The algorithm uses image derivatives to segment nuclei in a densely packed 3D tissue. The input file is set of z-slices imaged over time stored in tiff or lsm format ( from Carl Zeiss microscope). After first level of segmentation based on image derivatives, 3D properties are computed. Using these properties image statistics is inferred for the segmented objects and the algorithm implements clustering methods (mind you, computationally time taking!!) to resolve fused nuclei into single ones.
For more details, please read "Object Segmentation and Ground Truth in 3D Embryonic Imaging" published research article (PLOS ONE | DOI:10.1371/journal.pone.0150853). Please refer to this work for citations.