Towards improved breast cancer diagnosis using microwave technology and machine learning

Today, the early detection of breast cancer for asymptomatic women primarily relies on generalised screening programmes with x-ray mammography. However, the long-term value of screening mammography has been questioned due to e.g. the high false positive rate, resulting in unnecessary biopsies and overdiagnosed cancers. In this context, a need exists for new breast screening modalities with greater specificity. In this thesis, a machine learning platform using microwave technology is investigated for the purpose of diagnosing breast cancer. The proposed platform is evaluated by means of numerical and experimental phantom sets designed and developed in this research. The proposed numerical tumour phantom set is designed to ensure tumour models are clinically-realistic; a validation procedure was undertaken with clinicians to verify the applicability of the models. Experimental tumour and breast models were also developed using tissue mimicking materials. Both the level of spiculation in the tumour models and varying levels of glandular content were included as novel elements of the phantom set. Breast cancer diagnosis was investigated through the development of a 3-stage automated platform to analyse backscattered signals, which includes: data acquisition; data pre-processing through tumour windowing and feature extraction; and diagnosis through a random forests classifier in a two-level architecture. The results demonstrate the usefulness of creating high-similarity groups of signals before the classification and how the extraction of features can capture the characteristics of a tumour, without the need for a priori information. Results also show that benign tumours are more often correctly classified than malignant tumours, suggesting that microwave breast systems may be capable of achieving high specificity rates. In the context of current approaches to breast cancer care management, the results of this work support the potential value of microwave breast diagnosis systems to impact patient outcomes