Automated prediction and early detection of breast cancer in mammograms

  • Georgia Valeria Ionescu

Student thesis: Phd

Abstract

Breast cancer is one of the most frequent forms of cancer that affects 1 in 9 women in the Western world. Screening programmes help towards the diagnosis of the disease when it is still asymptomatic and cancers can be detected at an early stage. Computer-based aids have been explored as a way of helping to predict and detect breast cancer due to shortages of medical experts. It is also recognised that radiologists' performance is variable, and computers may be able to improve consistency of early cancer detection. There is also scope to automatically predict women at risk of developing breast cancer so that preventive strategies can be put in place, and to stratify women according to risk rather than using a one-size-fits-all breast screening model. In this thesis we investigate two aspects of computer aided mammographic screening: (a) the impact of prompting errors on observer performance and (b) methods for assessing breast density, an important risk factor for breast cancer. The first direction of work focuses on offering a deeper insight into the way observers interact with prompting systems and how their performance varies depending on the true and false prompts produced by a computer aided detection (CAD) system. We employed a citizen science approach and simulated a CAD environment through an on-line game. We showed that observers are distracted by false prompts and that observer performance can be adversely affected by prompting parameters. The second part of the thesis proposes fully automated methods for breast cancer risk assessment and breast density assessment using deep learning. Our breast density method achieves performance equivalent to that of experts and our predictions are highly predictive of future development of breast cancer.  Our approach outperforms existing commercial solutions used in clinical practice and would make a pragmatic method for population-based screening stratification.
Date of Award1 Aug 2020
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorSusan Astley (Supervisor) & Johan Hulleman (Supervisor)

Keywords

  • deep learning
  • mammography
  • breast density
  • breast cancer risk

Cite this

'