A Hybrid Segmentation Framework using Level Set Method for Confocal Microscopy Images

Based on variational and level set approaches, we present a hybrid framework with quality control for confocal microscopy image segmentation. First, nuclei are modelled as blobs with additive noise and a filter derived from the Laplacian of a Gaussian kernel is applied for blob detection. Second, nuclei segmentation is reformulated as a front propagation problem and the energy minimization is obtained near the boundaries of the nuclei with the Fast-Marching algorithm. For each blob, multiple locally optimized points are selected as the initial condition of the front propagation to avoid image under-segmentation. In order to achieve higher accuracy, a graphical interface is provided for users to manually correct the errors. Finally, the estimated nuclei centres are used to mesh the image with a Voronoi network. Each mesh is considered as a Geodesic Active Contour and evolves to fit the boundaries of the nuclei. Additional post-processing tools are provided to eliminate potential residual errors. The method is tested on confocal microscopy images obtained during trophoblast elongation in ruminants. Experimental results show that cell nuclei can be segmented with controlled accuracy and difficulties such as inhomogeneous background or cell coalescence can be overcome.