Cell Segmentation In Microscopy Images For Biomedical Quantitative Analysis

Abstract

In the field of pathology, examination of organs and tissues at the cellular level is an important recurring task, especially for the research and detection of cancer. Automated solutions to the cell segmentation problem on microscopy images have been an important task for a long time, and many different solutions have been proposed. However, most, if not all, methods proposed have been met with hurdles, especially when faced against variance of cell morphologies, such as color, shape, size, or structure of the tissue and distribution of the cells. Deep learning based solutions have been proposed, which has shown to perform exceptionally well compared to traditional methods, but even these are far from perfect. In this thesis, we design a deep-learning-based method for the automated cell segmentation problem, where, given a microscopy image of H&E stained tissue specimen, our task is to divide the image into multiple regions, where each region contains a cell. We test the method on a multi-organ dataset, out of which there are organs that were not present during training. Our method has shown to achieve decent result, with the ability to generalize well, being able to achieve similar result even on images of organs that were not seen during training.