Segmenting subcellular structures in histology tissue images

Pathologists often rely on features of the subcellular structures, i.e., nucleus and cytoplasm, to differentiate different types of immune cells. Accurate segmentation of the sub-cellular structures can help classify cell types. However, segmentation in histology tissue images is very challenging, due to, intra-class variations and inter-class similarity, complex and crowded background, and noisy data. In this paper, we propose a novel energy minimization framework with constraints, that can easily incorporate various photometric and geometric priors to alleviate such challenges. We also propose a novel layered graph model to solve our energy minimization problem optimally in polynomial time. Experiments on clinical data demonstrate our method obtain better results on both segmentation, and classification of cell types using features extracted from the segmented subcellular structures, than other popular state-of-the-art general segmentation methods.