GPU Enabled Parallel Touching Cell Segmentation Using Mean Shift Based Seed Detection and Repulsive Level Set

Automated image analysis of histopathology specimens could potentially provide support for the early detection of breast cancer. Automated segmentation of cells in the digitized tissue microarray (TMA) is a prerequisite for quantitative analysis. However touching cells bring significant challenges for traditional segmentation algorithms. In this paper, we propose a novel algorithm to separate touching cells in hematoxylin stained breast TMA specimens which have been generated using a standard RGB camera. The algorithm starts with an accurate and fast object center localization approach using mean shift based seed detection. The final results are obtained using a multiphase level set with repulsive force. We compared our results with the most current literature. The segmentation results are evaluated by comparing the pixel wise accuracy between human experts’ annotation and the automatic segmentation algorithm. The method is tested using 100 image patches which containing more than 1000 touching cells. As the voting method of the seed detection is the most time consuming procedure, the algorithm is parallelized using graphic processing units (GPU) and 22 times speed up is achieved when compared with the C/C++ implementation.