Combined Super Resolution and Partial Fourier Reconstruction for 3D Magnetic Resonance Imaging with varying Partial Fourier settings

Motivation: Current approaches using neural networks for combined super-resolution and partial Fourier (PF) reconstruction require separate models for each PF factor and direction resulting in increased training and maintenance efforts. Goal(s): Develop a single deep learning model capable of handling any PF range and direction, eliminating the need for multiple networks. Approach: Input images for training are zero-padded in the Fourier domain to simulate randomly varying PF from 75% to 100% in all three acquisition directions. Results: Compared to zero-filling and models trained for specific PF factors, the proposed approach shows improved sharpness, reduced ringing artifacts, and enhanced quantitative metrics. Impact: Our unified network performs super-resolution and PF reconstruction across a large range of PF factors applied in arbitrary directions. This removes the need for multiple dedicated networks trained for specific PF factors and simplifies pre- and post-processing operations.