Texture Features Based Detection of Parkinson’s
Disease on DaTSCAN images
F J Martinez-Murcia
1
, J M G´orriz
1
, J Ram´ırez
1
, I A Ill´an
1
, C G Puntonet
2
,
and the Parkinson’s Progression Markers Initiative
⋆
1
Department of Signal Theory, Networking and Communications,
Universidad of Granada, Spain.
2
Department of Computer Architecture and Technology,
Universidad de Granada, Spain.
Abstract.
In this work, a novel approach to Computer Aided Diagnosis
(CAD) system for the Parkinson’s Disease (PD) is proposed. This tool is
intended for physicians, and is based on fully automated methods that
lead to the classification of Ioflupane/FP-CIT-I-123 (DaTSCAN) SPECT
images. DaTSCAN images from the Parkinson Progression Markers Ini-
tiative (PPMI) are used to have in vivo information of the dopamine
transporter density. These images are normalized, reduced (using a mask),
and then a GLC matrix is computed over the whole image, extracting
several Haralick texture features which will be used as a feature vector
in the classification task. Using the leave-one-out cross-validation tech-
nique over the whole PPMI database, the system achieves results up to
a 95
.9% of accuracy, and 97.3% of sensitivity, with positive likelihood ra-
tios over 19, demonstrating our system’s ability on the detection of the
Parkinson’s Disease by providing robust and accurate results for clinical
practical use, as well as being fast and automatic.
Keywords:
Parkinson’s Disease, DaTSCAN images, Computer Aided
Diagnosis, Haralick Texture Features, Support Vector Machines, Super-
vised Learning
1
Introduction
Parkinsonian Syndrome (PS), also known as Parkinsonism, is a neurological syn-
drome characterized by tremor, hypokinesia, rigidity and postural instability [3].
It is considered as the second most common neurodegenerative disease, with a
prevalence of 1-3% in the population over 65 years of age [11]. A wide range of
etiologies may lead to the PS, while the most common cause is the neurodegen-
erative condition called Parkinson’s Disease (PD). This disease originates due to
⋆
Data used in the preparation of this article were obtained from the Parkinson’s Pro-
gression Markers Initiative (PPMI) database (www.ppmi-info.org/data). As such,
the investigators within PPMI contributed to the design and implementation of
PPMI and/or provided data but did not participate in the analysis or writing of this
report. PPMI investigators include (complete listing at PPMI site).
the progressive loss of dopaminergic neurons of the nigrostriatal pathway, which
connects the substantia nigra to the striatum. As a result, the dopamine content
of the striatum decreases, and consequently, dopamine transporters (DAT) are
lost. Other possible causes include some toxins, a few metabolic diseases, and a
handful of non-PD neurological conditions [2].
As the PD is related to a loss of dopamine transporters in the nigrostriatal
pathway, the study of its status by means of brain imaging techniques has been
suggested to increase the diagnostic accuracy in the case of parkinsonian syn-
dromes [3]. Ioflupane/FP-CIT-I-123 (better known as DaTSCAN) is a tracer
that binds to the dopamine transporters in the striatum, allowing the obtention
of Single Photon Emission Computed Tomography (SPECT) images that show
a reduced uptake of the tracer in the striatum in patients with PS [1].
A wide range of supervised learning techniques have been combined to gener-
ate Computer Aided Diagnosis (CAD) systems that allow to detect neurodegen-
erative diseases, such as Alzheimer’s Disease [9] or Parkinson [16]. Techniques
range from the use of selection of Regions of Interest (ROIs) [5], or Single Value
Decomposition strategies (SVD) [14] to more complex approaches such as Empir-
ical Mode Decomposition (EMD) combined with Principal Component Analysis
(PCA) combined method in [13].
In this work we use the Haralick Texture analysis, proposed in [4], that
provides several texture features, which we will use to characterize some patterns
of the Parkinson’s Disease. To do so, we calculate a 3D Gray Level Co-occurrence
(GLC) matrix [12], required for the computation of these features. Finally we
make use of a Support Vector Machine (SVM) [18] binary classifier to test the
ability of these features in the PD pattern detection.
2
Methodology
2.1
Test data and preprocessing
Data used in the preparation of this article were obtained from the Parkinson’s
Progression Markers Initiative (PPMI) database (www.ppmi-info.org/data).
