Epilepsy is the most common neurological condition in children and the most common cause is cortical malformations. We have been developing high quality quantitative structural Magnetic Resonance Imaging (MRI) to allow these to be more optimally detected. However detection of abnormalities remains challenging in a significant proportion of patients. This may be attributable to the current resolution of MRI (typically ~1x1x1mm) being unable to resolve cortical structure (4-6 layers within 2-4mm which needs scans at 0.5x0.5x0.5mm resolution).
In this project image resolution for quantitative structural MRI will be dramatically increased by:
1. Developing improved image acquisition protocols
2. Using new higher field strength 7T MRI
3. Applying machine learning methods from computer vision to combine data from 3T and 7T to enhance the image contrast at 7T / resolution of the 3T MRI.
The approaches developed will be piloted in epilepsy patients without visible abnormalities in current clinical 3T MRI images.
WP1 Development of high resolution imaging protocol.
We will first implement the existing variable flip angle multi parameter mapping (MPM) sequence used by us at 3T and our collaborators at 7T (N Weiskopf, MPI, Leipzig) which uses fast Low angle single shot (FLASH) images. This can obtain 0.5mm isotropic PD,T2*, and T1 maps in ~1 hour at 7T.
The current MPM implementation uses a Cartesian trajectory with each shot acquiring a single k-space line with multiple T2* weightings. However, for high resolution imaging a greater k-space area needs to be sampled, for which longer durations are required and dynamic correction of motion and B0 fluctuations will be critical. Coupled with increased gradient performance this means there are potential gains in efficiency by utilising non-Cartesian 3D trajectories. In particular, hybrid trajectories such as radial cones  dramatically reduces the number of ‘shots’ required to obtain 3D data and therefore allows for reduced scan times for high resolution 3D data acquisition and they are also inherently ‘self-navigatied’ using the central radial spoke to estimate and correct for motion and off resonance errors. In particular, we will build on the image acquisition, reconstruction and motion corrections methods developed for quantitative cardiac imaging by the project team  allowing for ultra-high resolution 7T imaging where the brain pulsates like a heart.
WP2 Resolution transfer from 7T to 3T
In this work package, we will utilise the rapid developments in the field of computer vision that have been increasingly applied in medical imaging. It has demonstrated that deep learning can achieve using two key image-to-image transformations: super-resolution imaging and domain transfer learning. These tasks map between two image domains, ‘X’ and ‘Y’, e.g. low-resolution and high-resolution images or where ‘X’ and ‘Y’ are images with different contrast. We will use a single image obtained at 7T with ultra-high resolution (e.g. 0.4mm isotropic) and a set of quantitative maps obtained at 3T (1x1x1mm isotropic) to generate quantitative maps at the higher resolution.
In addition, to the data obtained within this project, ~150 datasets already obtained at 1mm resolution are available for algorithm training purposes (e.g. contrast-to-contrast mapping) and existing 400µm 7T data (Collaboration with Leipzig) that can be used as additional training data for high-low resolution mapping. The nature of these algorithms  is that it is a local operation in image space, so there are many examples within a single (low resolution) image, dramatically reducing data requirements (every image contains many training examples).
WP3 Pilot clinical application
Clinical feasibility of the proposed approach will be tested on n=5 patients with focal epilepsy and PET visible abnormality but no MRI visible lesion using clinical standard scans. Where possible patients undergoing surgical evaluation with stereotactic intracranial EEG (SEEG) will be recruited because they may go on to have surgical resection allowing for histological validation of imaging findings.
1. Weiskopf N, Suckling J, Williams G, Correia MM, Inkster B, Tait R, Ooi C, Bullmore ET, Lutti A. Quantitative multi-parameter mapping of R1, PD(*), MT, and R2(*) at 3T: a multi-center validation. Front Neurosci. 2013 Jun 10;7:95.
2. Johnson, K. M. (2017). Hybrid radial-cones trajectory for accelerated MRI. Magnetic Resonance in Medicine, 77(3), 1068–1081.
3. Cruz G, Atkinson D, Henningsson M, Botnar RM, Prieto C. Magn Reson Med. 2017 May;77(5):1894-1908. doi: 10.1002/mrm.26274. Highly efficient nonrigid motion-corrected 3D whole-heart coronary vessel wall imaging.
4. Tanno, R., Worrall, D., Ghosh, A., Kaden, E., Sotiropoulos, S. N., Criminisi, A., & Alexander, D. C.. Bayesian Image Quality Transfer with CNNs: Exploring Uncertainty in dMRI Super-Resolution. https://arxiv.org/pdf/1705.00664.pdf