“Active surfaces” or deformable models have been proposed for the segmentation of anatomic structures in MRI data. Such algorithms are dependent on a good initial approximation of the target shape. The purpose of this work was to develop a reliable method for automatic generation of a starting point for segmentation of the lateral ventricle. The algorithm uses a parametric representation of an average lateral ventricle, which is customized for each individual by modulating the parametric coefficients based on the brain parenchymal fraction. The method was developed with a training set of 6 healthy controls and 25 patients with multiple sclerosis, and tested on an additional set of 10 patients. Compared to the average ventricle, this new approach provided a closer approximation to the manually segmented ventricular shape in 81% of the cases in the training set and 100% of the additional test set.