Robust Segmentation for Low Quality Cell Images from Blood and Bone Marrow

Biomedical image is often complex. An applied image analysis system should deal with the images which are of quite low quality and are challenging to segment. This paper presents a framework for color cell image segmentation by learning and classification online. It is a robust two-stage scheme using kernel method and watershed transform. In first stage, a two- class SVM is employed to discriminate the pixels of object from background; where the SVM is trained on the data which has been analyzed using the mean shift procedure. A real-time training strategy is also developed for SVM. In second stage, as the post-processing, local watershed transform is used to separate clustering cells. Comparison with the SSF (Scale space filter) and classical watershed-based algorithm (those are often employed for cell image segmentation) is given. Experimental results demonstrate that the new method is more accurate and robust than compared methods. The analysis of blood and bone marrow slides is a powerful diagnostic tool for the detection of leukemia. In order to determine the type of leukemia, the different lineages and maturity levels of white blood cells, which come from peripheral blood or bone marrow, need be recognized and counted. Whereas systems for the analysis of stained blood cells (either using flow cytometric methods or panoptically stained blood slides) that yield a pre-classification are commercially available, the analysis of bone marrow smears is much more difficult (1).