AUTOMATED EXTRACTION OF NUCLEI IN KIDNEY ELECTRON MICROGRAPHS

Abstract

This thesis concerns a study on the development of methods that can automatically extract cell nuclei in kidney tissue images. A survey of the current methodologies that have been employed in the segmentation of tissue images is first presented and various techniques in this specific area are reviewed. This survey suggested that classical segmentation techniques are still the main approaches in handling tissue sections. Advanced techniques such as fractal geometry and active contour models are shown to have potential in tissue image segmentation. Due to the complexity of the image, little work has been done on kidney image segmentation. A study on the methodologies in the extraction of cell nuclei in kidney sections is described in this thesis. A threshold that emphasizes high contrast edges over low contrast edges for intensity thresholding is first presented. Inspired by the fact that fractal geometry provides a means to characterize textured image surfaces, the relative box subtraction method is adopted to depict the tissue image surface. The fractal dimension map distribution is fully analyzed with regard to identify cell nuclei, and proves to be effective in characterizing tissue image features. Active contour modeling is a top-down strategy for the extraction of salient object boundaries. An adaptive greedy algorithm for active contours is introduced to improve the accuracy of extraction of external nuclear boundaries. By adding a negative force to initialize a new snake, we incorporate a data-driven algorithm into the extraction procedure. In comparison, morphological processing still plays an important role in effectively identifying the nuclear regions. The segmentation methods achieve good results when applied to electron micrograph images. The work in the thesis suggests that advanced and newly developed tools provide alternative approach and improved methods of solving the recognition problem in complicated tissue sections in combination with classical techniques.