Neural network approach towards pattern classification using fMRI activation maps

The activity patterns in fMRI data represent execution of different physical and mental tasks. Each of these patterns is unique and located in specific location in the brain. The main aim of analyzing these datasets is to localize the areas of the brain that have been activated in a given experiment. The basic analysis involves carrying out a statistical test for activation at thousands of locations in the brain. In this paper an attempt is made to develop and train classifiers based on the subjectspsila fMRI sequences in order to predict the tasks performed. The fMRI data set is huge and also the data size for different tasks is dimensionally dissimilar. Dimensionality reduction of high dimensional data is useful for three general reasons; it reduces computational requirements for subsequent operations on the data, eliminates redundancies in the data, and, in cases where the feature data set dimensionality doesnpsilat match then a common dimension is to be arrived at with the available data. All three reasons apply here, and motivate the use of Principal Component Analysis (PCA), a standard method for creating uncorrelated variables from best-fitting linear combinations of the variables in the raw data. The depth information data is extracted using Statistical Parametric mapping (SPM). The templates comprising of principal components represent individual activity. These are then fed to the back propagation training algorithm. The trained network is capable of classifying the test pattern into the corresponding defined class.