Principal Investigator: Professor Vicente Grau
Institution: University of OxfordTags: 40161, cardiac magnetic resonance imaging, convolutional neural networks, Machine Learning, motion analysis
The heart requires a complex combination of mechanical, chemical and electrical phenomena to produce and sustain an efficient beating action. Motion of cardiac walls during contraction and subsequent relaxation can be captured by modern imaging methods such as Cardiac Magnetic Resonance. It is known that such motion patterns are affected by disease, but current quantification methods leave substantial room for improvement.
Machine learning, and in particular the more recent methods known as deep learning, has revolutionised several areas of research such as computer vision. Its effective use requires the availability of large data sets, which had been lacking until the appearance of large data collection initiatives such as UK Biobank.
The proposed research aims at developing deep learning methods for the analysis of motion patterns in hearts from the UK Biobank, learning models of normal motion as well as the natural variability in the Biobank cohort, establishing links between these models and different subject characteristics, and eventually allowing the prediction of cardiac risks.
The proposed research plan will be carried out initially over a 36 month period, with the expectation that further research questions are developed over that time leading to subsequent scientific questions and potential new projects
The present project aims at developing deep learning methods for the characterisation of cardiac motion from Cardiac Magnetic Resonance (CMR) scans. The main research questions and aims are: a) ellucidating whether convolutional neural networks can be used to represent motion patterns from CMR incorporating their natural variability; b) analysing differences dependent on baseline variables as learned by the methods; c) analysing the relationship between CMR-derived characteristics and electrophysiological parameters extracted from the ECG, together with baseline variables ; and d) investigating the potential of such learned models to predict future cardiovascular events.