regional uptake has been
performed in the past. Our objective is to enable further
enhanced clinical utility in the diagnosis as well as tracking
of progression in Parkinson’s disease via quantification
based on pattern recognition. We developed and
implemented two such frameworks: first, we utilized
shape/texture metrics that did require registration to a
common structure/template; e.g. 3D moment-invariants,
Haralick texture features, and multiple others. We also used
a surface registration algorithm, which falls under the broad
class of Large Deformation Diffeomorphic Metric Mapping
(LDDMM). In this latter framework, we obtain a common
coordinate system for the entire population based on MR
images, and compare SPECT intensities across subjects in
this common coordinate system. This method has the
advantage of estimating population-based templates for each
structure individually rather than using a predetermined
collective atlas for all regions, as is customary. In this
common coordinate system, we then used Principal
Component Analysis (PCA) on intensities to obtain sub-
regions (set of voxels inside the structure of interest) with
highest variance in SPECT intensities across subjects. We
show that the healthy and diseased populations can be
subsequently distinguished. Via these methods, we also
aimed to assess correlations with different clinical measures
(e.g. UPDRS score, disease duration). In addition to enabling
enhanced diagnostic task performance, these methods have
considerable potential as biomarkers of PD progression.
I. INTRODUCTION
Imaging of the dopaminergic system with SPECT has
become widespread in Europe and has entered a new
active phase in the US since 123
I-ioflupane-dopamine
transporter (DAT) SPECT was approved by the FDA in
2011. Visual interpretation is the common assessment
approach [1-3]. However, more objective assessment can
be performed using quantitative analysis [4] involving
manual or automated ROI drawing and analysis [5, 6].
Our proposed approach is based on the observation that
Manuscript received December 06, 2014. S. Jain and L. Younes are with the Center for Imaging Science,
Johns Hopkins University, Baltimore, MD, USA 21287. Y. Salimpour and Z. Mari are with the Department of Neurology,
Johns Hopkins University, Baltimore, MD, USA 21287. G. Smith is with the Department of Psychiatry, Johns Hopkins
University, Baltimore, MD, USA 21287. V. Sossi is with the Department of Physics & Astronomy, University
of British Columbia, BC, Canada V6T1Z1. A. Rahmim is with the Departments of Radiology, and Electrical &
Computer Engineering, Johns Hopkins University, Baltimore, MD, USA 21287.
SPECT and PET images convey important information at
the voxel level, whereas commonly used ROI-averaging
vastly oversimplifies the available spatial uptake
information. We propose to explore a novel quantification
paradigm applied to SPECT neurochemical imaging that
utilizes pattern recognition to quantify uptake alterations
with disease, expected to detect subtle regional DAT
losses: this is especially important as DAT binding can be
a marker of PD at the presymptomatic stage [7, 8] and the
proposed approach can enhance abilities to discriminate
among different populations. Furthermore, we seek to
move beyond classification tasks in clinical DAT SPECT
imaging [9], to also arrive at enhanced markers of
progression. This is a significant step, and is for instance
consistent with the aims of the Parkinson’s Progressive
Marker Initiative (PPMI) to identity biomarkers of PD
progression, a critical step in the development of novel
and enhanced treatments for PD [10].
We have in the past pursued pattern recognition
techniques, including shape and texture analysis in the
context of quantitative brain PET imaging, in studies of
PD [11-13] and neuroinflammation [14]. These
techniques have the advantage of not requiring
normalization/registration of all regions of interest (ROIs)
to a common structure. In the present work, we turn our
attention to DAT SPECT imaging, given its increasingly
popular clinical usage. Furthermore, we have evidence
that application of shape and texture metrics can
appropriately extend from the higher resolution spectrum
of PET imaging to the domain of SPECT imaging [15].
In parallel, we have also utilized a distinct and
powerful normalization framework, enabling generation
of a population-based template of ROIs, to which
individual subject ROIs are mapped, enabling application
of other pattern recognition tools, such as principal
component analysis (PCA), and generation of sub-
structure maps of correlation with clinical manifestations.
