AUTOMATIC DECISION MAKING PROCESS FOR CLASSIFICATION OF FMRI PATTERNS INTEGRATING THE PRINCIPAL COMPONENT ANALYSIS AND THE LATERALITY INDEX IN PEDIATRIC EPILEPSY

Purpose

This study reports a new application of the Principal Component Analysis (PCA) as a data driven decision mechanism to automatically extract brain activation patterns from a given population that is asked to perform an Auditory Description Decision Test (ADDT) paradigm with no previous knowledge of the population.

Method

Functional Magnetic Resonance Imaging (fMRI) on 64 control and 38 epileptic subjects were processed. These datasets were acquired from 5 medical institutions.  Each subject's 3D fMRI activation map was structured into 1D vector, then a 2D matrix containing the whole population was created. The PCA was applied on this 2D matrix. The position of the subject on this matrix is arbitrary, but an index table was created for tracking purposes. The PCA coefficients were fed into a distance decision making algorithm to generate the primary clusters based on the first 2 significant eigenvectors with the largest eigenvalues.

Results

A 2D plane was used to depict the clusters on the eigenspace using only 2 eigenvectors as shown in Fig. 1. To validate the clustering technique for this specific ADDT paradigm, the mean activation for the members of the resulting clusters were calculated. Fig. 2 illustrates the mean activation pattern of each cluster with relevant activation slices for visual appreciation. Cluster 1 (67% of the population) presents high activation on the left hemisphere, which is considered typical behavior. Cluster 2 (8% of the total population and 24% of the epileptic population) exhibits a right hemisphere dominant response, which is considered atypical. Cluster 3 shows much stronger activation on the left hemisphere than cluster 1, especially in Broca's area; this is considered another typical behavior pattern. Table 1 shows the statistics of LI together with the PCA clusters. Fig. 3 plots the LI distribution based on the location of the activation.

 

Conclusion

This research shows that PCA is more effective than LI in terms of brain asymmetry activation description, since PCA reveals the actual spatial activation patterns, making evident the atypical language network behavior. LI lacks spatial and graphical information, and the use of different masks may generate totally different results.