Self-Supervised Pretraining of Joint Acquisition and Reconstruction for Fast Quantitative MRI

Motivation: Fourier transform based quantitative MRI (qMRI) relies on decoupled spatial and physical encoding, which limits scan efficiency, thus impedes clinical adoption of qMRI. Goal(s): A new framework was explored to learn optimally efficient spatial-physical encoding purely from data. Approach: An end-to-end model was developed for contemporary spatial-physical acquisition and reconstruction. The model was learnt with large-scale pretraining using self-supervised objective and unlimited natural image data. Results: Results demonstrated such reconstruction model is capable of untangling mixed spatial-physical encoding, and robust to zero-shot transfer from natural image to medical data. With joint learning framework, reconstruction automatically guides the discovery of optimal acquisition. Impact: Large-scale pretrained end-to-end model can be an alternative to long standing Fourier Transform foundation for qMRI. The new capability of contemporary spatial-physical encoding opens up new possibility to push current speed limit of qMRI.