Medical image segmentation often involves variants of deformable models to account for both the variability of object shapes and variation in image quality. Segmentation quality, however, highly depends on the initial estimate, and human guidance is often needed to guarantee acceptable results. For automating segmentation, our method employs a quality-of-fit function associated with a finite element model of shape in a search for the optimum parametrisation. A global search with an evolutionary strategy is employed to determine the set of optimum pose parameters for initialisation of the shape models. A local search subsequently optimises the nonrigid shape parameters by employing the deformable model paradigm. Experimental results are presented for different medical applications, which include object detection, localisation and segmentation, and show the good performance of our approach.