Detection of nuclei by unsupervised manifold learning
Muhammad Arif, Nasir Rajpoot · Warwick Research Archive Portal (University of Warwick) · 2007
Shape related signatures of nuclei in a tissue section are important for diagnosis and prognosis of cancer. Understandably, the process of demarcation of nuclei for cytometry with high degree of confidence is the most difficult part as the tissue section is fraught with staining artifacts and frequently contains other objects such as overlapping nuclei, nuclear debris, and extracellular structures. In this paper, we address this problem using a novel clustering algorithm for various shapes in prostate histopathology images using an unsupervised manifold learning paradigm. Experimental results with two-dimensional embedding of the shapes using diffusion maps demonstrate that various shapes in the tissue section are organized in accordance to the degree of complexity of their boundaries. This important observation can be exploited in the development of computerized techniques for image based cytometry.