Machine learning for breast cancer diagnosis in developing countries

Breast cancer is one of the leading causes of cancer deaths among women all over the world. Use of conventional imaging such as mammography, ultrasound and magnetic resonance imaging has shown reduction in mortality rates of breast cancer, especially in developed countries. However, implementing conventional imaging in developing countries poses challenges of high equipment cost and requirement of high-skilled staff. Embarrassment to undress in front of a technician for imaging is also found to be a barrier in developing countries like India. This thesis reconsiders infrared thermography as an imaging modality for breast cancer detection since it has advantages of being low-cost, portable, non-invasive, radiation-free and privacy aware imaging, where technicians need not see the undressed woman. However, manual interpretation of breast thermography had earlier produced low accuracy if not interpreted by a skilled thermographer and there are stringent warnings on its usage. This thesis proposes an end-to-end automated system with machine learning to alleviate the subjectivity in interpretation and to improve the accuracy of breast cancer detection. The proposed system is designed to improve clinician interaction with the help of interpretable features, annotated images and a cyclic feedback between the expert and different modules in the proposed system.