Segmentation of Cells with Partial Occlusion and Part Configuration Constraint using Evolutionary Computation Masoud S. Nosrati and Ghassan Hamarneh {smn6, hamarneh}@sfu.ca Medical Image Analysis Lab., Simon Fraser University, BC, Canada Abstract. We propose a method for targeted segmentation that iden- tifies and delineates only those spatially-recurring objects that conform to specific geometrical, topological and appearance priors. By adopting a “tribes”-based, global genetic algorithm, we show how we incorpo- rate such priors into a faithful objective function unconcerned about its convexity. We evaluated our framework on a variety of histology and mi- croscopy images to segment potentially overlapping cells with complex topology. Our experiments confirmed the generality, reproducibility and improved accuracy of our approach compared to competing methods. 1 Introduction Histology and microscopy image analysis plays a crucial role in studying diseases such as cancer and in obtaining reference diagnosis (e.g. biopsy histopathology). Automatically segmenting cells in such images is one of the preliminary steps toward automatic image analysis and computer-aided diagnosis. In spite of re- cent advances in segmenting cells based on some homogeneity and smoothness characteristics, segmenting complex cells with a non-homogeneous appearance (with multiple internal regions) remains challenging. This problem becomes even more challenging when these complex cells overlap. Previous works addressed cell overlapping, for single-region cells, using post-processing [14, 15, 11] (e.g. finding connected components and using parameter sensitive morphological operations [11]). However, cells in histology and microscopy images typically consist of mul- tiple regions (e.g. membrane, nucleus, nucleolus), each with a unique appearance model (intensity, color or texture) and unique geometric characteristics (e.g. cell size and shape prior). Furthermore, well defined spatial interactions usually exist between different regions of a cell (e.g. membrane contains nucleus, and nucleus contains nucleolus). Most existing methods have only considered simple struc- tured cells and ignored their complex composition [1, 3, 2]. There are many types of priors that benefit the segmentation of spatially- recurring cells with appearance inhomogeneity along with cell-overlapping. Many state-of-the-art image segmentation methods are formulated as optimization problems, which are capable of incorporating multiple criteria (or priors) as energy terms in the objective function and examining the relative performance