Segmentation of Tumor Regions in Microscopic Images of Breast Cancer Tissue

Abstract

Nowadays, advances in the domain of digital pathology gave the means to replace the old optical microscopes by the Whole Slide Imaging (WSI) scanners. These scanners enable pathologists viewing conventional tissue slides on a computer monitor. Currently, several applications that aim to analyze human tissue are evolving remarkably. Segmentation of tumor regions in microscopic images of breast cancer tissue in one of the application that researchers are investigating extensively. Indeed, researchers are interested in such application not only because breast cancer is one of the pervasive cancers for human beings, but also segmentation is one of the basic and frequent tasks that pathologists have to perform in order to perform tissue analysis.

In this thesis, we addressed the task of segmentation of tumor regions in microscopic images of breast cancer tissue as a machine learning task. We developed different supervised and unsupervised learning frameworks. Our proposed frameworks encompass five steps: (1) pre-processing, (2) feature extraction, (3) feature reduction, (4) supervised and unsupervised learning, and (5) post-processing. We focused on the extraction of textural features, as well as utilization of supervised learning techniques. We investigated individually the MR8Fast, Gabor, and Phase Gradient features, as well as a combination of all these features. We investigated also different classifiers which are Naive Bayes, Artificial Neural Network, and Support Vector Machine, as well as a combination of the supervised learning results.

We conducted different experiments in order to compare the different proposed frameworks. Therefore, we developed different conclusions. The MR8Fast features are the most discriminating features compared to the Gabor and Phase Gradient that come in the second and third place respectively. Furthermore, the Naive Bayes classifier and the combination of classification results, that have been overlooked for the segmentation of tumor regions in microscopic images of breast cancer tissue, achieved better results compared to the Support Vector Machine classifier which has been extensively employed for this task. These promising conclusions promote the need for further work to investigate other textural features as well as other classifiers.