Student: Ke Wen
1st Supervisor: Sonia Nielles-Vallespin, Imperial College London
2nd Supervisor: Daniel Rueckert, Imperial College London
Additional Supervisors: Guang Yang, Andrew Scott, Pedro Ferreira, Imperial College London
Clinical Champion: Ranil de Silva, Prof Dudley Pennell, Imperial College London
Aim of the PhD Project:
- 3-times accelerated whole-heart in vivo diffusion tensor cardiac magnetic resonance (DT-CMR).
- Development of 3D and simultaneous multi-slice (SMS) DT-CMR acquisitions
- Development and optimisation of AI algorithms to minimise artefacts in 3D and SMS DT-CMR images.
- Integration of these algorithms in the CMR image reconstruction pipeline for clinical translation.
Project Description / Background:
In vivo diffusion tensor cardiac magnetic resonance (DT-CMR) provides a means for non-invasive interrogation of the 3D microarchitecture of the beating heart, which no other clinical test allows1,2. This unique and innovative technology offers tremendous potential to improve clinical diagnosis through novel microstructural and functional assessment. DT-CMR data acquisition is currently very inefficient; ~10 minutes are needed to acquire a single 2D slice through the heart at one timepoint in the cardiac cycle. Whole-heart DT-CMR coverage is essential for accurate diagnosis in many cardiac diseases such as myocardial Infarction (MI) which affects the heart in a heterogeneous fashion. Improved efficiency is therefore key to DT-CMR clinical translation.
3D acquisitions allow increased coverage and SNR. Currently, 2D DT-CMR methods use single-shot readouts that acquire all the data needed for one image in one cardiac cycle. 3D DT-CMR would require segmented readouts that acquire all the data needed for one volume over several cardiac cycles. Segmentation combined with diffusion encoding leads to phase error artefacts. Furthermore, longer 3D readouts lead to distortion artefacts2.
Simultaneous multi-slice (SMS) techniques acquire multiple 2D slices simultaneously but reconstruct them separately, improving imaging efficiency tremendously. While SMS techniques have been successfully used in neuroimaging3, so far only proof-of-concept SMS DT-CMR studies have been published4,5. Since the heart is a smaller organ than the brain, the distance between multiple slices is smaller. The SMS algorithm often fails in these circumstances and information from one slice erroneously ends up on a different slice (“interslice-leakage” artefact).
Artificial Intelligence (AI) algorithms can learn complex relationships or patterns from training images and make accurate predictions on newly acquired images. AI algorithms can be trained to reconstruct highly undersampled 3D acquisitions, correct distortion artefacts and minimise motion-induced phase errors. SMS is also well poised to benefit from AI algorithms because multi-slice training data sets can easily be acquired. Ex vivo DT-CMR data can provide an excellent test case devoid of cardiac and respiratory motion, while in vivo DT-CMR data can be acquired in healthy volunteers and as well as patient cohorts.
In this project, Generative Adversarial Networks and other novel AI algorithms will be developed to minimise/eliminate artefacts in 3D DT-CMR and SMS DT-CMR ex vivo and in vivo data. These novel algorithms will be optimised to allow for the maximum possible speed-up factors, aiming for 3-fold accelerated whole-heart coverage (i.e. a whole-heart 9-slice protocol which currently takes ~90 minutes would require ~30 minutes with the targeted SMS acceleration factor of 3). These AI enabled in vivo whole-heart DT-CMR techniques will then be integrated in the clinical CMR image reconstruction pipeline for clinical translation. Reproducibility of these AI algorithms will be tested in healthy volunteers and patients with MI.
This project is suitable for a motivated student with a background in physics, engineering, computer science or mathematics and experience in coding (Python, C++).
Figure 1: Multi-slice and SMS DT-CMR acquisitions in an ex vivo heart. As the distance between simultaneously acquired slices decreases, the image quality of the diffusion weighted images degrades due to interslice-leakage artefacts. Even at the maximum distance, although the diffusion weighted images appear artefact-free, the maps clearly demonstrate interslice-leakage artefact is still present.
References:
- Nielles-Vallespin JACC 2017. doi:10.1016/j.jacc.2016.11.051
- Nielles-Vallespin JMRI 2019. doi:10.1002/jmri.26912
- Setsompop MRM 2012. doi:10.1002/mrm.23097
- Lau MRM 2015. doi:10.1002/mrm.25200
- Mekkaoui Radiology 2018. doi:10.1148/radiol.2016152613