CEST MRI: a neural network reconstructs images in 1.7 seconds that previously took 45 minutes, enabling a scan of the entire brain in 3 minutes

Optimised for MRI saturation transfer imaging, the Movienet network enables the acquisition of four times less data whilst maintaining a correlation of 0.965 with the reference method.

Chemical exchange saturation transfer magnetic resonance imaging (CEST-MRF) enables the mapping of fine tissue properties, but the scan and its processing are time-consuming. Researchers have proposed on arXiv a method to speed up the technique using Movienet, a deep learning network adapted to exploit temporal and inter-receiver-antenna correlations in subsampled radial acquisitions.

Eight rays instead of 34

Ten healthy volunteers underwent repeated scans. The authors trained models with a number of acquisition beams ranging from 1 to 34 per scan, and compared their reconstructions to the iterative reference method GRASP-LLR. With 8 rays, the median correlation reached 0.947 and the error was 8.2 per cent; in cross-validation across the ten subjects, the median image correlation was 0.965 with an error of 7.5 per cent.

The time saving is spectacular: a series of 30 slices is reconstructed in 1.7 seconds, compared with around 45 minutes using GRASP-LLR. The team concludes that a whole-brain CEST-MRF scan is now feasible with an acquisition time of 3 minutes and reconstruction in less than two seconds.