To do the latter, we investigated the use of surface
matching based on Large Deformation Diffeomorphic
Metric Mapping (LDDMM) [16], as a method for
obtaining a common coordinate system to compare
SPECT intensities across subjects. Nonlinear mapping of
intensities, commonly known as normalization, is
performed in popular neuroimaging analysis packages
such as SPM by mapping MRI images to predetermined
atlases. Compared to more typical normalization
methods, our approach is based on surface matching, as
opposed to image matching, which is more robust under
noise, since it is not affected by noise once we obtain an
appropriate segmentation. Furthermore, our method has
the advantage of estimating population-based templates
for each structure individually rather than using a
predetermined collective atlas for all regions, as is
customary. We used our LDDMM framework, described
in Section II, to first obtain a population-based template
from the caudate and putamen segmented from MRI
scans, and then obtain non-linear maps between each
subject and the template. These mappings are then used to
map the SPECT intensities in caudate and putamen to the
common coordinate system of the template. We then
perform PCA of the intensities in the common coordinate
system to obtain low dimensional features, which
correspond to regions in the caudate and putamen in
which the intensities vary the most across subjects.
For consistency, subjects were selected from the PPMI
database such that SPECT data were acquired on the
same kind of scanner, while having high-resolution 3T
MRI images, and age was limited to <70 to minimize
further age-related confounds. With this criteria, we
analyzed 128 subjects, 48 of which were healthy controls
(HCs) while 80 were diagnosed with Parkinson’s Disease
(PD). We show in Section V that for this particular
population, the uptake in putamen is enough to classify
HC from PD using a simple linear classifier in the low
dimensional space of principal components. However,
what is of greatest interest to us in the present work is the
ability to quantify progression and not merely enable
classification, as we discuss later in this work.
II. LDDMM FRAMEWORK
In the LDDMM framework, a non-linear map
between two surfaces, say and , is obtained by
integrating a time varying vector field on the
underlying Euclidean space , so that it is a solution of
the differential equation
, , , with the initial condition , 0 for all . The
solution to this equation for a given vector field , is
denoted by . The optimal mapping is obtained by
minimizing the cost functional
| | | , 1 | , where the first term penalizes the size of the vector field
in an appropriate norm, which in turn, controls the
smoothness of , and the second term penalizes the
mismatch between the mapped surface , 1 and the
target surface T. The details of this algorithm can be
found in [16].
The template is generated via a Bayesian generative
model described in [17], by mapping a hyper-template to
a template using a deformation parameterized with a
random vector field, as above, and maximizing the joint
likelihood of the observed target surfaces, which are
considered to be noisy versions of the deformed template.
III. TEXTURE AND SHAPE ANALYSIS
METHODS
In parallel to utilization of a normalization (LDDMM)
framework, we pursued other pattern recognition tools
that do not require normalization. We have elaborated
upon these in past application to PET imaging [14], and
briefly outline here:
(a) 3D moment invariants (3D-MIs): spatial
descriptors designed to be invariant to scaling, translation
and rotation, with extensive use in areas such as pattern
recognition, and as we have utilized in brain imaging.
These metrics included 2nd
order J1, J2 and J3 as well as
3rd
order B3 and B4 metrics.
(b) Intensity histogram analysis: e.g. standard
deviation, skewness and kurtosis. We also performed
cumulative histogram analysis leading to the intensity-
volume histogram [18].
(c) Neighborhood gray tone difference (NGTD)
analysis: which enables quantification of local variations
in uptake. Five textural metrics, corresponding to visual
properties of texture [19], were computed: (1) coarseness;
(2) contrast; (3) busyness; (4) complexity; (5) strength.
(d) The gray-level spatial-dependence (GLSD) matrix
(the co-occurrence matrix [8]). We considered twelve so-
called Haralick texture measures [20, 21]: (1) energy, (2)
entropy, (3) correlation, (4) contrast (also known as
inertia [22, 23]), (5) variance, (6) sum mean, (7)
agreement [24] (also known as Cohen’s [25] а-statistic),
(8) cluster shade, (9) cluster tendency (or prominence),
(10) homogeneity, (11) max probability, and (12) inverse
variance.
