CEA/NeuroSpin & Inria Parietal, CEA Research Director
Date(s) : 22/01/2021 iCal
14 h 30 min - 15 h 30 min
Reducing acquisition time is a major challenge in high-resolution magnetic resonance imaging (MRI) that has been successfully addressed by the Compressed Sensing (CS) theory. Slow CS-MR image reconstruction is however a major bottleneck. In this talk, we address both aspects for collecting data faster (first part) and speeding up MR image reconstruction (second part).
First, most of Fourier encoding schemes nowadays consist in downsampling existing k-space trajectories. Recently, we have overcome this issue by proposing the Spreading Projection Algorithm for Rapid K-space sampLING (SPARKLING) for T2* 2D non-Cartesian imaging and extended this approach to 3D imaging using the usual stacking strategy of 2D SPARKLING (SoS) sampling patterns. In this talk, I will present recent advancements based on a globally optimized 3D SPARKLING extension  and will show how this version outperforms the SoS strategy both on phantom and in vivo human brain data collected at 3 Tesla.
Second, current solution to faster image reconstruction rely on deep learning (DL) and big data. I will present our XPDNet DL architecture that has been ranked second in the 2020 brain fastMRI challenge . I will summarize the main results and showcase XPDNet’s transfer learning capacity on out-of-distribution 7T T2 images collected at NeuroSpin. Last, I will explain how XPDNet can easily handle non-Cartesian data such as SPARKLING.
 Chaithya GR, Weiss P, Massire A, Vignaud A, Ciuciu P. Globally optimized 3D SPARKLING trajectories for high-resolution T2*-weighted Magnetic Resonance Imaging. https://hal.inria.fr/hal-03090471/document
 Muckley MJ, Riemenschneider B, Radmanesh A, Kim S, Jeong G, Ko J, Jun Y, Shin H, Hwang D, Mostapha M, Arberet S. State-of-the-art Machine Learning MRI Reconstruction in 2020: Results of the Second fastMRI Challenge. arXiv preprint arXiv:2012.06318. 2020 Dec 9. https://arxiv.org/pdf/2012.06318.pdf