1st supervisor: Claudia Prieto, King’s College London
2nd supervisor: Julia Schnabel, King’s College London
Magnetic Resonance Imaging (MRI) has become an important non-invasive tool for risk assessment and treatment monitoring of cardiovascular disease. Conventional MR images are qualitative measurements that depend on different parameters such as the longitudinal T1 and the transverse T2 relaxation times. Recently, T1 and T2 mapping techniques are emerging to provide quantitative tissue characterization and objective assessment of myocardial tissue properties1. These techniques are changing the MR paradigm from visualization to quantification and offering the promise of early disease detection and monitoring over time or in response to therapy.
However, clinically used cardiac parametric mapping methods still present several limitations in terms of accuracy, precision, robustness, reproducibility, coverage (usually limited to 2D images), spatial-resolution and long acquisition times. Moreover, T1 and T2 maps need to be acquired sequentially, further increasing the scan time. Magnetic Resonance Fingerprinting (MRF) is a novel technique that promises to alleviate most of these problems2. The key idea of MRF is to use a pseudo-randomized acquisition scheme that causes the signals from different tissues to have a unique signal evolution or “fingerprint”. Matching the measured MR signal response (highly undersampled time-point images) to a previously generated dictionary of fingerprints allows MR tissue identification and parameter estimation. Fingerprints are designed as a simultaneous function of multiple tissue parameters and therefore several MR parameters can be reconstructed from the same single acquisition.
MRF has shown initial promising results for 2D myocardial tissue characterisation3,4, however several challenges need to be tackled to allow the application of MRF in 3D cardiac imaging. Physiological respiratory motion influences the MR signal formation, changing the corresponding tissue specific MRF signals, and therefore need to be taken into consideration. Highly accelerated time-point images, required to achieve clinically reasonable scan times, introduce bias on the parametric maps, and thus efficient and effective undersampled reconstructions need to be employed to reconstruct these images. Conventionally used pixel-wise template matching for mapping generation is computationally inefficient and lacks scalability, and thus fast and accurate map reconstruction compatible with clinical workflow is required.
Here we propose a new deep learning based MRF framework to enable efficient 3D multiparametric myocardial characterization from a single free-breathing acquisition. This will be achieved by:
1) developing 3D respiratory motion compensated cardiac MRF;
2) investigating machine learning as well as deep learning reconstruction techniques to generate good quality time-point images from highly-undersampled 3D cardiac MRF;
3) investigating spatiotemporal convolutional neural networks to enable parametric maps reconstruction in the order of a few seconds;
4) validating the proposed approach in healthy subjects and a reduced cohort of patients with cardiovascular disease, in comparison to current 2D clinical gold standards.
This project encompasses topics from both the Emerging Imaging and AI in Medical Imaging themes of this CDT, and thus will permit the student to work at the interface of both sub-disciplines.
References:
1) Ferreira V et al. Myocardial Tissue Characterization by Magnetic Resonance Imaging: Novel Applications of T1 and T2 Mapping. J Thorac Imaging 2014; 29: 147-154.
2) Ma D et al. Magnetic resonance fingerprinting. Nature 2013; 495:187-192.
3) Hamilton J et al. MR fingerprinting for quantification of myocardial T1, T2 and M0. Mag. Reson. Med. 2017; 77:1446-1458.
4) Jaubert O, Prieto C et al. MORE-MRF: Towards Motion Resolved Cardiac Multi-Parametric Mapping with Magnetic Resonance Fingerprinting. Proc. Intl. Soc. Mag. Reson. Med. 26 (2018), 4261.
5) Oksuz I, et al. Magnetic Resonance Fingerprinting using recurrent neural networks. Submitted to ISBI 20199.
