Efficient texture descriptors for localisation of masses in local dense background of a mammogram

Author: Shelda Sajeev

Sajeev, Shelda, 2019 Efficient texture descriptors for localisation of masses in local dense background of a mammogram, Flinders University, College of Science and Engineering

Terms of Use: This electronic version is (or will be) made publicly available by Flinders University in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. You may use this material for uses permitted under the Copyright Act 1968. If you are the owner of any included third party copyright material and/or you believe that any material has been made available without permission of the copyright owner please contact copyright@flinders.edu.au with the details.


Breast cancer is considered to be one of the major health problems and leading causes of death among women worldwide. Screening mammography is the most used and cost-effective tool for detecting early stage cancer. However, detecting cancerous masses in the dense background of a breast is a particularly challenging task, even for an experienced radiologist. This stems from the similarity of intensity between the masses and the overlapped dense normal tissues. Mammographic sensitivity is less than 50% in women with dense breasts. The need for improved diagnosis of breast cancer in women with dense breast is further emphasized by the greater risk of breast cancer in this population. Women with dense breast have four to five times higher risk of getting breast cancer compared to women with little or no dense tissues. Computer-aided detection (CAD) has been developed to assist radiologists in early breast cancer detection and diagnosis. Although many CAD techniques have been developed for mass classification/detection, the CAD sensitivity in dense breast is still low.

This study aims to improve detection of cancerous masses localised in the dense background of breasts by characterising the textures of masses, based on primitive micropatterns (at pixel level) and their macro level (superpixel) representations. A new paradigm for texture analysis, based on superpixel tessellation, is the main contribution of this thesis. The paradigm enables new mechanisms for understanding complex texture structures in images. Both pixel and superpixel level micropatterns are used in this study to distinguish breast masses from normal dense tissues. The results indicate that the proposed textural features can produce highly effective and efficient descriptors of breast masses, localised in a dense background.

The effectiveness of the proposed approaches is validated on two datasets (DDSM and BSSA) using performance measures such as Dice Index, Hausdorff distance, receiver operating curve (ROC), area under the curve (AUC) and free receiver operating curve (FROC).

The experimental results indicate that the proposed methods can classify masses with AUC score up to 0.97 and can localise masses with sensitivity of 80% with only 2.7 false positives per image.

Keywords: Graph Modelling, Superpixel tessellation, Superpixel Local Binary Patterns, Dense Breast, Mammography, Adaptive CLAHE, CAD

Subject: Mathematical Sciences thesis

Thesis type: Doctor of Philosophy
Completed: 2019
School: College of Science and Engineering
Supervisor: Gobert Lee