Molecular and metabolic pattern classification for diagnosis of recurrent brain gliomas using support vector machine learning model

spraytownspeakerAI and Robotics

Oct 16, 2013 (3 years and 9 months ago)

90 views

M
olecular
and metabolic
pattern classification for
diagnosis of
recurrent brain
gliomas

using
support vector machine learning m
odel


Authors:

Farzin Imani, MD, PhD, Fernando Boada, PhD, Frank S. Lieberman, MD, Erin

Deeb, BS,
and James M. Mountz, MD, PhD


O
bjectives:

The aim of this study was to
systematically combine
multimodal imaging
data

to improve diagnostic accuracy of
glioma
progression

in post
-
therapy patients using
Support Vector Machine (SVM) learning model.
The
SVM
generates a
n optimal

separatin
g
hyperplane in an n
-
dimensional space

between two classes of data

(e.g.,
pos
itive

vs. neg
ative
)
.

We demonstrate that
SVM outperform
s

single
-
parameter cutoffs

obtained from r
eceiver operating characteristic

(ROC) plot analysis

of

proton
Magnetic
Resonance Sp
ectroscopy (MRS
)

and FDG PET

data
.


Methods:

12

post
-
therapy patients (5
m
, 7

f
,
25

70 y
), initially with histolo
gy proven
gliomas (6 grade II and 6

grade III
)
who presented with

contrast enhancing lesions on
MRI

and

clinical symptoms suggestive but not co
nclusive of recurrence were
selected
.
FDG uptake
of lesions
was rated 0 (no uptake), 1 (<WM), 2 (=WM), 3 (>WM, <GM), 4
(=GM), 5 (>GM).

Choline (Cho) over Creatine (Cr) ratios of the lesions were
normalized
to the

contralateral hemisphere
.

C
linical follow
-
u
p

(> 12 mo)
and sequential MR
I

studies
were used as the reference standard.

An SVM with
linear classifier

and
dot

kernel
was
established

and computed
using quadratic programming to maximize the distances
between the hyperplane to the close
st

points of eith
er class (supporting vectors)
.

The
ROC
plots and optimal cutoff values f
or FDG uptake and Cho/Cr were calculated.


Results:

The SVM with

4 supporting vectors
was able
t
o classify

all

recurrent cases

(n=8
)

without false positive

results
(accuracy 100%)
.
The

equation of maximal margin
separating
hyperplane

was

0.5
F+
0.86
C
=
2.23
, where F and C

represent FDG uptake

and
normalized Cho/Cr ratio
, respectively. The accuracy of PET and MRS based on optimal
cutoff value
s
(FDG 2.5 and Cho/Cr 1.455)
using ROC analyses we
re 83
% and 75%

respectively
.


Conclusion:

SVM
technique
, in spite of

small number of cases, was able to effectively
combine MRS and FDG PET data and successfully classify recurrent gliomas more
accurately than single
-
parameter cutoffs.


Research Support:

N
ational Cancer Institute, Cancer Center Support Grant Supplement Award,
Imaging
Response Assessment Teams

Supporting Data





Figure 1.
Recurrent glioma
classification

by calculated

cutoff values
from
ROC analysi
s
of FDG PET

and normalized
Cho/Cr

MRS

dat
a
, as
well as SVM classification of
the
combined data.

The arrows point to the supporting vectors.





FDG PET

Cho/Cr MRS

SVM

Sensitivity

75%

75%

100%

Specificity

100%

75%

100%

NPV

67%

60%

100%

PPV

100%

86%

100%

Accuracy

83%

75%

100%


Table 1.
Resul
ts of

the data analysi
s
. (
*
F and C represent FDG uptake and normalized
Cho/Cr ratio, respectively)




No.

Age (y)

Sex

Pathology

WHO

1

25

F

Astrocytoma

2

2

29

F

Anaplastic Oligodendroglioma

3

3

33

M

Anaplastic Astrocytoma

3

4

60

F

Anaplastic Astrocytoma

3

5

30

M

Oligodendroglioma

2

6

54

F

Astrocytoma

2

7

54

M

Oligodendroglioma

2

8

30

F

Oligodendroglioma

2

9

70

F

Oligodendroglioma

2

10

29

M

Anaplastic Astrocytoma

3

11

36

F

Anaplastic Astrocytoma

3

12

48

M

Anaplastic Astrocytoma

3



Refrences:


Bo
ser
B.,
Guyon
I., and
Vapnik
V. 1992.
A training algorithm for optimal margin
classifiers
. In:
Annual Workshop on Computational Learning Theory archive

Proceedings of the fifth annual workshop on Computational learning theory
.
Assn for
Computing Machinery,

Pittsburgh
, PA. pp.

144
-
152
.



Vapnik V., Golowich S., and Smola A. 1997. Support vector method

for function
approximation, regression estimation, and signal

processing.
In: Advances in Neural
Information Processing Systems 9, ed. by M. Mozer and M. Jor
dan and T. Petsche,
MA,
MIT

Press, Cambridge
. pp. 281
-
287.


Byvatov E, Schneider G.

Support vector machine applications in bioinformatics.

Appl Bioinformatics. 2003;2(2):67
-
77.