For up-to-date information on the study, visit www.ppmi-info.org.
The images in this database were imaged 4 ± 0.5 hours after the injection
of between 111 and 185 MBq of DaTSCAN. Subjects were also pretreated with
saturated iodine solution (10 drops in water) or perchlorate (1000 mg) prior to
the injection. All subjects had a supplied
57
Co line marker affixed along the
canthomeatal line, which will facilitate subsequent image processing and allow
the core lab to accurately distinguish left and right in the face of multiple image
file transfers. These markers are only evident in the
57
Co window and hence do
not contaminate the
123
I-DaTSCAN brain data [8, 6].
Raw projection data are acquired into a 128 × 128 matrix stepping each 3
degrees for a total of 120 projection into two 20% symmetric photopeak windows
centered on 159 KeV and 122 KeV with a total scan duration of approximately
30 - 45 minutes. Other scan parameters (collimation, acquisition mode, etc) are
selected for each site. The images of both the subject’s data and the cobalt stri-
atal phantom are reconstructed and attenuation corrected, implementing either
filtered back-projection or an iterative reconstruction algorithm using standard-
ized approaches [6]. After the processing, images used are spatially and intensity
normalized, and of a 91 × 109 × 91 size.
All images in the databases have been spatially normalized (using the SPM8
software and a custom DaTSCAN template -see Fig. 1, which depicts some
cuts of both normal and affected subjects-) and intensity normalized using the
Integral Normalization
algorithm. This method is based on the obtainment
of an intrinsic parameter from the image, I
p
, and the estimation of the binding
activity as:
t
′
= t/I
p
(1)
where t denotes the spatially normalized image, and t
′
the image normalized
spatially and in intensity. In this case, all intensity values on the image are
summed as an aproximation of the expression I
p
=
t, which results in an
integral value of the intensity.
Axial
20
40
60
20
40
60
80
Sagital
20
40
60
20
40
60
80
Coronal
20
40
60
10
20
30
40
50
60
(a) PD subject (PPMI)
Axial
20
40
60
20
40
60
80
Sagital
20
40
60
20
40
60
80
Coronal
20
40
60
10
20
30
40
50
60
(b) Normal Control (PPMI)
Fig. 1: Sample image from (a) a PD patient from PPMI database and (b) a
healthy subject from PPMI database.
2.2
Mask
At this point, all the images are spatially and intensity normalized. These im-
ages feature two major changes in their intensity levels: the change between the
noisy background and the whole brain, and the increase difference between brain
intensity levels and those of the striatum.
We assume that the change between the background and the brain should
not be significative enough for the diagnosis of the PD, though its influence on
the texture of the image is clear. Given this assumption, it would be desirable
to remove all background pixels from the images. To do so, we use a masking
process
.
As the computation of the GLC matrix (which will be explained more exten-
sively in the following section) needs an cuboid image, our purpose is to extract
the biggest box which contains only brain pixels. Thus, we establish an intensity
threshold, I
th
, and extract a mask from the mean image of all images in the
database using:
M ask = I > I
th
(2)
Then, the coordinates of the largest box that fits into that mask is obtained,
and this area is selected in all images. We have used intensity thresholds ranging
from 0% (all image is selected) to 50% of the highest intensity of the mean image.
2.3
Haralick Texture Features
A co-occurrence matrix is a matrix that is defined over an image to be the distri-
bution of co-occurring values at a given offset. Mathematically, a co-occurrence
matrix C is defined over an n × m image I, parameterized by an offset (∆x,∆y),
as:
C
∆x,∆y
(i, j) =
n
p=1
m
q=1
1,
if I(p, q) = i and I(p + ∆x, q + ∆y) = j
0,
otherwise
(3)
Note that the (x,y) parameterization makes the co-occurrence matrix sen-
sitive to rotation. We choose one offset vector, so a rotation of the image not
equal to 180 degrees will result in a different co-occurrence distribution for the
same (rotated) image. This is rarely desirable in the applications co-occurrence
matrices are used in, so the co-occurrence matrix is often formed using a set of
offsets sweeping through 180 degrees (i.e. 0, 45, 90, and 135 degrees) at the same
distance to achieve a degree of rotational invariance.
