Surface extraction from noisy volumetric images is a problem commonly encountered in medical image analysis. Deformable surface models can, in principle, solve the problem in an automatic manner. However, it is often essential that a reasonably close initialization and good parameter values for deformable models are provided. In this paper, novel algorithms for global minimization of the energy of deformable meshes are presented. We demonstrate that global optimization by these algorithms reduces the sensitivity of the deformable mesh to its initialization and its parameter values. Consequently, it becomes easier to automate the initialization process and the selection of parameter values. As the second contribution, the internal energy function is derived in a novel way in the framework of deformable surface models. The construction of the internal energy in this way features a simple way to derive the variants of our global optimization algorithm. The experiments with synthetic images are performed to compare variants of the proposed optimization algorithm. Also, we present a practical application of our deformable model to automatic segmentation of positron emission tomography images.
Combining discriminative and generative methods for 3D deformable surface and articulated pose reconstruction
