Comparison of Supervised Classification Methods with Various Data Preprocessing Procedures for Activation Detection in fMRI Data - Computational Neuroscience - page 82

Table 5.1: Mean classification error of different classifiers with different preprocess-
ing procedures
Preprocess FLD Linear SVM Nonlinear SVM GNB Corr k NN

None 2.261 0.9 2.25 50.824 62.522 1.168

SVD 36.228 1.05 2.03 13.766 64.775 1.055

LDA 2.858 0.9 1.692 1.572 11.288 1.198

LDA+SVD 2.826 0.82 1.078 1.768 17.066 1.108

PCA 2.38 0.9 2.098 1.842 2.17 1.06

PCA+SVD 3.128 0.6 2.02 1.57 3.022 1.106

approaches. Since correlation performs so much worse than the chance level in the
first two cases of preprocessing procedures (without preprocessing and with SVD
only), we will not consider it in observations of these two cases. Denoising the data
by SVD in first two rows helped all algorithms besides linear SVM and FLD (actu-
ally increased FLD classification error by 34%). For instance, the accuracy of GNB
was enhanced by 37% and the accuracy of k NN and nonlinear SVM were a little
enhanced. Dimension reduction by PCA or LDA enhances performance on the ma-
jority of classifiers reported. It is clear that the classification error of all classifiers
after reducing the data dimension via PCA is almost the same. So we can use any of
the classifiers after applying the PCA without any concern about the results.

5.6 Conclusions

In this chapter we have compared different classification methods with various
preprocessing procedures for detecting activation in fMRI data. The experimental
results presented here demonstrate the feasibility of training classifiers to distin-
guish between active and inactive voxels of fMRI data. kNN and SVM perform
equivalently and better than other approaches. They can classify the voxel with the
classification error of less than 1.2%. Further work could include the use of other
dimension reduction methods such as independent component analysis (ICA). It is
also of interest to examine other classification algorithms.

References

1. Franc, V., Hlavac, V. Statistical pattern recognition toolbox for matlab. Center for Machine
Perception, Czech Technical University, Prague, Czech (2004)
2. Group, T.F.M. Spm5 manual. (2005)
3. Isabelle, G., Andre, E. An introduction to variable and feature selection. J Mach Learn Res 3 ,
1157-1182 (2003)
4. Ku, S.-P., Gretton, A., Macke, J., Logothetis, N.K. Comparison of pattern recognition methods
in classifying high-resolution bold signals obtained at high magnetic field in monkeys. Magn
Reson Imaging 26 , 1007-1014 (2008)