The method used here to expand the co-occurrence matrix to a tridimensional
space is defined in [12], introducing another variable in Eq. 3. A 3D co-occurrence
matrix C is defined over an n × m × k three-dimensional image I, parameterized
by an offset (∆x,∆y, ∆z), as:
C
∆
(i, j, k) =
n
p=1
m
q=1
k
r=1
1,
if I(p) = i and I(p + ∆) = j
0,
otherwise
(4)
where
p
= (p, q, r), and ∆ = (∆x, ∆y, ∆z)
(5)
Twelve Haralick texture features [4] are then extracted from the GLC ma-
trix computed in Eq. 4, in this order: Energy, Entropy, Correlation, Contrast,
Variance, Sum Average, Inertia, Cluster shade, Cluster prominence, Homogene-
ity, Maximum probability and Inverse variance. As 13 spatial directions and 10
distances are considered, a total number of 130 co-ocurrence matrices are com-
puted, and therefore, 130 values for each of the Haralick texture features have
been computed. These values have been used as an input vector to the following
classifier.
Fig. 2: Spatial representation of the thirteen direction vectors used to compute
the thirteen different GLC matrices.
2.4
Classifier
The classification step is performed as follows. A predictive model is derived
from a set of training data from two different classes -in our case, patients with-
out dopaminergic deficit (NOR) and Parkinson’s Disease (PD)-, and the test
image is then classified. To build our predictive model, we make use of the Sup-
port Vector Machines paradigm. Support Vector Machine (SVM) [17] is a recent
class of statistical classification and regression techniques playing an increasing
role in applications to detection problems in various engineering problems, no-
tably in statistical signal processing, pattern recognition, image analysis [9], and
communication systems. SVM with linear discriminant functions define decision
hypersurfaces or hyperplanes in a multidimensional feature space, that is:
g(x) = w
T
x
+ ω
0
= 0,
(6)
where w is known as the weight vector and ω
0
as the threshold. The weight
vector w is orthogonal to the decision hyperplane and the optimization task
consists of finding the unknown parameters ω
i
, i = 1, ..., n defining the decision
hyperplane.
Let x
i
, i = 1, 2, ..., n be the feature vectors of the training set, X. These belong
to either ω
1
or ω
2
, the two classes. If the classes were linearly separable, the
objective would be to design a hyperplane that classifies correctly all the training
vectors. Among the different design criteria, the maximal margin hyperplane is
usually selected since it leaves the maximum margin of separation between the
two classes. Since the distance from a point x to the hyperplane is given by
z = |g(x)|/ w , scaling w and w
0
so that the value of g(x) is +1 for the nearest
point in ω
1
and −1 for the nearest points in ω
2
, the optimization problem is
reduced to minimizing a cost function J(ω) = 1/2||ω||
2
subject to:
f
svm
(x) =
N
S
i=1
α
i
ω
i
Φ(s
i
) · Φ(x) + ω
0
(7)
where α
i
are the solution of a quadratic optimization problem that is usually
determined by quadratic programming or the well-known sequential minimal op-
timization algorithm, and Φ(s) or Φ(x) denote the transformation of the feature
vectors into the effective feature space. This basic SVM classifier produces a
linear separation hyperplane.
SVM classifiers with linear kernels are based on a solid theoretical background
thus leading to reproducible performance and finally, showing a good robustness
to noisy or mislabeled data. Therefore, they are usually applied to evaluate the
separability of different features, providing low generalization error even with
small learning sample datasets.
3
Results and Discussion
To perform an evaluation of our proposed CAD system, we have independently
evaluated the effect of using each of the 12 Haralick Texture Features. Thus,
regarding their ability to correctly interpret the texture of the normalized DaT-
SCAN image. All directions are considered but the influence of the distance d
at whith the GLC matrix is calculated has been also considered, using all the
feature values extracted in a range of 1 < d distance from the central voxel.
In each of these experiments, the images are previously reduced using masks,
with an specific intensity threshold I
th
starting at 0% (whole image) up to 50%
of the highest intensity value, as commented in Section 2.2.
3.1
Evaluation
The proposed methodology has been tested on the PPMI databse (see Sec. 2.1).
