A probabilistic framework for simultaneous segmentation and classification of multiple cells in multi-marker microscopy images

Segmentation and classification of cells in biological data are important problems in bio-medical image analysis. This paper outlines a novel probabilistic approach to simultaneously classify and segment multiple cells of different classes in a multi-variate setting. Superpixels are extracted from the input vector-valued image, and a 2D hidden Markov model (HMM) is set up on the superpixel graph. HMM emission probabilities are used to ensure high confidence in local class selection based on superpixel feature vectors. Spatial consistency of labels is enforced by proper choice of transition probabilities, which are conditioned on the feature vectors of neighboring superpixels at each location. Optimal superpixel-level class labels are inferred using the HMM, and are aggregated to obtain global multiple object segmentation. The performance is demonstrated on a challenging microscopy dataset. Experiments show, both quantitatively and qualitatively, that the proposed approach effectively segments and classifies cells, outperforming related techniques.