Automatic image analysis of histopathology specimens would help the early detection of blood cancer. The first step for automatic image analysis is segmentation. However, touching cells bring the difficulty for traditional segmentation algorithms. In this paper, we propose a novel algorithm which can reliably handle touching cells segmentation. Robust estimation and color active contour models are used to delineate the outer boundary. Concave points on the boundary and inner edges are automatically detected. A concave vertex graph is constructed from these points and edges. By minimizing a cost function based on morphological characteristics, we recursively calculate the optimal path in the graph to separate the touching cells. The algorithm is computationally efficient and has been tested on two large clinical dataset which contain 207 images and 3898 images respectively. Our algorithm provides better results than other studies reported in the recent literature.