diagnosis of schizophrenia using multivariate analysis of ... · diagnosis of schizophrenia using...
TRANSCRIPT
International Journal of Engineering Trends and Technology (IJETT) – Volume 37 Number 4 - July 2016
ISSN: 2231-5381 http://www.ijettjournal.org Page 221
Diagnosis of Schizophrenia using Multivariate
Analysis of Brain MR Images Sushrutha Bharadwaj. M
1, Anand Prem Rajan
2
1&2School of Biosciences and Technology, VIT University, Vellore - 632014, India
1Department of Medical Electronics, Dayananda Sagar College of Engineering, Bengaluru, India - 560078
Abstract- Current laboratory procedures for the diagnosis of
Schizophrenia are mere subjective methods i.e. based on
symptom profiles. Neural correlates do not provide sufficiently
useful criteria. The existing univariate image analysis
procedures are sub-optimal for the identification of specific
neural correlates of schizophrenia. This paper aims at
developing a mathematical tool for the diagnosis of
schizophrenia using multivariable analysis of brain MR
images. To identify the correlative pattern of gray matter
differences between schizophrenics and healthy controls,
Multivariate Linear Model (MLM) and Voxel based
Morphometry (VBM) in Statistical Parametric Mapping (SPM)
are used. Five images each of patients and controls are
acquired and subject to multivariate analysis. The perceivable
differences between the brain anatomies of the two groups are
extracted and the regions of interest are recorded using the
Talairach and Tournoux atlas. The pattern of the Eigen image
is characterized by positive loadings indicating gray matter
decline in the patients as well as the negative loadings
reflecting gray matter increase in the patients. A learning
algorithm using Support Vector Machines (SVM) is proposed
and a classifier model is to be built. The differences extracted
from the multivariate results have to be input to the model.
When a random image is then input to the model, it has to
accurately classify the incoming image as schizophrenic or a
healthy one. The accuracy is predicted to be around 80%.
These findings suggest that the multivariable analysis of brain
images could be an effective method for the diagnosis of
schizophrenia.
Keywords: Multivariate analysis, Schizophrenia, Statistical
Parametric Mapping, Support Vector Mapping, Voxel Based
Morphometry
I. INTRODUCTION
Current operational diagnostic systems for major
psychiatric disorders such as schizophrenia are based
solely on clinical manifestations and associated psycho-
social impairments [1-3]. There is no single objective
approach for the diagnosis of schizophrenia. No
laboratory test for schizophrenia currently exists.
Current image analysis approaches based on univariate
methods have been found to be sub-optimal to identify
potential specific neural correlates for schizophrenia.
Researchers are now looking towards multi-variable
analysis methods for an efficient objective analysis of
the disease. Current research focuses on developing a
clinical diagnosis procedure for schizophrenia. In
univariate analysis the time series of each voxel in the
brain is separately modeled and tested statistically for a
condition of interest, usually in the framework of the
General Linear Model [4-5].
In contrast, multivariate analysis considers the
measurement of signals from many locations
simultaneously. It is based on finding those patterns of
the brain across many voxels which are characteristic of
Schizophrenics versus controls to find perceivable
characteristics of schizophrenics using MR images.
Support vector machine (SVM) is a data classification
algorithm that has been increasingly employed as a
multivariate method for brain response classification [6].
SVM procedures can reveal coherent brain activations
for individual subjects. The research intends to build a
classifier using SVM that mathematically classifies a
random image as a schizophrenic or a control.
II. METHODS
The block diagram of figure 1 shows a series of
processes that are employed in statistical analysis that
are used in this research. MR images of the subjects
were acquired and pre-processing algorithms were
applied on the images. Statistical analysis of the images
was performed using Statistical Parametric Mapping
software. The Multivariate Methods toolbox was used to
process and extract multiple features from the images.
These features were used by the Support Vector
Machine tool to design a learning algorithm which
learns about the differences in the images of
schizophrenics and controls. Finally, a random image
input to the algorithm will be classified as a
schizophrenic or a normal image.
IMAGE ACQUISITION AND ANALYSIS
For experimental purpose, five images each of
schizophrenic patients and healthy controls were
acquired using MRI technique. The patients were
recruited from National Institute of Mental Health and
Neuro Sciences (NIMHANS), Bangalore, India. All the
subjects were of same age, sex, education and had
similar socio-economic status. Most importantly, all the
patients were right handed and anti-psychotic naïve.
The MRI images were acquired using a 3.0 Tesla
Philips (Achieva R2 release) MRI scanner. T1 weighted
International Journal of Engineering Trends and Technology (IJETT) – Volume 37 Number 4 - July 2016
ISSN: 2231-5381 http://www.ijettjournal.org Page 222
three-dimensional magnetization prepared rapid
acquisition gradient echo sequence was performed
yielding a total of 165 slices in the saggital plane. The
imaging parameters were: TR= 8.1 ms; TE= 3.7 ms;
Nutation angle= ; Field of view= 256mm; Slice
thickness= 1mm without any inter slice gap; NEX= 1;
Matrix size= 256*256; Voxel size= 1mm X 1mm X
1mm.
