Segmentation of bone marrow stromal cells in phase contrast microscopy images

Abstract

The morphology of bone marrow stromal cells (BMSCs) and the nature of phase contrast images make segmentation of such images challenging as many standard segmentation approaches do not work. This presents an obstacle to the development of systems that could use pattern recognition (PR) techniques to assess culture quality, since successful segmentation is an important precursor to successful pattern recognition. A method is presented for image normalisation and segmentation of cell regions within sub-confluent cell cultures of human bone marrow stromal cells, including a novel method of dealing with the halo associated with phase contrast images. The proposed method was evaluated by measuring its effect on the accuracy of a subsequent PR stage that was trained to discriminate between two BMSC cultures of differing quality. The accuracy achieved averaged 93% across four commonly used PR algorithms, corresponding to an overall accuracy gain of 17% compared to non-normalised, unsegmented images.