​Deep learning is an artificial intelligence approach in which artificial neural networks are trained to recognize an object, an image for example. Applied to magnetic resonance imaging (MRI), deep learning allows, like traditional reconstruction algorithms, to reconstruct images from a reduced number of collected data according to compressed acquisition schemes and reduced examination times. There is a true dual benefit to using deep learning: it allows to recover images of a quality similar to those obtained from complete data sets and more notably to significantly accelerate (100 times faster) this reconstruction process compared to previous methods.

Deep learning in MRI is starting to offer promising results. Many artificial neural networks are being developed, but how can we judge the performance of each one? Comparisons remain hard because the frameworks used are not the same among studies, the networks are not properly re-trained, and the datasets used are not always the same among comparisons. 

In order to develop deep learning for MRI, Facebook AI Research and NYU Langone Health recently launched a collaborative research project (fastMRI). In this framework, NYU Langone Health has made public complete sets of raw knee MRI data (several hundred examinations with different imaging sequences: T2 weighting with or without fat saturation, different imagers; different acquisition contexts: single or multi-channel antenna to improve signal reception). The data are freely available and have recently been supplemented by cerebral MRI scans (about 7000 exams, see https://fastmri.med.nyu.edu/ for further details).

This recent release encouradged researchers from NeuroSpin (PARIETAL INRIA-CEA team) and Cosmostat (Irfu CEA-CNRS) to write a consistent benchmark of several deep neural networks for MR image reconstruction, using two sets of data: raw knee MRI data from fastMRI and brain MRI data from OASIS, also freely available. In total, they compared the performance of four deep neural networks, some acting in a single space (in the k-space of data acquisition or in the image space, single-domain networks), others acting in the two spaces alternately (hybrid learning, cross-domains networks). In a paper published in Applied Sciences, the authors unveil the results of their analysis comparing the networks and link the open source implementation available on Github of all these networks in Keras. Their main conclusion is, on the one hand, that cross-domains networks offer better performance than single-domain networks and, on the other hand, that it seems more advantageous to perform more iterations between the two spaces than to have a deeper network in one of the two spaces.

The neural network architectures now proposed by the NeuroSpin team (see updnet_v3 by Nspin on https://fastmri.org/leaderboards for multi-channel knee imaging) are at the top of the ranking established according to the NMSE image quality criterion and in 4th position according to the SSIM score (two complementary measures reflecting the quality of the reconstructions).

All this work was made possible thanks to the support of the D3P (DRF-CEA)[1] (F. Boillot-Cerneux and Ch. Calvin for the Numerical Simulation, HPC and AI sector) to obtain a fast access to the Jean Zay supercomputer with a significant number of hours of calculation on Nvidia V100 multiple GPU architecture.