A Mixed Precision Conjugate Gradient Method for Efficient Iterative MRI reconstruction

Motivation: Iterative MRI reconstruction algorithms are computational expensive requiring expensive compute to implement. Goal(s): Investigate the feasibility of mixed-precision arithmetic for an iterative MRI reconstruction algorithm. Approach: Different mixed precision versions of conjugate gradient method were implemented with 32-bit floating point, 16-bit floating point, and 8-bit floating point representations and the quality of the reconstructed image was assessed. Results: The mixed-precision floating point 8-bit algorithm shows numerical instability but decent performance with an increased in regularization compared to 32-bit and 16-bit floating point versions. Impact: Iterative reconstruction algorithms could possibly be implemented on low performance and cheaper compute hardware with further algorithmic stability enhancements.