Detection and Classification of Mammographic Abnormalities

Authors Organisations
Type

Student thesis: Doctoral ThesisDoctor of Philosophy

Original languageEnglish
Awarding Institution
Supervisors/Advisors
Award date2019
Links
Show download statistics
View graph of relations

Abstract

Breast cancer is one of the major causes of death in women all around the
world and early detection and diagnosis improves life expectancy. Computer Aided Diagnosis (CAD) systems have been developed to assist radiologists to improve diagnosis and ultimately the treatment process. Many algorithms have been developed to detect and classify mammographic abnormalities. The primary goal of our research is to develop CAD tools for the detection and classification of specific abnormalities in mammograms. We have developed the following novel histogram-based approach for mass segmentation, four different approaches for the classification of benign and malignant micro-calcifications, one approach for the classification of benign and malignant mass, and two approaches for the classification of normal and abnormal mammograms. In addition we have proposed a novel method for Content Based Image Retrieval (CBIR) for mammogram patches. The datasets that we used in the experiments are the segmented micro-calcification images for the classification of benign and malignant micro-calcifications and mammogram patches for the classification and segmentation of mass from two different publicly available databases. For two variants of the topological modelling developed for the classification of benign and malignant micro-calcifications, we achieved best Classification Accuracy equal to 91% for the DDSM dataset and 85% for the MIAS database. Similar results have been achieved by using other approaches for the classification of micro-calcifications. We achieved a Classification Accuracy equal to 96% and 83% for benign and malignant mammographic masses using two different texton-based variants. For the classification of normal and abnormal mammograms several texture and intensity features have been explored and when using different feature sets the Classification Accuracy is 87% for normal and abnormal mammograms. The Classification Accuracy using the proposed CBIR is 86% for classifying normal and abnormal class of mammograms. All the developed approaches gave good and reliable results in addition the methods are generic and could be extended to other application areas.

Keywords

  • Breast cancer, Computer Aided Diagnosis (CAD), Content Based Image Retreival (CBIR)