Machine learning approaches to enabling ultra-fast diagnostic MRI protocols for neurology

1. Brief description of the context of the research including potential impact

Magnetic resonance imaging (MRI) is proven as the diagnostic imaging method of choice for a wide range of neurological conditions. However, it is used less often than competing modalities such as CT due to its expense and the longer time taken to acquire images. Additionally, the number of installed MRI machines is lower than the number of CT scanners, with many healthcare organisations having older machines not capable of the latest high-quality imaging.

This mixture of challenges around the use of MRI means that it is used less often than CT, even though in many cases it is the more appropriate choice, providing greater diagnostic sensitivity and specificity.

An important area where this has impact is in the diagnosis of Alzheimer's disease (AD) where NICE guidelines specify using structural imaging to rule out reversible causes of cognitive decline and to assist with subtype diagnosis. Ideally, this means scheduling an MRI scan, due to its lack of ionising radiation, excellent soft-tissue contrast, and superiority over other imaging techniques in identifying vascular dementia or when the subtype is uncertain.

This information allows differential diagnosis, which may alter management and enhance prognostication, unlike CT.

If the scan time, availability, and cost for an MRI scan were comparable to a CT scan, its benefits mean that MRI would be used in almost all cases. The key to solving each of these problems is a substantial reduction in the duration of a diagnostic scan for Alzheimer's disease. Shorter scans would be easier to schedule in the overall diagnostic patient pathway, thereby improving availability of appropriate imaging to patients and providers.

Shorter scans are also less expensive, as cost is driven to a large degree by the scan time. The patient experience would also be improved by less time in the scanner as anxiety is reduced, and more time is available for staff.

However, achieving shorter times has been problematic to date because the required scan time reduction leads to unacceptable image quality degradation. Additionally, if older MRI machines acquisitions could be improved in quality, then the overall availability of suitable scanning would also be increased. A key scientific challenge of this PhD project is the development of a combination of new ultra-fast MRI and machine learning methods for reconstruction and analysis that can provide equivalent diagnostic information to conventional diagnostic MRI.

2. Aims and Objectives The specific objectives are to: Evaluate existing machine learning methods to accelerate MRI acquisition

Develop and evaluate the use of Image Quality Transfer (IQT) methods as applied to MR image reconstruction and quality improvement Apply and assess IQT methods on ultra-fast acquired MRI scans for the detection of Alzheimer's disease Apply and assess IQT methods on ultra-low field acquired MRI scans for the detection of Alzheimer's disease

Develop and maintain tools and workflows for efficient application of IQT and related methods in a translational research environment. 3. Novelty of Research Methodology

Until now there has been no previous attempt of using Image Quality Transfer methods to enable accelerated scans for dementia. The challenge of creating standard of care images (T1, T2, SWI) from significantly degraded rapid scans will require new machine learning methods to be developed, implemented on clinical images, and subsequently evaluated.

4. Alignment to EPSRC's strategies and research areas Aligned with the EPSRC themes on Artificial Intelligence and Healthcare Technologies 5. Any companies or collaborators involved There is a possibility that there may be collaboration with MRI scanner manufactures, but this is yet to be confirmed.