The latter approach (Haralick texture features; d) were
utilized in a recent study [9] for automatic detection of
PD. We supplement this with other methods, as outlined
above (a, b, c), and furthermore move beyond
classification tasks, to identify features that best capture
progression.
IV. MRI-DRIVEN SEGMENTATION AND
MAPPING
Both abovementioned overall frameworks (Sec. II and
III) utilized the following MRI-bases analyses:
1. Segment the high-resolution MRI images to obtain
the boundaries of the caudate and putamen (both left
and right), utilizing a multi-atlas segmentation
method [26].
2. Resample the SPECT image on the MRI grid
performing rigid mapping, using the FSL utility
FLIRT [27].
Figure 1 shows the output of a typical segmentation and
resampling of SPECT onto the MRI grid.
Following these steps, texture/shape analysis methods
of Sec. III were applied. For the normalization framework
(Sec. II), the following steps were also needed:
3. Use the LDDMM method to obtain templates, one
each for the caudate and putamen, and find the
optimal mappings from the template to each subjects.
4. Map the SPECT intensities for each subject to the
template coordinates using the mapping from 3.
Following this, we performed PCA on the mapped
SPECT intensities to obtain the regions with maximum
variation of the signal across subjects
replaced by more sophisticated non-linear
reduction techniques such as manifold lea
direction we intend to explore in the future
Figure 1: Same slice of the MRI (left) and SPECT
the cross-sections of the segmented caudate and
shown by curves.
V. RESULTS
We first demonstrate results for classvs. PD in Sec. V-A. However, the major
the present research effort is to provide
track progression of disease, as pursued in
A. Classification LDDMM normalization framework (S
shows the two-dimensional spaces obtai
which best distinguish the diseased fro
patients. In the case of the putamen, the fir
components best separated PDs from HCs
was not successful utilizing the caudate,
routine observations.
We note that in this analysis, we
affected side, i.e. with lower uptake, fo
This is consistent with the asymmetry of
the putamen, and formation of a gradient
the putamen. In fact, an approach in past w
subdivide the putamen into 2 or 3 sub-re
intermediate and posterior). In any case
moves beyond this conventional approa
PCA images that explain the variability b
populations (Figure 3).
Figure 2: A scatter plot of the putamen in the 2D
components (PC1 and PC3) showing HC (x and PD
s. PCA can be
r dimensionality
arning. This is a
e.
T image (right) with
d putamen surfaces
sification of HC
goal pursued in
methodology to
n Sec. V-B.
ec. II): Figure 2
ned from PCA,
om the healthy
rst two principal
s. The separation
consistent with
used the more
or each subject.
f uptake loss in
especially along
work has been to
egions (anterior,
e, our technique
ach by seeking
between subject
D space of principal
D (o).
Figure 3: (Left vs. Right) PCA loa
principal components (PC1 and PC
putamen. Darker regions show higher
therefore signify regions that correspo
the corresponding principal component
Shape/texture analysis (
application of two histogram
skewness and kurtosis, also sh
HC and PD, as seen in Fig. 4.
Figure 4: A scatter plot of two h
putamen (skewness and kurtosis) for PD
B. Progression It is likely that while utiliz
(of putman) is preferred for cl
using the less affected side
enhanced tracking of progr
following initial asymmetry in
side can more severely progre
such can provide a higher dyna
seen in Fig. 2-4 of [28]). In th
we did not detect a significant
when utilizing more affected
(thus results are only shown for
In our studies, we correl
against 7 different clinical mea
1) The unified Parkinson
(UPDRS)
2) Gait
3) Tremor
4) Bradykinesia
5) Rigidity
6) Disease duration (DD) –
adings corresponding to the two
3) plotted in Figure 2 for the
r loadings in absolute value, and
ond to the variance explained by
t.
