Normalization of magnetic resonance images with a given reference is a common preprocessing task which is rarely discussed. We review and address this question for a specific neuro-imaging problem of practical huge interest. We investigate the influence of the location of region of interest used for normalization of perfusion maps obtained with perfusion magnetic resonance imaging in the framework of the study of acute stroke. We demonstrate that a slice by slice normalization based on the whole hemisphere strategy optimally reduces the variability of the predictive value of the different perfusion maps. Interestingly, this is obtained for all the tested perfusion maps both from numerical simulation of perfusion MRI and from perfusion maps of real patients through a Neyman-Pearson detection strategy. These are important results to ease the quantitative assessment of stroke lesion from perfusion MRI on cohorts of patients. The proposed methodology could easily be transposed to other medical imaging problems where normalization of images is necessary.