Functional magnetic resonance images (fMRI's) provide high-resolution datasets which allow researchers to obtain accurate delineation and sensitive detection of activation areas involved in cognitive processes. To preserve the resolution of this noninvasive technique, refined methods are required in the analysis of the data. In this paper, the authors first discuss the widely used methods based on a statistical parameter map (SPM) analysis exposing the different shortcomings of this approach when considering high-resolution data. First, the often used Gaussian filtering results in a blurring effect and in delocalization of the activated area. Secondly, the SPM approach only considers false alarms due to noise but not rejections of activated voxels. The authors propose to embed the fMRI analysis problem into a Bayesian framework consisting of two steps: (i) data restoration and (ii) data analysis. They, therefore, propose two Markov random fields (MRF's) to solve these two problems. Results on three protocols (visual, motor and word recognition) are shown for two SPM approaches and compared with the proposed MRF-approach.