New Approaches to Simultaneous Multislice Magnetic Resonance Imaging : Sequence Optimization and Deep Learning based Image Reconstruction
Eickel, Klaus
Universität Bremen: Physik/Elektrotechnik
Magnetic Resonance Imaging, MRI, Deep Learning, Image Reconstruction, Perfusion Imaging, Tracer Kinetics, Pharmacokinetics, Simultaneous Multi-Slice Imaging, SMS, Parallel Imaging, Medical Physics, Medical Imaging
Magnetic resonance imaging (MRI) is a versatile imaging modality in clinical diagnostics. Despite the impressive range of application, a main drawback of MRI is its inherently low acquisition speed. However, scan time is crucial for many applications and also for an efficient utilization of MRI in clinical routine. Two developments have influenced MRI recently: Simultaneous multislice imaging (SMS) and deep learning (DL). Simultaneous multislice imaging is a paradigm shift in MRI which has re-emerged in the early 2010'. It yields improved image quality compared to in-plane parallel imaging, because it benefits from increased signal-to-noise ratio and robustness for higher accelerations. SMS sequences accelerate data acquisition by undersampling along the slice dimension and specific algorithms allow reconstruction of these undersampled data. In the first part, SMS was extended to measure multiple image contrasts in contrast-enhanced dynamic MRI. Therefore, a bespoke MRI sequence was developed to accelerate segmented echo-planar imaging of three echoes. Dynamic in-vivo data with sufficient spatial coverage were acquired in an animal model. Data acquisition were fast enough to sample the arterial input function which is essential for pharmacokinetic modeling. Imperfections in the excitation of multiple slice and their relevance for reconstruction algorithms were closely investigated and evaluated for processing of multi-contrast data. This work connects SMS and deep learning. Today, the application of deep learning in medicine assists decision making in medical diagnosis, analysis of radiologic data or personalized medicine in genomics. In MRI however, deep learning has just entered the stage. With two abstracts matching the search term 'deep learning' at the ISMRM 2016, the number of abstracts rose to 42 in 2017 and to 139 in 2018. Most of the early contributions to DL in MRI concern image processing and data evaluation. Image reconstruction itself is mostly conducted in standard fashioned way. Common algorithmic approaches applying deep neural networks for (some) processing steps have shown impressive results and can often be generalized to similar problems. In the second part, the separation of overlapping slice content after SMS was performed by an artificial neural network. This novel reconstruction technique, termed SMSnet, does not require any reference data for calibration of the MR machine's receiver characteristics. Omitting the need for reference data could extend the use of modern accelerated imaging sequences to a broad spectrum of applications. Potential and limitations of this approach were investigated in various experiments accounting for image quality, robustness, sensitivity and how the network generalizes. The discussion at the end summarizes and relates the results of this work to state-of-the-art techniques and recent developments in MRI and gives an outlook to future work on SMS and DL-based reconstructions.
New Approaches to Simultaneous Multislice Magnetic Resonance Imaging : Sequence Optimization and Deep Learning based Image Reconstruction
[RefWorks] [Bibtex ]
[EndNote ] [RefMan(RIS) ]