We have used a cross-validation method called leave-one-out to extract some
evaluation parameters which will allow us to compare and evaluate the perfor-
mance of these proposed systems. Parameters like accuracy, sensitivity, speci-
ficity, Positive Likelihood (PL) and Negative Likelihood (NL) ratios have been
estimated using this method.
Leave-one-out method provides us with a mean of using almost all images for
the training of the classifier and still get an unbiased error estimate [7]. However
this estimate might be affected by the database topology and the classifier used.
Anyway, this is one of the most used methods for system validation, and so will
be used in this work.
The accuracy, sensitivity and specificity parameters are calculated as:
Acc =
T P + T N
T P + T N + F P + F N
, Sens =
T P
T P + F N
, Spec =
T N
T N + F P
(8)
where TP, TN, FP and FN are the number of true positives, true negatives, false
positives and false negatives, respectively. Accuracy measures the proportion of
correctly classified samples. Sensitivity and specificity are used to measure the
proportion of actual positives or negatives which are identified correctly (e.g. the
percentage of PS patients, or normal controls who are identified as such).
Although sensitivity and specificity are very important to reveal the abil-
ity of a system on detecting PD/NOR patterns, they are prevalence-dependent.
This means that in an positive prevalent database (those where there is a higher
number of positives), sensitivity will be higher, and similarly occurs with speci-
ficity. So even if they are a good estimate of the goodness of the classifier, other
parameters such as Positive and Negative Likelihood ratios (PL and NL), which
are prevalence independent, are computed.
P L = Sensitivity/(1 − Specificity)
(9)
N L = (1 − Sensitivity)/Specificity
(10)
These parameters are also widely used in clinical medicine, where values of PL
greater than 5 or NL values less than 0.2 can be applied to the pre-test probabil-
ity of a patient having the disease tested to estimate a post-test probability of the
disease state existing [10]. A positive result for a test with PL of 8 adds approx-
imately 40% to the pre-test probability that a patient has a specific diagnosis.
These parameters are computed with Eq. 9 and 10.
3.2
Results and discussion
In this work, we have tested the performance of a CAD system which makes use
of the Haralick Texture Features to extract relevant textural information from
images, and then using these features to classify each image. We have used a
mask to reduce dimensionality and improve the GLC matrix computation, and
the whole system has been evaluated using the widely available PPMI database,
to better facilitate the reproducibility of results.
As commented in Section 2.2, a mask has been added to the normalized
images, to extract smaller box-shaped images that allows us to both perform
the following tests more efficiently and optimize the calculation of the GLC
matrix by removing all non-brain pixels from the source image. Therefore, this
masking step might seem profitable.
However, it shows one major drawback: its dependence on the manual oper-
ation of setting the mask threshold. Although large number of border detection
algorithms exist, and most of them could be applied here to select the brain area,
and thus, the intensity threshold, in this work we have established some manual
thresholds -see Sec. 2.2- for better evaluating the performance of our system.
To better illustrate the influence of using masks, Table 1 shows the accuracy
results obtained by our system using either no mask (mask = 0% ) or different
levels for the intensity threshold I
th
.
Most of the best values are obtained using the Homogeneity as an input
vector, except for when I
th
> 0.35 ∗ I
max
. This can be easily explained due to
the image characteristics. As we commented before, one of the reasons of using
masks was to select the biggest image box that contains only brain pixels. This is
achieved in these integral-normalized images when the intensity threshold values
are around this percentage of the maximum intensity of the image. Therefore,
more automatic methods to detect this threshold in all types of images should be
Feature
0
5%
15%
25%
35%
45%
Energy
0.822
0.822
0.784
0.862
0.929
0.952
Entropy
0.807
0.807
0.792
0.881
0.933
0.914
Correlation
0.851
0.851
0.870
0.929
0.933
0.955
Contrast
0.818
0.818
0.833
0.937
0.948
0.937
Variance
0.762
0.762
0.784
0.825
0.888
0.914
Sum Average
0.810
0.810
0.844
0.900
0.937
0.933
Inertia
0.773
0.773
0.810
0.833
0.937
0.941
ClusterShade
0.818
0.818
0.833
0.937
0.948
0.937
Cluster Tendency
0.926
0.926
0.911
0.941
0.933
0.918
Homogeneity
0.926
0.926
0.929
0.944
0.952
0.907
MaxProbability
0.766
0.766
0.747
0.770
0.818
0.814
InverseVariance
0.743
0.743
0.796
0.870
0.874
0.914
Table 1: Accuracy results for each of the 12 Haralick texture features (see Sec.
