Conceptual data sampling for image segmentation- an application for breast cancer images

Abstract

At the present time data analytics have become a buzzword for the in- formation technology sector. In an attempt to analyze data; one may follow various paths. Be it deploying sophisticated technologies to process big data or using commodity hardware while applying data reduction/sampling techniques to draw meaningful insights from a data. In this thesis, we aim to reduce data size in terms of th e number of tuples/objects for a given data. Our method has driven its roots from formal concept analysis (FCA); which is a mathemat- ical framework for data analysis. The proposed transformation is preserving functional dependencies/implications in a database. Consequently, we can gen- erate a much smaller data sample that is able to help in making decisions. In this study, we analyze a variety of reduction methods in order to recognize the best one(s), including randomized object selection procedures. The accu- racy of the decision s made on generated sample is comparable to accuracy of the decision made of whole/original data. To illustrate the concept we have chosen data from medical image domain. The data used for experimentation contains microscopic images of breast cancer that need to be segmented into two categories; i.e. benign or malignant. Extensive set of experiments have been performed to show the strength of the proposed reduction method.