Information Technology Reference
In-Depth Information
different features are used to describe the pattern of BOLD effect, such as peak
time, peak value, change rate, cumulative changes, and bold change value at
each time point. Among these features, cumulative changes might describe the
change rate, peak value, and peak time synthetically.
In this paper, we compare two feature extracting methods for encoding fMRI-
sequence (t, t+n) as inputs to the classifier. The first one is to extract cumulative
BOLD changes of each voxel located in predefined regions, in which the input
vector is a 748-dimensional vector. Supposing V =[
v 748 ), where
v i is the sum of BOLD changes from scan2 to scan7 of the ith voxel. The second
one is to extract the change value at each time point of BOLD time course, in
which the classifier's input vector is a 6-dimensional vector. Supposing T =[
v i ]=(
v 1 ,
v 2 ,...,
t j ]
=(
t 1 ,
t 2 , ...,
t 6 ), where
t j means the BOLD signal change value at the j+1 scan
after stimulus onset (
t 1 corresponding to scan2,
t 2 corresponding to scan 3, and
so on).
Classification
Two classifiers based on SVM are compared in the present study, one is called
multi-voxel-classifier (MVC), in which the 748-dimensional vector of MVC is
used as the future vector of the classifier. The other is called single-voxel based
voter classifier (SVVC), in which the 6-dimensional vector for each selected voxel
.. .
voxel 2
voxel n
voxel 1
v n
v 1
v 2
.... .
(a) Input vector
.... .
.... .
Training samples
v 1 v 2 v 3 v 4 v n
Task type
v i
sample 1
Training
.... .
sample 2
sample 3
. ... .
sample
n
.... .
.... .
Test Data
Task type
?
?
?
v j
v 1 v 2 v 3 v 4 v n
Predicting
Data 1
Data
2
Data n
(b) Training classifier and predicting
Fig. 4. TheschemeofMVC
Search WWH ::




Custom Search