Development of Genetic Probabilistic Neural Network for the Analysis of fMRI Data

Abstract

In fMRI, various MR scan images of slices of the brain are recorded depending on the tasks the subject is performing. The data are normally hard to interpret even by an experienced physician due to artifacts, noise etc. In this talk, we describe a statistical tool, Genetic Probabilistic Neural Network (GPNN), for evolutionary optimization of the topology and parameters in the framework of density estimation. The basic idea of applying GPNN to fMRI data is to find feature waveforms and estimate the probability of a class of feature waveform conditional on a given waveform of voxel according to the tasks. In this algorithm, we use a complicated density function to approximate the distribution of the group of waveforms to be classified. Hence, for every task a very specific waveform should be extracted and thus it should be possible to locate the neural activation in the brain.