The application of spectroscopy techniques to the study of normal and pathological tissue is currently limited by the difficulties of acquiring precisely localized spectra and analyzing the resulting, relatively low signal-to-noise data. Interpretation of multivoxel spectroscopy data is greatly facilitated by estimating peak parameters and representing their spatial distribution as metabolite images. We present here the new algorithm MIMSTATS, which performs an analysis of these images and determines whether the variations in the images are statistically significant. The algorithm has been developed and tested by application to 31P CSI data from the human forearm and brain. Our results demonstrate that MIMSTATS can detect signal at a very low signal-to-noise ratio in a reliable and reproducible fashion and provides a sound basis for testing hypotheses concerning differences in distribution of metabolites.