Figure 1: Basic flow of processes for multivariate
analysis
Image analysis was performed using the Statistical
Parametric Mapping (SPM) 5 software developed by
Wellcome Department of Cognitive Neurology, Institute
of Neurology, London, UK
(http://www.fil.ion.ucl.ac.uk/spm). This software was
implemented in MATLAB 7.01 (Mathworks Inc.,
Sherborn, MA, USA).
Image processing and analysis was performed
according to the standard VBM protocol [7]. The first
step is to realign the data to 'undo' the effects of subject
movement during the scanning session. The spatial
normalization involved transforming MRI images of all
the subjects to a template that approximated the
stereotactic space of Talairach and Tournoux [8]. The
spatially normalized images were resliced to a final
voxel size of 1×1×1 mm3 and partitioned into gray
matter, white matter, cerebrospinal fluid, and other
compartments. Modulated segments of gray matter were
finally smoothed with a FWHM Gaussian kernel.
Statistical parametric mapping entails the
construction of spatially extended statistical processes to
test hypotheses about regionally specific effects [9].
Statistical parametric maps (SPMs) are image processes
with voxel values that are, under the null hypothesis,
distributed according to a known probability density
function, usually the Student's T or F distributions. We
analyze each and every voxel using any standard
(univariate) statistical test. The resulting statistical
parameters are assembled into an image - the SPM.
SPMs are interpreted as spatially extended statistical
processes by referring to the probabilistic behavior of
Gaussian fields [10-13]. In the present research,
statistical t-test was conducted on the two groups of data
and the results were estimated. These results were then
used for multivariate analysis. The regional differences
in gray matter volume were preserved in these results.
MULTIVARIATE ANALYSIS
Next, we extracted the gray matter distribution that
differed most between schizophrenics and healthy
controls. We used the MLM (Multivariate Linear
Model) Package for this analysis (MMtoolbox, SHFJ-
CEA, Orsay, France,
http://www.madic.org/download/MMTBx/). The general
format of the multivariate protocol is shown in (1)
[Ds, u, Nvox] = MM (argfile, typeAnal) (1)
where Ds – Eigen values; u – Eigen vector; Nvox –
Total number of voxels considered; TypeAnal – Type of
analysis.
The standard way of using the MM package is to first
perform a standard spm analysis that will provide a first
apriori model, the corresponding estimated parameters
and residual sum of square images. The temporal filter
or the normalization chosen is of great importance for
meaningful interpretation of the MM results. The MLM
method is based on singular value decomposition of the
matrix Z, where Y is the data, X is the linear model, and
Σ represents the temporal covariance matrix of the data
[14].
(2)
The MM toolbox analyses the results of SPM
analysis and computes the dispersion matrix using the
beta and Eigen images as shown in figure 2. By default,
the package computes up to five Eigen images. Since
IMAGE ACQUISITION
PRE PROCESSING
STATISTICAL ANALYSIS
MULTIVARIATE PROCESSING
EXTRACTION OF FEATURES
LEARNING ALGORITHM
RESULT
International Journal of Engineering Trends and Technology (IJETT) – Volume 37 Number 4 - July 2016
ISSN: 2231-5381 http://www.ijettjournal.org Page 223
there is no temporal covariance for VBM data, the
matrix was simplified to . This matrix was
then decomposed into an Eigen image (Figure 3) that
represents the variance between the data and a set of
regressors.
Figure 2: Flowchart of MM analysis
MM embeds the concept of several spatial spaces such
that if the regression model is performed separately on
several subjects (1...N) or regions of interest having the
same temporal space, MM allows us to consider the data
as a matrix with dimension as in (3):
Common-time-dim X (subject1-space-dim + subject2-
space-dim + ... + subjectN-space-dim)
(3)
In such a case, the result of the analysis for the first
component is one time dimension Eigen vector and one
space dimension vector with size given in (4):
(Subject1-space-dim + subject2-space-dim + ... +
subjectN-space-dim) (4)
This can also be considered as one Eigen image per
subject (or region of interest). More often, the MM is
performed on a matrix with dimension given in (5):
(Subject1-time-dim + subject2-time-dim + ... +
subjectN-time-dim) X common-space-dim (5)
III. RESULTS AND DISCUSSION
The results obtained are particularly important for
characterizing the difference between a group of patients
and a group of healthy controls. The gray matter
increase in figure (3) is indicated by the positive
loadings which can also be noted in the Eigen values in
figure 4 indicating an increase in gray matter in the
corresponding regions of schizophrenics’ brains.
Alternatively, the negative loadings in Eigen images
indicate a decrease in gray matter in schizophrenics.
This observation is very useful in differentiating
between a schizophrenic brain and a normal brain.
Figure 3: Eigen images resulting from MM analysis.
