Ultra-acceleration: compounding AI reconstruction methods with end-to-end training to achieve over fifty-fold acceleration.

Motivation: Faster MRI is needed for MRI-guided interventional procedures, MRI-guided radiation therapy, and to reduce the time and economic cost-burden of MRI on health systems. Goal(s): To test the limits of accelerated MRI acquisition by compounding multiple techniques simultaneously in a method we term 'ultra-acceleration'. Approach: We developed an ultra-acceleration model based on AI to simultaneously reconstruct and apply super-resolution at 50× and 70× ultra-acceleration. Results: The normalised mean-square-error (mean ± standard deviation) was 0.056 ± 0.009 and 0.065 ± 0.011 for 50× and 70× ultra-acceleration respectively. Ultra-acceleration produced images with high visual quality and minimal aliasing. Impact: Our work demonstrates compounding acceleration techniques can shorten the acquisition time of MRI by over 50 times. This finding is a step towards real-time volumetric MRI for interventional guidance.