Back to Projects

AI-enabled Imaging, Emerging Imaging

Artificial Intelligence for ‘Self Driving’ MRI

Project ID: 2021_027

Project category: Core

1st Supervisor: Shaihan Malik, King’s College London
2nd Supervisor: Jo Hajnal, King’s College London
Additional Supervisors: Jorge Cardoso, King’s College London
Clinical Champion: Vicky Goh, King’s College London

Aim of the PhD Project:

  • To develop an AI framework that allows MRI scanners to ‘learn’ from experience, reducing/removing the need for time consuming calibration to be performed while the patient is in the scanner.
  • Achieve a step change improvement in workflow, which is one of the major obstacles for ultrahigh field MRI in particular

Project Description / Background:

Magnetic Resonance Imaging is today a vital clinical diagnostic tool, offering high spatial resolution and excellent soft tissue contrast for investigation of an increasing number of clinical indications. Typically image acquisition takes a few minutes per scan, and a single clinical examination consists of multiple scans acquired using different pulse sequences to generate complementary contrasts, typically taking 30-40 minutes. A key challenge with MRI is its expense, and related to this its relative lack of speed in image formation. One reason for this lack of speed is that the underlying nuclear magnetic resonance (NMR) signal is relatively weak. Technological development over the last decades has resulting in scanners with increasingly high main magnetic field strengths (B0), motivated by the supra-linear1 increase this yields in signal strength.

Increasing field strength does however come with some technical challenges. As noted by the RCR2, while 3T scanners provide higher quality images under optimal conditions, image quality is generally more variable compared to 1.5T because of subject induced variability in the B0 field of the scanner and the radiofrequency (RF) fields used for signal excitation (denoted B1+). Both B0 and B1+ fields are distorted by electromagnetic (EM) interactions with the human in the scanner, and these effects typically worsen as the B0 strength is increased. Consequential issues include contrast variation (leading also to quantitation errors), local loss of signal, failure of fat suppression, and geometric distortions. Body imaging is particularly problematic at 3T, for example in the breast, liver, wide FOV whole-body imaging and the fetus in-utero. These technical challenges are experienced to an often greater degree for any type of imaging at 7T.

Previously our group has worked to create patient-adaptive MRI methods that can increase efficiency3 and reduce image artefacts4–6 for 3T and 7T MRI. These methods rely on taking in-situ measurements of EM fields, and calculate the optimal settings for the scanner accordingly. They essentially treat the scan process as a ‘physics experiment’ which uses real world measurements and a physical model to optimize input settings for the scanner. This is a robust and intuitively understandable approach, however it: (i) necessitates taking multiple measurements and performing often time consuming calculations at scan time; and (ii) requires an expert operator. The workflow issue is actually true for more standard simpler ‘calibration steps’ used in day-to-day MRI practice (such as ‘B0 shimming’) which can also absorb significant fractions of available scan time.

This project proposes to replace the ‘physics experiment’ with a ‘learning examination’ paradigm: an AI based approach to MRI that is trained by performing time-consuming optimizations on large quantities of historic data, adapting quickly to new patients at scan time. This will lead to better imaging with less inter-subject variability in quality, and higher efficiency. The highly challenging environment of 7T MRI will be used to develop these methods; improvements will have a significant impact on this (smaller) field. However the overall aim of the project is to produce more general capabilities to impact on clinical MRI workflow in general.

A diagram illustrating the concept of the project. On the left is an MRI image, in the middle is a cartoon representation of a neural network, and on the right a cartoon of an MRI scanner. The meaning is that the MRI image will be used to "drive" the scanner, via a neural network used for processing.

Figure 1: The aim of this project is to allow MRI scanners to ‘learn’ from the data they acquire to inform how they operate in future. Artificial Intelligence techniques will be used to infer optimized settings for subsequent imaging from information that the scanner would acquire anyway, therefore allowing it to become ‘self driving’. This will save significant time by dispensing with calibration that is usually done with the patient in-situ, improving workflow, saving time and hence reducing costs.

References:

  1. Pohmann, R., Speck, O. & Scheffler, K. Signal-to-noise ratio and MR tissue parameters in human brain imaging at 3, 7, and 9.4 tesla using current receive coil arrays. Magn. Reson. Med. 75, 801–809 (2016).
  2. Graves, M., Malcolm, P., Lipton, A., Scurr, E. & Horne, D. Magnetic resonance imaging (MRI) equipment, operations and planning in the NHS Report from the Clinical Imaging Board. (2017).
  3. Beqiri, A., Price, A. N., Padormo, F., Hajnal, J. V & Malik, S. J. Extended {RF} shimming: Sequence-level parallel transmission optimization applied to steady-state free precession {MRI} of the heart. {NMR} Biomed. e3701 (2017). doi:10.1002/nbm.3701
  4. Malik, S. J., Larkman, D. J., O’Regan, D. P. & Hajnal, J. V. Subject-specific water-selective imaging using parallel transmission. Magn. Reson. Med. 63, 988–97 (2010).
  5. Malik, S. J., Keihaninejad, S., Hammers, A. & Hajnal, J. V. Tailored excitation in 3D with spiral nonselective (SPINS) RF pulses. Magn. Reson. Med. 67, 1303–1315 (2012).
  6. Beqiri, A., Hoogduin, H., Sbrizzi, A., Hajnal, J. V. & Malik, S. J. Whole-brain 3D FLAIR at 7T using direct signal control. Magn. Reson. Med. 80, 1533–1545 (2018).
  7. Gras, V., Vignaud, A., Amadon, A., Le Bihan, D. & Boulant, N. Universal pulses: A new concept for calibration-free parallel transmission. Magn. Reson. Med. 77, 635–643 (2017).
  8. Ianni, J. D., Cao, Z. & Grissom, W. A. Machine learning RF shimming: Prediction by iteratively projected ridge regression. Magn. Reson. Med. 1–11 (2018). doi:10.1002/mrm.27192
  9. Bollman, S. et al. Improving FLAIR SAR efficiency by predicting B1-maps at 7T from a standard localizer scan using deep convolutional neural networks. in Proc. Intl. Soc. Mag. Reson. Med. 1129 (2018).
  10. Loktyushin, A., Ehses, P., Schoelkopf, B. & Scheffler, K. Estimating B0 inhomogeneities with projection FID navigator readouts. in Proc. Intl. Soc. Mag. Reson. Med. 98 (2017).
  11. Meliadò, E. F. et al. On-line Subject-Specific Local SAR Assessment by Deep Learning. in Proc. Intl. Soc. Mag. Reson. Med. 293 (2018).
  12. Tomi-Tricot, R. et al. SmartPulse, a machine learning approach for calibration-free dynamic RF shimming: Preliminary study in a clinical environment. Magn. Reson. Med. 1–16 (2019). doi:10.1002/mrm.27870

Back to Projects