Towards 10-minute Magnetic Resonance Imaging scans in children with machine learning

Now Closed

1.    Primary Supervisor: Dr. Jennifer Steeden (UCL Centre for Cardiovascular Imaging, Institute of Cardiovascular Science)
2.    Secondary Supervisors: Dr. Owen Arthurs  (Great Ormond Street Hospital) and Prof. Vivek Muthurangu (UCL Centre for Cardiovascular Imaging, Institute of Cardiovascular Science)

Project summary
A 4-year PhD studentship is available in the UCL Centre for Cardiovascular Imaging, working closely with Great Ormond Street Hospital (GOSH). The funding covers an annual tax free stipend and tuition fees. As this is a BRC/EPSRC funded studentship the standard EPSRC eligibility criteria apply, please see the EPSRC website for details. The successful candidate will be enrolled onto the UCL CDT in Intelligent Integrated Imaging in Healthcare (i4health) and benefit from the being part of a cohort of PhD students as well as participation in the activities and events organised by the centre.

Background
Magnetic Resonance Imaging (MRI) has enabled significant advances in the diagnosis and management of many diseases. However, MRI is challenging in the pediatric population as it is time consuming (~1 hour to perform) and requires patient cooperation. Hence it is often necessary to use general anesthesia (GA) in children below 8 years of age, which is both costly and carries some risk. 
One way of overcoming these problems would be to speed up the MRI scans so children do not have to keep still or hold their breath. The simplest way of doing this is to acquire less data (data undersampling), however this results in artefacts that make the images unusable. The resulting artefacts are dependent on the data undersampling scheme used. Current reconstruction methods for removing these artefacts, allow limited acceleration, or use time consuming algorithms which hamper their clinical uptake. A new approach is Machine Learning that aims to 'learn' how to remove undersampling and motion artefacts. In particular, convolutional neural networks (CNN) have been shown to be well suited to MRI reconstruction. 
The purpose of this project is to develop fast MRI acquisitions, with rapid Machine Learning reconstruction technologies for use in childhood diseases of the abdomen.
Building on work in motion correction, super resolution and deep artefact suppression, this project aims to reframe the reconstruction of MRI data, to remove aliases caused by data undersampling and motion corruption, as an image de-noising problem that can be initially performed by a CNN. This strategy requires specific sampling patterns that produce noise-like aliasing and high quality, application-specific training data. Because some pediatric diseases are rare, we may not have access to large amounts of training data, and part of this proposal is to look into the use of transfer learning techniques to enable development of accurate networks from small prospective or retrospective data sets. 
By the end of this project, the student will have an excellent understanding of machine learning algorithms particularly for reconstruction of MRI data. The student will also be able to design and implement their own MRI sequences, and have an understanding of traditional and state-of-the-art reconstruction algorithms. All work packages will be integrated into standard clinical workflow to enable clinical validation studies, and simple translation into routine clinical practise.

Research Aims
This study aims to develop novel accelerated magnetic resonance imaging (MRI) technologies which will allow scan times to be reduced from ~1 hour to ~10 minutes in children with diseases within the abdomen. This will be achieved this through development of optimised MR acquisition strategies combined with Machine Learning (ML) reconstruction techniques. 
Specific work packages include development of rapid, undersampled, non-Cartesian MRI sequences for abdominal imaging, development of machine learning techniques to correct for artefacts caused by respiratory motion, as well as development of transfer learning techniques to enable fast and accurate reconstruction of abdominal MRI’s where there is only a limited training data sets. The resulting networks will be integrated into standard clinical workflow to enable clinical validation studies, as well as simple translation into routine clinical practise.

To Apply: Please send a CV and Covering Letter expressing your interest to Dr Jennifer Steeden jennifer.steeden@ucl.ac.uk 

Deadline: 18th June 2020