Sec. III): An example
m-based metrics, namely
howed separation between
histogram-based features for the
D vs. HC.
zing the more affected side
lassification/diagnosis, that
may be the choice for
ression. This is because
n uptake, the less affected
ess over the years, and as
amic range (e.g. as we can
he present work, however,
t difference in performance
side vs. less affected side
r one case).
ated image-based metrics
sures::
n's disease rating scale
– from diagnosis
7) DD – from symptoms
Items (2-5) are motor sub-scores within the UPDRS
motor score.
We found that 16, of the various metrics considered,
depicted more substantial correlations against clinical
scores than other metrics did, thus we limit our summary
to these in what follows. The 16 metrics are as follows:
1) Conventional mean (normalized to mean reference
uptake in the occipital cortex).
2) 3D-MIs: B3.
3,4,5) Histogram-based: standard deviation; skewness;
Kurtosis.
6,7,8) NGTD: Coarseness; Busyness; Strength.
9,10,11,12) GLSD (haralick): Energy; Homogenuity;
Cluster-Shade; Agreement.
13,14,15,16) PCA analysis: First four PC components.
Fig. 5 depicts, for n=128 subjects, the performance of
each of the 16 abovementioned metrics in relation to the
UPDRS score. On top of each plot three correlation
indices are reported: i) Pearson (p), ii) Kendall (k), and
Spearman (s). The first one assumes linear correlation,
while the latter two are non-parametric.
It is seen that conventional mean uptake (normalized
to reference region) results in high correlation (e.g. p=-
0.65), and the two non-conventional metrics comparable
to it are skewness (p=0.52) and Cluster-Shade (p=0.59).
Fig. 6 depicts a summary of correlation values (p) for
a figure such as Fig. 5, in the case of all seven clinical
measures. It is seen that along with conventional mean
uptake metric (#1), other texture/shape metrics also show
considerable correlations, including metrics #4 and #12,
namely skewness and Cluster-Shade.
Fig. 7 is very similar to Fig. 5, except that PD-only
subjects (n=80) are only considered. It is seen that the
correlations with UPDRS score are substantially
dampened. This means that HC vs. PD discrimination
plays a considerable role in introducing correlations, and
sheds light on the fact that prognosis of PD-only subjects
is a difficult task. Similar plots as Fig. 7 for PD-only
subjects are shown in Fig. 8, this time correlating imaging
to DD (from symptoms), instead of UPDRS. Finally, Fig.
9 summarizes correlation values (p) in the PD-only case
for all seven clinical measures.
It is seen that some metrics may outperform
conventional mean uptake metric when applied to PD-
only subjects in terms of correlation with clinical
measures. In any case, the correlations are seen to be
weak in this case. We believe this may be enhanced (as
we pursue in ongoing work) by performing more
effective SPECT to MRI mappings. This is a challenging
problem, given the blurred nature of SPECT and the
concentrated nature of uptake in DAT imaging.
Another logical step is to consider multi-metric
analysis. We performed preliminary analysis in this
context, using the multiple correlation coefficient R, in
the case of conventional mean vs. feature-of-interest (e.g.
cluster-shade) vs. their joint combination. Here are some
results:
PD+HC subjects - DD (symptoms): Conventional -
0.65; Cluster-Shade - 0.59; Combined: 0.70.
PD only - DD (diagnosis): Conventional - 0.04;
Strength - 0.29; Combined - 0.32.
PD only - DD (symptoms): Conventional - 0.13;
Strength - 0.09; Combined - 0.25.
PD only - UPDRS: Conventional - 0.05; Strength -
0.23; Combined - 0.27.
Thus, it is seen that use of multi-metric analysis,
incorporating conventional mean uptake in addition to
novel shape/texture analysis can enhance the ability to
track disease via imaging.
Figure 5: Plots of metric values vs. UPDRS score, for 16 different metrics as described in the text. The first 12 are shape/texture metrics applied to the
original patient-space, while the last 4 are PCA components as generated in the normalized/template space. HC+PD subjects were considered.
