Quentin Raynaud¹, Jérôme Yerly^2,3, Ruud B. van Heeswijk², and Antoine Lutti^1
1Clinical Neuroscience, Laboratory for Research in Neuroimaging, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, ²Diagnostic and Interventional Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, ³Center for Biomedical Imaging, Lausanne, Switzerland

Cardiac-related noise reduces the sensitivity of relaxometry data to brain microstructure properties such as iron or myelin concentration. To address this effect, we aim to characterize cardiac pulsation effects in in-vivo data. We combined a Cartesian pseudo-spiral sampling trajectory with compressed sensing image reconstruction using a temporal total variation regularization to obtain 3D multi-echo images at 12 points of the cardiac cycle. We show that 76% of the full k-space is required in each cardiac bin to maintain 90% of voxels above a 60% sensitivity to physiological noise in the reconstructed images. 

Moritz Blumenthal¹, Guanxiong Luo¹, Martin Schilling¹, Markus Haltmeier², and Martin Uecker^1,3,4
1University Medical Center Göttingen, Göttingen, Germany, ²Department of Mathematics, University of Innsbruck, Innsbruck, Austria, ³Institute of Medical Engineering, Graz University of Technology, Graz, Austria, ⁴DZHK (German Centre for Cardiovascular Research), Göttingen, Germany

In this work, we propose NLINV-Net, a neural network architecture for jointly estimating the image and coil sensitivity maps of radial cardiac real-time data. NLINV-Net is inspired by NLINV and solves the non-linear formulation of the SENSE inverse problem by unrolling the iteratively regularized Gauss-Newton method, which is improved by adding neural network based regularization terms. NLINV-Net is trained in a self-supervised fashion, which is crucial for cardiac real-time data which lack any ground truth reference. NLINV-Net significantly reduces noise and streaking artifacts compared to reconstructions using plain NLINV.

Xiaoke Wang¹, Danial Ludwig², Michael Rawson², Radu V Balan², and Thomas Ernst^1
1Diagnostic Radiology, University of Maryland-Baltimore, Baltimore, MD, United States, ²Mathematics, University of Maryland-College Park, College Park, MD, United States

Image reconstructions involving neural networks (NNs) are generally non-iterative and computationally efficient. However, without analytical expression describing the reconstruction process, the compuation of noise propagation becomes difficult. Automated differentiation allows rapid computation of derivatives without an analytical expression. In this work, the feasibility of computing noise propagation with automated differentiation was investigated. The noise propagation of image reconstruction by End-to-end variational-neural-network was estimated using automated differentiation and compared with Monte-Carlo simulation. The root-mean-square error (RMSE) map showed great agreement between automated differentiation and Monte-Carlo simulation over a wide range of SNRs.

Ricardo Otazo^1,2, Ramin Jafari¹, and Masoud Zarepisheh^1
1Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, United States, ²Department of Radiology, Memorial Sloan Kettering Cancer Center, New York, NY, United States

Machine learning, and specifically deep learning, has recently demonstrated high-performance reconstruction of undersampled k-space data. However, training of deep learning reconstruction methods does not take advantage of the thousands of previously acquired images that are stored in clinical databases. Inspired by the eigenfaces method from computer vision, where an image model derived from thousands of pre-existing images is used to identify new faces, this work presents an image reconstruction method named EigenMRI that learns a data-driven regularization approach using thousands of images extracted from a clinical database to reconstruct 3D brain images with 2D acceleration

Salman Ul Hassan Dar^1,2, Yilmaz Korkmaz^1,2, and Tolga Cukur^1,2,3
1Department of Electrical and Electronics Engineering, Bilkent University, Ankara, Turkey, ²National Magnetic Resonance Research Center (UMRAM), Bilkent University, Ankara, Turkey, ³Neuroscience Program, Aysel Sabuncu Brain Research Center, Bilkent University, Ankara, Turkey

Deep neural network models have demonstrated state-of-the-art performance in MR image reconstruction. These models require information regarding imaging operators during training, which limits their generalization. A recent framework is based on zero-shot learned generative models that learn MR image priors during training and couple imaging operator during inference on test acquisitions. Such models are however based on convolutional architectures that suffer from sub-optimal capture of long range dependencies. Here, we propose a novel architecture based on zero-shot learned generative adversarial transformers that enables efficient capture of long range dependencies via cross-attention transformers while removing reliance on imaging operator during training.