Dynamic analysis of individual cells in sheet-forming tissues is important to understand wound healing and cell-cell-contact regulation. Given is a temporal sequence of images of the endothelial cell culture, which is interpreted as a voxel image X with two spatial and one temporal dimension. The voxel grid is assumed to be a three-dimensional lattice, i.e. an undirected Graph, where each node corresponds to one picture element. The goal is to find a segmentation S, such that each node is assigned one of two possible labels, one for "background" and another for "cell nucleus". The correlation of an image sequence X and its segmentation S, i.e. the joined probability P(X, S), is modelled statistically as a Gibbs random field. It comprises an unspecific motion model, i.e. it assumes a smooth slow motion of segments without modelling a specific direction of that motion. An optimal segmentation is found as Bayesian decision with a local additive cost function. Experiments show cell nuclei are properly segmented. The main advantage of this approach is that in principle all parameters of the model can be learnt unsupervised, which is shown practically for certain parameters. Another advantage is that the temporal context of an image in the sequence is taken into account and influences the segmentation results. The segmented images can be used for dynamic analysis of shape, orientation and trajectories of cells and cell groups.