WHEN THE BRAIN MEETS THE STARS: KNOWLEDGE MADE VISIBLE TO THE NAKED EYE BY DR PHILIPPE CIUCIU (CEA, NEUROSPIN) The data acquisition mechanisms in radio-astronomy and in magnetic resonance imaging (MRI) present similarities. In either case, the acquisition system produces data in a mathematical space, called the Fourier space, and the transition from incomplete measurements to good quality images is far from obvious (see middle figure in the image below). Experts in radio interferometry and MRI have teamed up to improve image quality in both scientific fields. They both have a common goal: getting the cleanest image possible in the fastest observation time. When teams at the CEA in France (Commissariat à l’énergie atomique et aux énergies alternatives) joined their efforts in the COSMIC project (Compressed Sensing for Magnetic resonance Imaging and Cosmology) they showed what scientific teamwork could do by winning second place in the Brain FastMRI Challenge. This international competition was held in October 2020 and co-sponsored by Facebook AI Research and NYU
Langone Health. The goal was to see how accurately an image could be made to resemble the original from information that was incomplete. Results showed that the CEA’s method provided an excellent means to reconstruct an image to be almost the same as what could be found in an original image, even with data that was far from being optimal (degraded and unclear). (See right figures in the image below.)
To do this, these two teams (one working on the brain and the other on astrophysics) used Compressed Sensing (CS), a mathematical theory invented in 2006 by two different groups in California (CalTech and Stanford).
Above: An overview of the acquisition and reconstruction process for both brain imaging and radio-astronomy. MRI and radio-telescope collecting data (left) and sampling in a mathematical space called Fourier domain (centre). In practice, we cannot have a full sampling of this space and we need to reconstruct images from subsampled data. Right, example of such reconstructions from retrospectively subsampled data with two different methods: a linear inversion method called GRAPPA that is implemented in actual MRI scanners (middle right) and the deep learning (bottom right). The original image is shown top right. Credit: Dr Philippe Ciuciu, CEA; SKAO, ICRAR, SARAO. Top right: The team, clockwise from left: Dr Jean-Luc Starck (Cosmostat, CEA, CNRS) and Dr Philippe Ciuciu (NeuroSpin, équipe PARIETAL Inria-CEA), and PhD student Zaccharie Ramzi. 25