Robust detection and classification of biomedical cell specimens from light microscope images

Abstract

Efforts to implement automated systems for quantitation and analysis of biomedical cells in 2-D light microscope images have been unsuccessful since they lack robustness under varying microscope settings. The thesis reviews these methods and proposes a novel method that significantly enhances the robustness of automated systems.The thesis makes the following contributions: Firstly, an edge detection scheme accurately detects boundary contours of cell specimens under varying image luminance, contrast and noise levels. Secondly, a concavity analysis method splits overlapping specimens of diverse shapes and sizes based on a set of rules. Lastly, texture features invariant to cell orientation, scale and contrast are extracted from each segmented cell and used to assign the cells to their respective classes.The various aspects of the method are validated through a series of experiments and also compared against existing methods. They have been successfully applied to natural and light microscope images of airborne spores and cytological specimens.