Figure 6: Plots of correlation (y-axis) vs. metric of interest (16 different metrics on the x-axis) for each of the 7 clinical measures as mentioned in the
text. HC+PD subjects were considered.
Figure 7: Plots of metric values vs. UPDRS score, for 16 different metrics as described in the text. Here, PD-only subjects were considered.
Figure 8: Plots of metric values vs. DD (from symptoms), for 16 different metrics as described in the text. Here, PD-only subjects were considered.
Figure 9: Plots of correlation (y-axis) vs. metric of interest (16 different metrics on the x-axis) for each of the 7 clinical measures as mentioned in the
text. PD-only subjects were considered.
VI. MORE DISCUSSION AND CONCLUSION
We also note that the normalization framework
enables the performance of an additional kind of analysis.
This work is in progress, but we show a sample analysis:
since all structures are registered to a common template,
it is now possible to ask which voxels correlate most with
a given clinical score. An example of this is shown in Fig.
10 for the particular case of bradykinesia (motor sub-
score within the UPDRS). This approach can be
performed for different clinical scores, and also for the
caudate, in addition to the putamen, and has the potential
to open-up new understandings about sub-regional
involvements with clinical manifestations in PD.
Overall, the presented approach of utilizing novel
pattern recognition as applied to imaging of PD could
have important clinical implications. Potentially enhanced
sensitivity to subtle neuroanatomical changes is expected
to provide novel insights into the relationship between
dopaminergic alterations and PD manifestations, while
extending the clinical usefulness of this imaging
technique. This approach can also be applied to images
from subjects at increased risk of PD (e.g. mutation
carriers or subjects with rapid-eye-movement sleep
behavior disorder) in an attempt to discern dopaminergic
patterns that might be involved in pathogenesis and to
assess the impact of novel disease modifying therapies.
Figure 10: Plots of correlation vs. Bradykineasia (motor-subscore of the
UPDRS) in the normalized template for n=128.
ACKNOWLEDGEMENTS
The authors wish to gratefully acknowledge grant support
from the Michael J. Fox Foundation (RRIA2013; PIs: A.
Rahmim, V. Sossi).
REFERENCES
1. Kupsch, A.R., et al., Impact of DaTscan SPECT imaging on
clinical management, diagnosis, confidence of diagnosis, quality
of life, health resource use and safety in patients with clinically
uncertain parkinsonian syndromes: a prospective 1-year follow-up
of an open-label controlled study. Journal of Neurology
Neurosurgery and Psychiatry, 2012. 83(6): p. 620-628.
2. Grachev, I., et al., Impact of DaTscan (TM) SPECT Imaging on
Clinical Management, Diagnosis, and Confidence of Diagnosis in
Patients with Clinically Uncertain Parkinsonian Syndromes: A
Prospective 1-Year Follow-Up Study. Neurology, 2012. 78.
3. Catafau, A.M. and E. Tolosa, Impact of dopamine transporter
SPECT using I-123-ioflupane on diagnosis and management of
patients with clinically uncertain Parkinsonian syndromes.
Movement Disorders, 2004. 19(10): p. 1175-1182.
4. Djang, D.S.W., et al., SNM Practice Guideline for Dopamine
Transporter Imaging with I-123-Ioflupane SPECT 1.0. Journal of
Nuclear Medicine, 2012. 53(1): p. 154-163.
5. Badiavas, K., et al., SPECT imaging evaluation in movement
disorders: far beyond visual assessment. European journal of
nuclear medicine and molecular imaging, 2011. 38(4): p. 764-773.
6. Koch, W., et al., Clinical testing of an optimized software solution
for an automated, observer-independent evaluation of dopamine
transporter SPECT studies. Journal of Nuclear Medicine, 2005.
46(7): p. 1109-1118.
7. Nandhagopal, R., et al., Progression of dopaminergic dysfunction
in a LRRK2 kindred A multitracer PET study. Neurology, 2008.
71(22): p. 1790-1795.