2.3), using different intensity thresholds (in %).
desirable. In Figure 3, the resulting area of using an I
th
= 0.30I
max
is depicted,
applied to the same subject on Fig. 1b. It is interesting to note that our goal of
selecting only internal brain voxels has been achieved with this threshold.
Axial
10
20
30
5
10
15
20
25
30
35
Sagital
10
20
30
40
50
5
10
15
20
25
30
35
Coronal
10
20
30
40
50
5
10
15
20
25
30
Fig. 3: Resulting image box when using a mask with a I
th
= 0.30I
max
.
As commented before, all features improve their accuracy results as the in-
tensity threshold increases, but there are some of them that perform particularly
well in this task, e.g. Homogeneity, Cluster Shade or Energy. Five of these best
features are depicted on Fig. 4. It is noticeable that, while some features increase
their performance significantly once I
th
exceeds some values (which we can con-
sider as an indication that only brain voxels have been selected), there are others
that offer good values almost independently of the chosen value for I
th
.
We have already focused on the performance of the system depending on
the mask applied and the feature used for classifying. Nevertheless, there is
one more parameter on the system that can affect the results: the distance d
at which the GLC matrix is calculated. To perform a deeper analysis of this
parameter, we have evaluated the behavior of the system using only the 13 values
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Accuracy over I
th
I
th
Accuracy
Homogeneity
Correlation
Energy
Cluster Shade
Cluster Tendency
Fig. 4: Accuracy obtained by our proposed system using each of the texture
features listed in the legend, using different values for I
th
.
of each Haralick texture feature extracted from the 13 GLC matrices (one in each
direction) computed at a distance d of the central voxel, with 1 < d < 10. When
using each of the five aforementioned texture features, the performance results
on our system are displayed on Fig. 5. Notice that, once a maximum is achieved
generally at d = 7, 8, the accuracy decreases, what might point out the softness
degree of the texture pattern of these images.
1
2
3
4
5
6
7
8
9
10
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Distance
Accuracy
Accuracy vs Distance
Homogeneity
Cluster Shade
Cluster tendency
Energy
Correlation
Fig. 5: Accuracy obtained by our proposed system using only the features ex-
tracted from the 13 GLC matrices (one in each direction) computed at a distance
d of the central voxel.
In Table 2 we compare this method with others from the bibliography and
the Voxels-as-Features approach, commonly used as an estimation of the per-
formance of visual analysis performed by experts [15]. Some of the methods
evaluated include a combination of intensity normalization strategies and clas-
sifiers (VAF-IN) [5], a SVD approach that independently decomposes each side
of the brain (as PD often show asymmetrical dopamine deficit) [14] and a EMD
of the images, then modeled using PCA or ICA to classify the images [13].
System
Acc
Sens
Spec
PL
NL
Homogeneity
0.959
0.973
0.949
19.22
0.028
Cluster Shade
0.955
0.964
0.949
19.01
0.038
Cluster Tendency
0.955
0.973
0.943
17.10
0.029
Correlation
0.941
0.946
0.937
14.92
0.058
Energy
0.937
0.964
0.918
11.73
0.039
VAF
0.840
0.807
0.862
5.88
0.224
VAF-IN
0,913
0.890
0.932
13.08
0.118
SVD
0.940
0.962
0.918
11.73
0.041
EMD-IMF3
0.950
0.951
0.948
18.28
0.051
Table 2: Comparison of our proposed system (using different texture features)
and some other methods in the bibliography: VAF system using the intensity-
normalized images, a combination of intensity normalization strategies and clas-
sifiers (VAF-IN) [5], a SVD-based approach [14] and EMD using the third inde-
pendent mode function (IMF3) [13].
Our method based on the Haralick Texture Features and a mask obtains
better results in almost every case (specially when using one of the three best
features: Homogeneity, Cluster Shade or Cluster Tendency), which also show a
high independence of the value used for the intensity threshold of the mask (I
th
).
