Deep Learning for Real-Time Motion-Correction in Parallel-Transmit, High-Resolution MRI of Patients with Movement Disorders at 7 Tesla

: MRI scanners, operating at 7T offer substantial advantages over conventional clinical MRI systems operating at 1.5T and 3T, including enhanced sensitivity, accelerated imaging, and superior spatial resolution. However, some patient groups, who would benefit from this technology, are prone to motion

disorders that limit the application of the technique, particularly with respect to high-resolution imaging. Thisincludes patients with pathologies such as Parkinson's disease, essential tremor, tardive dyskinesia, and Huntington's Chorea. A further challenge for 7T MRI is the inhomogeneity of the radiofrequency

(RF) transmit field, which results in images with poor uniformity that can hamper diagnosis. In recent years, the University of Glasgow, in collaboration with Siemens Healthineers, has sought to address the challenge of poor image uniformity by exploring the utility of parallel transmission (pTx) using a range of novel head coils, developed locally by Dr Shajan Gunamony (University of Glasgow

and MR CoilTech Limited) [1]. This work has demonstrated that a substantial improvement in image uniformity can be achieved across the whole head for multiple image contrasts by using dedicated pTx RF pulses, which apply waveforms independently to eight transmit elements within the head coil [2,3]. The process of designing these waveforms is computationally demanding and in some cases this is done

individually for each subject during the scanning session, adding to the examination time. The University MRI Physics Group has also been investigating real-time motion correction in MRI by using image registration during the scan to update scanner parameters with sub-second temporal resolution. A recently completed PhD project [4] demonstrated that this method can substantially reduce

motion sensitivity in 7T functional neuroscience studies, which are easily corrupted by motion. The project also demonstrated how motion correction and pTx methods can be integrated effectively in a single acquisition [4,5].

The proposed PhD project aims to extend our work in pTx and motion correction by exploiting deep learning methods to substantially reduce data-processing times, which we anticipate will have a major impact on real-time computation. Firstly, this will allow advanced subject-specific pTx waveforms to be calculated with a further improvement in image uniformity, but without an increase in measurement

time compared to standard techniques. Secondly, it will reduce the time required for real-time calculations in motion correction, so that scanner updates can respond to the higher frequency motion components in subjects with movement disorders. Ultimately, we seek to develop an MRI acquisition framework that can improve clinical outcomes in patients with movement disorders, such as Parkinson's

disease, where high-resolution, high-precision imaging is crucial in surgical planning of deep brain stimulation.