Convolution model for automated localization of brain activity in fMRI images

In this article, we present an efficient technique for automated detection of neural activity from time-varying three-dimensional fMRI images. We develop a hemodynamic response model based on the impulse response function and the stimuli onset pattern that results in a linear combination of incomplete gamma functions. In order to improve the detection, we employ a spatial filtering approach based on a convolution model that improves the signal-to-noise ratio by suitably adapting to the local activity patterns. Then we develop a correlation model of the convolution signals and propose a method to detect the active regions based on the maximum correlation values. It eventually results in a nonlinear optimization problem with correlation maximization as the objective and bounds on impulse response parameters as the constraints. We propose an efficient method for the solution. We then implemented and tested it on real set of images. Experimental results show the effectiveness of our method for localization of brain activity in fMRI images