All methods clearly outperforms the VAF approach, and only the EMD based
method obtains similar values to our system. In the latter case, our system shows
a clear advantage: its simplicity in terms of computation. While we make use of
the directly-computed texture features from the GLC matrix, after a reduction
of the image using a mask, the EMD-based method combines gaussian filtering,
a Multidimensional Ensemble EMD extraction of some slices, feature reduction
using PCA and classification using SVM.
4
Conclussion
The development of new computer-based diagnosis software is a promising area
of research. The proposed system aims to provide a fully automated method
for physicians to help them in the diagnosis task of Parkinson’s Disease (PD),
eliminating expensive manual operations, in the sense of requiring an expertise
degree of the operator, as well as reducing time costs. Moreover, as the pro-
cess is automatic, it might not suffer from the pitfalls of investigator-dependent
methods.
The presented work makes use of several widely known techniques that,
combined, demonstrate their ability in the detection of some PD patterns in
DaTSCAN imaging. Particularly, the use of a mask to select subimages that
contains only brain voxels has a great impact on the computation of a GLC
matrix, facilitating a subsequent texture analysis. Therefore, our system that
combines a voxel selection based on mask and the analysis of the textural fea-
tures of the image demonstrates its ability on the detection of PD patterns in
DaTSCAN imaging, providing robust and accurate results up to a 95.9% of ac-
curacy, and 97.3% of sensitivity, with a Positive likelihood ratio over 19, a very
good indicator of its robustness on PD detection.
The results hereby presented are very promising, so new approaches to the
usage of textural information for the pattern characterization of neurodegener-
ative disorders can be made.
Acknowledgements
PPMI –a public-private partnership– is funded by The Michael J. Fox Founda-
tion for Parkinson’s Research and funding partners, including Abbott, Biogen
Idec, F. Hoffman-La Roche Ltd., GE Healthcare, Genentech and Pfizer Inc.
This work was partly supported by the MICINN under the TEC2008-02113
and TEC2012-34306 projects and the Consejer´ıa de Innovaci´
on, Ciencia y Em-
presa (Junta de Andaluc´ıa, Spain) under the Excellence Projects P07-TIC-02566,
P09-TIC-4530 and P11-TIC-7103.
References
1. Bhidayasiri, R.: How useful is (123I) beta-CIT SPECT in the diagnosis
of parkinson’s disease? Reviews in Neurological Diseases 3(1), 19–22 (2006),
http://www.ncbi.nlm.nih.gov/pubmed/16596082, PMID: 16596082
2. Christine,
C.W.,
Aminoff,
M.J.:
Clinical
differentiation
of
parkin-
sonian
syndromes:
Prognostic
and
therapeutic
relevance.
The
American
Journal
of
Medicine
117(6),
412
–
419
(2004),
http://www.sciencedirect.com/science/article/pii/S0002934304003626
3. Eckert, T., Edwards, C.: The application of network mapping in differential di-
agnosis of parkinsonian disorders. Clinical Neuroscience Research 6(6), 359 – 366
(2007),
http://www.sciencedirect.com/science/article/pii/S1566277207000023,
neural Networks in the Imaging of Neuropsychiatric Diseases
4. Haralick, R., Shanmugam, K., Dinstein, I.: Textural features for image classifica-
tion. IEEE Transactions on Systems, Man and Cybernetics 3(6), 610–621 (1973)
5. Ill´
an, I., G´
orriz, J., Ram´ırez, J., Segovia, F., Jim´enez-Hoyuela, J., Ortega Lozano,
S.: Automatic assistance to parkinsons disease diagnosis in datscan spect imaging.
Medical Physics 39(10), 5971–5980 (2012)
6. Initiative, T.P.P.M.: PPMI. Imaging Technical Operations Manual, 2 edn. (June
2010)
7. Kohavi, R.: A study of cross-validation and bootstrap for accuracy estimation
and model selection. In: Proceedings of International Joint Conference on AI. pp.
