A Level Set Based Hybride Framework for Confocal Image Segmentation

Based on level set approaches, a hybrid framework with quality control for nuclear segmentation of confocal images is presented. To overcome non homogeneous background, nuclei are firstly modeled into circles with some additive noise and Laplacian of Gaussian filter as a blob-detector is applied. Then, nuclei centers are obtained by energy minimization of fast marching towards the boundaries of desired objects. Here, multiple optimal points are selected as the initial condition to avoid undersegmentation. In order to achieve higher accuracy, the system is designed in a hybrid-structure so that selectable modules will permit manual adjustment to prevent errors propagation. The appropriate centers of nuclei divide the original image into Voronoi meshes. In each mesh, geodesic active contour evolves toward the minimum energy, and the influence of internal and external forces fit the accurate nuclear edge. The algorithm is successfully applied 3D nuclei segmentation from bovine trophoblast. Experiments show that noise in images can be effectively reduced and touching in clusters can be naturally managed.