Information Technology Reference
In-Depth Information
head motion, smoothed to reduce the effect of high-frequency noise on the analysis,
and spatially normalized to allow for intersubject comparisons within SPM5 soft-
ware. After that the activation map was obtained and was used as a gold standard
in training the classifiers in this study. Figure 5.1 shows the activation map. For fur-
ther analysis we use a global threshold in order to identify those voxels with highest
overall responsiveness. Figure 5.2 shows the activation map after thresholding.
Fig. 5.1: The activation map.
Fig. 5.2: Active regions are shown as white pixels.
5.3 Data Preprocessing
To estimate the generalization ability of the classification methods, we split each
data set into two nonoverlapping subsets: the training set on which each classifier
was trained and the test set on which each classifier was tested. The procedure was
repeated many times for different random partitions of the data, and results were
averaged across the results. In order to enhance classifier performance a variety of
data preprocessing steps were employed. Two of them were done for all voxels. First
the data were normalized by subtracting the mean value and dividing by the overall
standard deviation. Thus, each voxel had mean activity of 0 and unit standard devia-
tion. Second the outliers were removed by setting all values that were beyond three