Figure 4: Comparison b/w observed and predicted Eigen
values.
MULTIVARIATE METHODS
DESCRIPTION OF INPUT
ARGUMENTS
GET INPUT ARGUMENTS
SET MODEL SUB SPACE OF
INTEREST
COMPUTE TIME x TIME
DISPERSION MATRIX
COMPUTE DISPERSION MATRIX
WITH BETA IMAGES
WRITE EIGEN IMAGES
International Journal of Engineering Trends and Technology (IJETT) – Volume 37 Number 4 - July 2016
ISSN: 2231-5381 http://www.ijettjournal.org Page 224
IV. CONCLUSION
In the current paper, SPM and MM toolbox have
been used to differentiate schizophrenia from healthy
control subjects using multivariate analysis methods
applied on MR images. Perceivable differences between
the two groups of images are extracted. The positive and
negative loadings in the Eigen images obtained using
multivariate analysis algorithm relate to the gray matter
changes in the brain that are particular to the
Schizophrenia images.
The future work focuses on extracting the
perceivable differences from the results of multivariate
analysis and to build a classifier model using Support
Vector Machines (SVM). When a random image is input
to the model, it is intended to mathematically classify it
as either schizophrenic or a healthy control.
LIBSVM, a library of support vector machines has
been found to be a powerful tool for building a classifier
model. It can be found at
http://www.csie.ntu.edu.tw/~cjlin/libsvm. The accuracy
of classification is also expected to be high compared to
other classification models. Further elaboration of these
methods may contribute for effective clinical diagnosis
of schizophrenia.
ACKNOWLEDGEMENTS
The authors thank Dr. G. Venkat Subramanian, Dept
of Psychiatry, NIMHANS for providing suitable data
and communicating about SPM & VBM methods; Dr.
Ferath Kherif, UCL, London for guidance with the MM
toolbox; Dr. Fumiko Hoeft, Stanford University for her
communications about SVM.
REFERENCES
[1] World Health Organization, 1993. The ICD-10 Classification of
Mental and Behavioral Disorders: Diagnostic Criteria for Research
[2] American Psychiatric Association, 1994. Diagnostic and Statistical Manual of Mental Disorders, 4th ed. American Psychiatric Press,
Washington, DC.
[3] Yasuhiro Kawasaki, Michio Suzuki, Ferath Kherif, Tsutomu Takahashi, Shi-Yu Zhou, Kazue Nakamura, Mie Matsui, Tomiki
Sumiyoshi, Hikaru Seto, Masayoshi Kurachi., 2007. Multivariate
voxel-based morphometry successfully differentiates schizophrenia patients from healthy controls; Neuroimage 34 (2007) 235-242
[4] Friston, K. J., Holmes, A. P., Poline, J. B., Grasby, P. J., Williams,
S. C., Frackowiak, R. S., and Turner, R. (1995b). Analysis of fMRI time-series revisited. Neuroimage 2, 45–53.
[5] Worsley, K. J., Liao, C. H., Aston, J., Petre, V., Duncan, G. H.,
Morales, F., and Evans, A. C. (2002). A general statistical analysis for fMRI data. Neuroimage 15, 1–15.
[6] Ze Wang, Anna R. Childress, Jiongjiong Wang, and John A.
Detre., 2007. Support vector machine learning-based fMRI data group analysis, NeuroImage 36, 1139-1151.
[7] Good, C.D., Johnsrude, I.S., Ashburner, J., Henson, R.N., Friston,
K.J., Frackowiak, R.S., 2001a. A voxel-based morphometric study of ageing in 465 normal adult human brains. NeuroImage 14, 21–36.
[8] Talairach, J., Tournox, P., 1988. Co-Planar Stereotaxi Atlas of the
Human Brain: Three-dimensional Proportional System. Thieme, New York.
[9] Friston KJ Frith CD Liddle PF and Frackowiak RSJ. (1991)
Comparing functional (PET) images: the assessment of significant change. J. Cereb. Blood Flow Metab. 11:690-699
[10] Adler RJ. (1981) In “The geometry of random fields”. Wiley New
York [11] Worsley KJ Evans AC Marrett S and Neelin P. (1992) A three-
dimensional statistical analysis for rCBF activation studies in human
brain. J Cereb. Blood Flow Metab. 12:900-918 [12] Friston KJ Worsley KJ Frackowiak RSJ Mazziotta JC and Evans
AC. (1994a) Assessing the significance of focal activations using their
spatial extent. Hum. Brain Mapp. 1:214-220 [13] Worsley KJ Marrett S Neelin P Vandal AC Friston KJ and Evans
AC. (1996) A unified statistical approach or determining significant
signals in images of cerebral activation. Hum. Brain Mapp. 4:58-73 [14] Worsley, K.J., Poline, J.B., Friston, K.J., Evans, A.C., 1997.
Characterizing the response of PET and fMRI data using multivariate
linear models. NeuroImage 6, 305–319.