8. Sossi, V., et al., Dopamine Turnover Increases in Asymptomatic
LRRK2 Mutations Carriers. Movement Disorders, 2010. 25(16): p.
2717-2723.
9. Martinez-Murcia, F.J., et al., Parametrization of textural patterns in
I-123-ioflupane imaging for the automatic detection of
Parkinsonism. Medical Physics, 2014. 41(1).
10. Parkinson's Progression Markers Initiative. Available from:
http://www.ppmi-info.org/.
11. Sossi, V., et al., New insights into levodpoa induced dopamine
release in Parkinson's disease. J. Cereb. Blood Flow & Metab.,
2012. 32 (suppl. 1): p. P114.
12. Gonzalez, M., et al., Novel spatial analysis method for PET data
using 3D moment invariants: applications to Parkinson's disease. J.
Cereb. Blood Flow & Metab., 2012. 32 (suppl. 1): p. P117.
13. Gonzalez, M.E., et al., Novel spatial analysis method for PET
images using 3D moment invariants: Applications to Parkinson's
disease. Neuroimage, 2013. 68: p. 11-21.
14. Rahmim, A., et al., Novel parametric PET image quantification
using texture and shape analysis. Proc. IEEE Nucl. Sci. Symp.
Conf. , 2012: p. 2227-2230.
15. Blinder, S., et al., Texture and Shape Analysis on High and Low
Spatial Resolution Emission Images. Proc. IEEE Nucl. Sci. Symp.
Conf. , 2014.
16. Vaillant, M. and J. Glaunès, Surface Matching via Currents, in
Information Processing in Medical Imaging, G. Christensen and
M. Sonka, Editors. 2005, Springer Berlin Heidelberg. p. 381-392.
17. Ma, J., M.I. Miller, and L. Younes, A Bayesian Generative Model
for Surface Template Estimation. International Journal of
Biomedical Imaging, 2010. 2010: p. 14.
18. El Naqa, I., et al., Exploring feature-based approaches in PET
images for predicting cancer treatment outcomes. Pattern
Recognition, 2009. 42(6): p. 1162-1171.
19. Amadasun, M. and R. King, Textural Features Corresponding to
Textural Properties. Ieee Transactions on Systems Man and
Cybernetics, 1989. 19(5): p. 1264-1274.
20. Haralick, R.M., Shanmuga.K, and I. Dinstein, Textural Features
for Image Classification. Ieee Transactions on Systems Man and
Cybernetics, 1973. Smc3(6): p. 610-621.
21. Conners, R.W., M.M. Trivedi, and C.A. Harlow, Segmentation of
a High-Resolution Urban Scene Using Texture Operators.
Computer Vision Graphics and Image Processing, 1984. 25(3): p.
273-310.
22. Conners, R.W. and C.A. Harlow, Toward a Structural Textural
Analyzer Based on Statistical-Methods. Computer Graphics and
Image Processing, 1980. 12(3): p. 224-256.
23. Oh, G., S. Lee, and S.Y. Shin, Fast determination of textural
periodicity using distance matching function. Pattern Recognition
Letters, 1999. 20(2): p. 191-197.
24. Parkkinen, J., K. Selkainaho, and E. Oja, Detecting Texture
Periodicity from the Cooccurrence Matrix. Pattern Recognition
Letters, 1990. 11(1): p. 43-50.
25. Cohen, J., A Coefficient of Agreement for Nominal Scales.
Educational and Psychological Measurement, 1960. 20(1): p. 37-
46.
26. Tang, X.Y., et al., Bayesian Parameter Estimation and
Segmentation in the Multi-Atlas Random Orbit Model. Plos One,
2013. 8(6).
27. Jenkinson, M. and S. Smith, A global optimisation method for
robust affine registration of brain images. Medical Image Analysis,
2001. 5(2): p. 143-156.
28. Nandhagopal, R., et al., Longitudinal progression of sporadic
Parkinson's disease: a multi-tracer positron emission
tomography study. Brain, 2009. 132(11): p. 2970-2979.