This paper proposes a parametric, multivariate method for the joint detection and segmentation of brain activation based on fMRI data. The proposed technique uses region based level sets to separate between the task-related and non-task-related regions and performs, at each iteration of level set evolution, a separate multivariate linear model (MLM) analysis in each of the two regions. Simulations using synthetic data generated based on typical experimental parameters and noise levels showed a false positive rate of 6% and a false negative rate of 2% for the results obtained with the proposed technique. The performance of the level sets method was further investigated by analysing empirical fMRI data from two subjects performing a visual and a motor task. Our results indicate that the proposed technique provides brain activation results comparable to those obtained by a standard univariate approach, with the advantage that it does not require the definition of a significance threshold.