Machine Learning For Rapid Cardiovascular MRI

MRI has become the reference standard method of assessment for many aspects of cardiac function and anatomy. However, MRI is an inherently slow methodology, resulting in long scan times (~1 hour) and sensitivity to motion. In compliant adults we can get good quality images through the use of breath-holding and ECG gating. However, there is great interest in speeding up scans to enable MRI to be used in children and subjects who are unable to perform breath-holding. Additionally, reduced scan times would make MRI more comfortable for patients, reduce costs of scanning, improve patient throughput and reduce NHS waiting lists. 

In this talk I will cover some of the ways in which we have been able to sped up cardiovascular MRI at UCL. These tools have enabled real-time imaging without the use of ECG gating or breath-holding, with total scan times of just 15 minutes. Our techniques rely on non-cartesian acquisitions with significant data undersampling. Historically we have used compressed sensing reconstructions, but more recently are heavily reliant on machine learning reconstructions to overcome the bottleneck caused by traditional slow, iterative reconstructions. We have translated these techniques into clinical and research services at UCL and I will demonstrate the use of these rapid CMR techniques in children, during active exercise and during cardiac intervention (catheter tracking). Additionally I will show how we use machine learning to enable rapid post-processing of cardiovascular MRI images (including pixel based segmentation, graph based segmentation and computational fluid dynamics).