Robust muscle cell segmentation using region selection with dynamic programming

Morphological characteristics of muscle cells, such as cross-sectional areas (CSAs), are critical factors that determine the health and function of the muscle. Automated cell segmentation is usually a prerequisite to calculate the CSAs. However, it is challenging for many traditional segmentation methods to efficiently and effectively separate muscle cells. In this paper, we proposed a region selection-based algorithm for automatic touching cell segmentation on Hematoxylin and Eosin (H&E) stained muscle images. The algorithm starts with generating a set of segmentation candidates and then assigns these candidates proper scores based on a learnt cell shape model and local features. Next, a subset of region candidates is selected as final segmentation based on an Integer Linear Programming scheme under the constraint that no any pair of selected regions can overlap. The algorithm is extensively tested on 60 H&E stained muscle images with over 10,000 cells. Compared with the recent state-of-the-arts, the algorithm provides the best performance.