Information Technology Reference
In-Depth Information
vast majority of published researches summarizes average fMRI responses when the
subject responds to repeated stimuli of some type (e.g., reading, mental imagery, re-
membering) [5]. Other researchers have since applied various multivariate methods
in analyzing distributed response patterns in the human fMRI data sets: approaches
include training machine learning classifiers to automatically decode the subjects'
cognitive state at a single time instant or interval [6]. These statistical pattern recog-
nition algorithms are powerful because they project the activity of multiple voxels
to achieve a discriminative separation of the activity patterns. Before performing
pattern recognition algorithms, there is a need to select a subset of voxels for further
analysis. The procedure of selecting particular voxels can greatly enhance classi-
fication performance [3]. The enhancement consists of avoiding the “curse of di-
mensionality” by reducing the dimension of the space of patterns to be labeled and
removing noise features that can only degrade performance [4]. Likewise, most clas-
sification of fMRI data depends on an effective feature selection procedure being
applied beforehand [5]. In a typical fMRI study, time courses of more than several
thousand voxels are simultaneously acquired. Many of these are uninformative and
could severely damage the performance of the algorithm. In order to perform pattern
recognition more efficiently, one should use a technique to find a reasonable subset
of voxels to feed the classifiers. The aim of this work is to exploit supervised clas-
sification techniques for the voxel selection procedure. The goal of these analyses
is to detect the activated voxels (those voxels with highest overall responsiveness).
In the present study we applied several pattern recognition techniques and data pre-
processing approaches to compare their performance in classifying active and in-
active voxels. We used five classification procedures: the Fisher linear discriminant
(FLD), support vector machine (SVM), Gaussian nave Bayes (GNB), correlation
analysis, and
k
-nearest neighbor classifier (
k
NN). This chapter is organized as fol-
lows. In the next section, we briefly explain the acquisition of fMRI used in the
application. In Section 5.3, we provide data preprocessing approaches. Section 5.4
describes the pattern recognition techniques. Results of experiments are presented in
Section 5.5.
5.2 Data Set
In the studies described in this chapter, a data set from the SPM site http://www.fil.
ion.ucl.ac.uk/spm/data/ was used which comprises whole brain BOLD/EPI images
acquired on a modified 2T Siemens MAGNETOM vision system. This data set was
the first ever collected and analyzed in the functional imaging laboratory (FIL). Each
acquisition consisted of 64 contiguous slices (64
×
×
×
×
3 mm voxels).
Acquisition took 6.05 s, with the repetition time (TR) set arbitrarily to 7 s. At whole
96 acquisitions were made from a single subject giving 16 42 s blocks. The condi-
tion for successive blocks alternated between rest and auditory stimulation, starting
with rest. Auditory stimulation was bi-syllabic words presented binaurally at a rate
of 60 per minute [2]. The images were then realigned to mitigate noise caused by
64
64 3
3
