A machine-learning approach to solving the SAR problem for ultrahigh field MRI

1st Supervisor: Shaihan Malik, King’s College London
2nd Supervisor: Jo Hajnal, King’s College London

Ultrahigh field (UHF) MRI has the potential to provide very high resolution anatomical images with new types of contrast compared with more conventional lower field MRI. This is already starting to bear fruit for neuroimaging, where the high resolution has been used to create unprecedentedly detailed functional maps or detection of previously unseen lesions associated with disorders such as epilepsy, for example.

Use of this powerful tool for body imaging has so far been less successful. This is because the high frequency radio-frequency (RF) fields needed for UHF MRI are highly spatially non-uniform inside the body. Research has shown that using multiple transmitters (so called parallel transmit, pTx) can be used to overcome this problem. However safety calculations required to quantify potential RF heating effects are made more complex because the fields from multiple transmitters can interfere to generate ‘hotspots’ that move inside the body. As a result, current body UHF-MRI requires conservative safety limits and expert operators, stopping the method from reaching its true potential.

The aim of this project is to use emerging electromagnetic (EM) calculation methods combined with machine learning to move towards population level characterisation of EM interactions with human subjects, providing confidence in safety estimation and solving the workflow issues that currently block progress for UHF MRI.

Safety calculations in MRI require prediction of electric fields that can be summarized by a ‘specific absorption rate’ quantifying the spatially variable rate of energy deposition in the body from the RF fields. SAR cannot easily be measured in-vivo, hence we rely on computer simulations. For standard MRI systems a few ‘standard’ human models are simulated, and then safety limits are set given an acceptable margin for error. This approach breaks down when using array transmit coils placed close to the subject’s body, since the fields produced are likely to be highly subject and position dependent. EM simulations are far too time consuming (they take hours or days) for a subject-specific model to be generated. Hence at the moment the only way around this is to use very conservative operational limits.

In this project we will use emerging new EM simulation methods to produce a large range of simulated data for a range of human body shapes and sizes and coil designs and positions, with the objective of generating the expected variability seen at a population level. We will then exploit this resource using machine learning methods to extract the main expected types of variation and identify features of the human subject that will allow us to predict the fields for a given human subject. The end objective is to provide an enabling development for UHF-MRI in the body: a bespoke safety assessment with quantified uncertainty for a given subject without extensive subject-specific calibration. If successful the project will lead to a transformation in how UHF-MRI can be applied to body imaging, for a huge range of potential applications.

Figure 1. Description is in the caption.