In multiple sclerosis (MS) research, burden of disease and treatments efficacy are mainly evaluated with lesion load and atrophy. The former being poorly correlated with patient's handicap, it is of interest to evaluate accurately the latter. A lot of methods to measure the brain atrophy are available in the literature. The brain parenchymal fraction (BPF) is one of these methods. It needs a precise segmentation of the brain and of the cerebro-spinal fluid. However, artefacts like partial volume effects (PVE) can impair this classification. According to some articles, the BPF may also be less precise in longitudinal studies. To address these points, this article proposes a new method to evaluate the BPF which is based on an expectation-minimization framework taking into consideration the PVE. Modifications of the workflow are also proposed to improve its reliability in longitudinal study. Experiments have been conducted on simulated pathological images that validate the different measures.