We present a level set based clustering technique to detect activation regions from functional brain images using contextual information. Earlier similar approaches have been primarily concerned with local spatial context. Our approach relies on the idea that voxels within a functional region have similar temporal behavior. Using a level set formulation, a two-dimensional curve is evolved with a speed proportional to a similarity measure between the fMRI signals of voxels lying on the curve and their neighbors in the direction of propagation. The correlation coefficient is used to quantify similarity in time series of adjacent voxels. Simulation results from synthetic images demonstrate that using spatio-temporal contextual information provides better segmentation than a context-free, voxel-wise technique. Results from a real fMRI experiment using auditory stimulation are also presented.