Morphology from texture in cytometry

A difficult challenge facing researchers in cellular biology is to automatically segment images of live human cells. This paper presents initial results from our method, which analyzes variation in the texture distribution in a semi-supervised fashion using thousands of unlabeled and a few hand-labeled images. We approach the segmentation problem by constructing a probability map over pixels while imposing constraints derived from biological knowledge. Performance of several algorithms is compared on the original benchmark, which consists of hundreds of cells with hand-traced cellular and nuclear membranes.