Using Artificial Intelligence to quantify cancer risk from breast Images

Breast cancer is the most common cancer amongst women worldwide. In this project we will develop methods to predict breast cancer risk by using artificial intelligence methods to investigate how images change over time. Accurate personalised breast cancer risk assessment will facilitate the introduction of preventative risk-reducing interventions, and also enable targeting of supplementary imaging modalities to those at higher risk, with the aim of earlier detection and improved survival rates.

This project will focus on analysis of large datasets of breast Magnetic Resonance Imaging (MRI) and x-ray mammography images from the UK, Canada and Norway. We will use data with multiple images taken over time, along with an available associated risk factor data. This will enable us to build models that can predict whether changes in appearance are those expected as women age, indicate changes in cancer risk or are responses to preventive interventions.

Deep learning has become the dominant method in artificial intelligence due to its high level of accuracy, particularly when classifying images. We will develop predictive methods using segmentation-type deep models (e.g. Ronneberger 2015) to produce risk maps of the breast, and then use regression or classification approaches to produce overall risk scores for each woman. As we will have multiple images from the same woman we will use time-dependent approaches such as recurrent neural networks (Hochreiter 1997) and dilated convolutional networks (Borovykh 2017). We expect to work in lower dimensional spaces shared between the different modalities, using approaches such as auto-encoders (Kingma 2013) so we can jointly use these data for risk prediction. We will also investigate whether registration methods (e.g. Garcia 2019) can effectively combine different modalities.

The project will involve curation of the large datasets needed, development of deep learning methodologies, and systematic evaluation on data previously unseen during the development phase.