The cell-graphs of brain cancer

In the current practice of medicine, the most reliable way to diagnose cancer is the pathological examination of a biopsy. For that, pathologists examine a tissue biopsy under a microscope and make assessments largely based on visual interpretation of cell morphology and cell distribution. Therefore, these assessments lead to a certain level of subjectivity. Computational diagnostic tools that operate on quantitative measures have the potential to reduce the subjectivity. Different computational approaches have been proposed to capture the deviations in cell morphology or cell distribution in order to identify the existence of cancer. Despite their promising results, these approaches suffer from the difficulty of segmentation or the high sensitivity to noise or the analysis of cell distribution based on only immediate neighborhood relations. The goal of this thesis is to develop a diagnostic tool that quantifies cell distribution addressing these problems. For this purpose, a novel graph-theoretical tissue representation (cell-graphs) is introduced. By deriving a set of distinctive topological features from cell-graphs, this approach automatically distinguishes cancerous tissues from their counterparts with high accuracy. In this cell-graph representation, nodes represent cell-clusters and edges represent spatial interrelations of these cell-clusters. Thus, this representation quantifies the spatial distribution of cell-clusters across a tissue. This thesis presents the methodology of cell-graph generation along with a theoretical framework and experimental demonstrations. It also introduces the definitions of different sets of distinctive cell-graph features. The experiments on clinical data demonstrate that the cell-graph approach is able to distinguish a cancerous brain tissue from non-cancerous brain tissues (from a healthy tissue and a benign inflammatory process) with high accuracy. This thesis also extends experiments to structured breast tissues to demonstrate that the cell-graph approach is not limited to represent unstructured brain tissues. The results are not as accurate as in case of brain cancer diagnosis, leading to the necessity of customizing the cell-graph approach to also consider the structure of tissues. Nevertheless, the experiments demonstrate that this approach significantly improves the diagnostic power of existing approaches for both brain and breast cancer diagnosis.