We present a method to segment femur bones in CAT images which allows a 3D reconstruction including the most difficult region, namely the proximal femur, in which the femur is in contact with the pelvis. Classical segmentation techniques fail in this region because there is no contrast between the different bones. The proposed method is based on a deformable model that allows to determine the shape of the femur even in those regions in which the contrast is low. The result of the segmentation at each slice provides a contour with ordered points which can be matched with the contour of adjacent slices to compute the 3D reconstruction.