1137–1145 (1995), http://citeseer.ist.psu.edu/kohavi95study.html
8. Marek, K., Jennings, D., Lasch, S., Siderowf, A., Tanner, C., Simuni, T., Coffey,
C., Kieburtz, K., Flagg, E., Chowdhury, S., Poewe, W., Mollenhauer, B., Klinik,
P., Sherer, T., Frasier, M., Meunier, C., Rudolph, A., Casaceli, C., Seibyl, J., Men-
dick, S., Schuff, N., Zhang, Y., Toga, A., Crawford, K., Ansbach, A., De Blasio, P.,
Piovella, M., Trojanowski, J., Shaw, L., Singleton, A., Hawkins, K., Eberling, J.,
Brooks, D., Russell, D., Leary, L., Factor, S., Sommerfeld, B., Hogarth, P., Pighetti,
E., Williams, K., Standaert, D., Guthrie, S., Hauser, R., Delgado, H., Jankovic,
J., Hunter, C., Stern, M., Tran, B., Leverenz, J., Baca, M., Frank, S., Thomas, C.,
Richard, I., Deeley, C., Rees, L., Sprenger, F., Lang, E., Shill, H., Obradov, S., Fer-
nandez, H., Winters, A., Berg, D., Gauss, K., Galasko, D., Fontaine, D., Mari, Z.,
Gerstenhaber, M., Brooks, D., Malloy, S., Barone, P., Longo, K., Comery, T., Rav-
ina, B., Grachev, I., Gallagher, K., Collins, M., Widnell, K.L., Ostrowizki, S., Fon-
toura, P., Ho, T., Luthman, J., Brug, M.v.d., Reith, A.D., Taylor, P.: The parkin-
son progression marker initiative (PPMI). Progress in Neurobiology 95(4), 629–635
(Dec 2011), http://www.sciencedirect.com/science/article/pii/S0301008211001651
9. Mart´ınez-Murcia, F., G´
orriz, J., Ram´ırez, J., Puntonet, C., Salas-Gonz´
alez, D.:
Computer aided diagnosis tool for Alzheimer’s disease based on Mann-Whitney-
Wilcoxon U-test. Expert Systems with Applications 39(10), 9676–9685 (August
2012)
10. McGee, S.: Simplifying likelihood ratios. Journal of General Internal Medicine
17(8), 646–9 (August 2002)
11. Moghal, S., Rajput, A.H., D’Arcy, C., Rajput, R.: Prevalence of movement dis-
orders in elderly community residents. Neuroepidemiology 13(4), 175–178 (1994),
http://www.ncbi.nlm.nih.gov/pubmed/8090259, PMID: 8090259
12. Philips, C., Li, D., Raicu, D., Furst, J.: Directional invariance of co-occurrence
matrices within the liver. International Conference on Biocomputation, Bioinfor-
matics, and Biomedical Technologies 0, 29–34 (2008)
13. Rojas, A., G´
orriz, J., Ram´ırez, J., Ill´
an, I., Mart´ınez-Murcia, F., Ortiz,
A., R´ıo, M.G., Moreno-Caballero, M.: Application of empirical mode de-
composition (emd) on datscan spect images to explore parkinson dis-
ease.
Expert
Systems
with
Applications
40(7),
2756
–
2766
(2013),
http://www.sciencedirect.com/science/article/pii/S0957417412012274
14. Segovia, F., G´
orriz, J.M., Ram´ırez, J., lvarez, I., Jim´enez-Hoyuela, J.M., Ortega,
S.J.: Improved parkinsonism diagnosis using a partial least squares based approach.
Medical physics 39(7), 4395–4403 (2012)
15. Stoeckel, J., Ayache, N., Malandain, G., Koulibaly, P.M., Ebmeier, K.P., Dar-
court, J.: Automatic classification of SPECT images of Alzheimer’s disease patients
and control subjects. In: Medical Image Computing and Computer-Assisted Inter-
vention - MICCAI. Lecture Notes in Computer Science, vol. 3217, pp. 654–662.
Springer (2004)
16. Towey, D.J., Bain, P.G., Nijran, K.S.: Automatic classification of 123I-FP-CIT
(DaTSCAN) SPECT images. Nuclear Medicine Communications 32(8), 699–707
(Aug 2011), http://www.ncbi.nlm.nih.gov/pubmed/21659911, PMID: 21659911
17. Vapnik, V.N.: Estimation of Dependences Based on Empirical Data. Springer-
Verlag, New York (1982)
18. Vapnik, V.N.: Statistical Learning Theory. John Wiley and Sons, Inc., New York
(1998)
Dostları ilə paylaş: |