Shape-constrained level set segmentation for hybrid CPU–GPU computers

Due to its intrinsic advantages such as the ability to handle complex shapes, the level set method (LSM) has been widely applied to image segmentation. Nevertheless, the LSM is computationally expensive. In order to improve the performance of the traditional LSM both in terms of efficiency and effectiveness, we propose a novel algorithm based on the lattice Boltzmann method (LBM). Using local region statistics and prior shape, we design an effective and local speed function for the LSM, from which we deduce a shape prior based body force for LBM solver. An NVIDIA graphics processing units (GPU) is used to accelerate the method. Our introduced algorithm has several advantages. First, it is accurate even if there are some geometric transformations (rotation angle, scaling factor and translation vector) between the object to be segmented and the prior shape. Second, it is local and therefore suitable for massively parallel architectures. Third, the use of local region information allows it to deal with intensity inhomogeneities. Fourth, including shape prior allows the method to handle occlusion and noise. Fourth, the model is fast. Finally the algorithm can be used without shape prior by means of minor modification. Intensive experiments demonstrate, objectively and subjectively, the performance of the introduced framework.