The segmentation of abdomen organs in volumetric medical images is difficult due to noisy and low contrasted images. Classical segmentation technics based on edge detection or thresholding lead to poor results. In this report, we use deformable meshes to perform segmentation. By using a template of the desired organ in our segmentation scheme, we introduce a prior knowledge of the shape to recover. We use noisy edge information to locally deform our model. Due to sparse edge data, we need to constrain our model so that it deforms smoothly. Our simplex meshes rely on a shape memory mechanism to regularize deformations. We are also using global transformations to provide additionnal constraints. An hybrid model provides a trade off between computationnal cost of complex global transformations and the number of freedom degrees of the model. We also study the use of a training set to built a more robust model with statistical knowledge of allowable deformations. Statistical information may be use for additionnal constraints or a fine tuning of the deformations parameters.
