Motion detection and correction in neuro MRI using RF sensors and learning from k-space data

Magnetic Resonance Imaging (MRI) continues to be affected by unintended and physiological movement of the subject. This is because image measurement is slow, tacking between 2 and 10 minutes for a single image. Images corrupted by motion lead to loss of expensive scan time, and reduced operation efficacity.
A number of solutions have been proposed; however, each are tuned to specific acquisitions or to the measurement of specific types of motion (rigid body / cardiac). A recent development uses the RF coils that exist as part of the MRI hardware (1-3). These coils are spatially located around the subject and their electrical characteristics change in relation to changes in nearby tissue (subject motion). Our preliminary work has shown that using a dedicated calibration scan, this data can measure head motion.

[Figure 1: head motion predicted using rf coils, demonstrating a high temporal resolution for motion information. This was achieved with a 15 minute subject specific calibration]
However, it is not possible or practical to calibrate these measurements for each subject. Other recent work has demonstrated that raw k-space acquisitions can be used to quantify subject motion(4) using redundant information from multiple receive channels. This method is limited to 3D acquisitions where sufficient parallel imaging redundancy exists, therefore alternative model based solutions will also be explored. Subject motion is typically characterized by few abrupt changes amongst relatively lengthy static positions. Compressed sensing established that such sparse innovations can be accurately determined from a number of measurements proportional to the number of abrupt changes. This project will combine techniques from compressed sensing with deep learning in order to automate motion detection and compensation so as to optimize the scan rate efficiency, reduce the need for rescanning, and improve overall diagnostic quality by making use of as much acquired data as possible.
This work proposes the application of machine learning methods to quantify motion, with high accuracy and high temporal resolution. This can be achieved by learning the relationship between the RF sensors and motion using the partial information available in the raw k-space data.
This work is at the interface between the academic interests - Hess with RF sensors for motion, Tanner with compressed sensing and deep learning including few-shot learning techniques, and Mailhe/Siemens in optimisation of MRI acquisitions.