MRI (Magnetic Resonance Imaging) is an important medical imaging method for disease diagnosis. Worldwidely more than 60 million investigations with MRI are performed each year. However, one challenging issue with MRI imaging is the huge amount of data that needs to be processed in order to get high resolution images. The inherently parallel processing of the large amount of data makes it a good application to run on a cluster. In this paper, we study the MRI applications by producing a parallelized version to run on our eight node prototype cluster. We use Stochastic activity networks (SANs) to model the MRI reconstruction algorithm running on our prototype clusters. We evaluate the dependability and scalability of the prototype MRI system. We also investigate how to improve the performance of the MRI system.