Unsupervised brightfield image segmentation with RPCA and spectral clustering

Microscopic image analysis is an important step in biological study. Biologists study the movements of cells under certain drug treatments through a sequence of time-lapse microscopic images to determine the effect of the treatments. The development of modern bright-field microscopes allows more detailed investigations of the cell activities. However, it also brings challenges for automatic cell image analysis because of the low-contrast nature of bright-field microscopic images. This paper presents contributions to automatic bright-field cell image segmentation. We propose ten methods for bright-field cell image segmentation. The ten methods are based on two well-known methods in computer vision, namely spectral clustering and robust principal component analysis (RPCA). The first three methods are based on spectral clustering. They determine the segmentations by classifying the k segments from spectral clustering into cell segments and background segments. The other three methods are RPCA-based methods. The cell image segmentation problem is treated as a background subtraction problem, where the cells are the sparse moving objects to be identified. Several modifications have been made to RPCA to improve the segmentations. In the last four methods, we combine spectral clustering and RPCA to solve the segmentation problem. The first two methods use the results from RPCA to help the segmentation based on spectral clustering. In the last two methods, we formulate the problem as a principal component pursuit with graph cut penalization, and obtained the segmentation results similar to the three RPCAbased methods presented previously. The last method outperforms all previous methods in terms of segmentation quality. We have applied these methods on a set of C2C12 cells in bright-field microscopy. Experimental results confirm that the proposed methods give accurate segmentation of cells in bright-field microscopy, which conventional image segmentation methods cannot attain.