Towards 10-minute Magnetic Resonance Scanning in Children - Developing Accelerated Imaging Using Machine Learning

Conventional Magnetic Resonance Imaging (MRI) is very time consuming (taking over 1hour/scan) and requires patients to remain still and perform multiple breath-holds. This is particularly difficult for children, and my work focusses on developing fast imaging techniques using Machine Learning (ML), to speed up MRI in children and reduce the need for breath-holding.

I have shown that it is possible to use ML to 'learn' the best way to speed up the collection of MRI data, allowing each image to be collected up to 50x faster. To achieve these speed-ups, it is necessary to reduce the amount of data that we collect, however this results in significant errors in the images. I have shown that it is possible to reconstruct clinically useful images very quickly using ML; up to 100x faster than current state-of-the-art mathematical methods.

In addition, I have developed fully-automated ML tools for analysis of the MRI images; enabling clinical metrics to be calculated whilst the patient is in the scanner, up to 7000x faster than conventional methods. Combined, these tools have allowed MRI scans in the heart and abdomen to be performed quickly (in as little as 10 minutes) and without the need for breath-holding in children, as well as in sick adults.

In initial studies, these tools have been tested in small patient groups. This work focusses on large-scale international clinical testing, to make sure that these ML tools work reliably and accurately across all children's diseases and in different hospital settings. This work will build trust in these tools, enabling them to be shared them with different hospitals across the world, to maximise the benefits to all patients and hospitals.

I will also continue benefitting from new developments in ML to further improve these technologies and overcome any limitations encountered during clinical testing.

Most hospitals only have conventional MRI scanners, with field strengths of 1.5T or 3.0T. However, recently low-field strength (0.55T) MRI scanners have become available. Although these have not yet been clinically established, they offer significant financial benefits, including lower initial cost (~50% of 1.5T), easier/cheaper installation (70% of 1.5T) and lower running/maintenance costs (~45% of 1.5T).

This makes these scanners highly desirable, and may enable MRI to become affordable in some countries for the first time. Low-field scanners, also address some of the remaining challenges of MRI in children; i) They have a bigger bore, so children find them less daunting, ii) They are much quieter, which means children may be able to remain asleep, and iii) There are less concerns over heating in the body.

However, the measured signal at low-field strength is <25% of that on conventional scanners, resulting in lower quality images. Therefore, the second part of this extension will build on the ML tools that I have developed for imaging children on conventional scanners, to enable good quality images from rapid scans at low-field MRI. This includes the use of ML to 'find' the best way to collect the data, and ML methods to improve image quality.

This extension will work towards clinical validation of these fast scans in children on conventional scanners, making MRI less difficult or daunting for children, improving availability and reducing risks. Quicker scans would help reduce waiting lists and costs for the NHS, and improve diagnostic accuracy and outcomes in childhood diseases. It would also mean that MRI scanning would be used far more often, so it could help many more children.

Additionally, by investigating the use of rapid imaging tools at low-field MRI, I will continue to push novel research ideas, enabling improved image quality near to the lungs, in the abdomen and in fetus', with lower risk to patients and significant cost benefits for hospitals.