Functional magnetic resonance imaging (fMRI) is an imaging technique that can be used to characterize brain physiological activity, usually presented as 3D volumes in function of time. In the context of our previous work in nonlinear association studies in electroencephalogram (EEG) time series, we were able to identify clinical relevant features useful in clinical diagnosis. The use of a similar approach in fMRI, now adapted for 3D time series, is both appealing and new. Such time series analysis imposes challenging requirements regarding computational power and medical image management. In this paper we propose a grid architecture framework to support the typical analysis protocol of association analysis applied to fMRI. The system, implemented using the gLite middleware, provides the necessary support to manage brain images and run non-linear fMRI analysis methods.