Expectation-Maximization inspired training for End-to-End MRI reconstruction without fully sampled data

Motivation: Current self-supervision via data undersampling (SSDU)-based MRI reconstruction algorithms struggle as the input undersampled data for training and inference stages are different. It also relies on the pre-estimated coil sensitivity maps, which may limit its performance. Goal(s): To develop a robust end-to-end model that overcomes these limitations in self-supervised MRI reconstruction. Approach: A method that integrates Expectation-Maximization (EM)-inspired training with automatic coil sensitivity estimation, built on an unrolled Alternating Direction Method of Multipliers (ADMM) reconstruction framework, was proposed. Results: The network achieves state-of-the-art performance, effectively reconstructing images and coil sensitivity maps using only undersampled k-space, and demonstrates significant advantages over traditional methods. Impact: This novel self-supervised training method does not require splitting the undersampled k-space and enables end-to-end MRI reconstruction without pre-estimated coil sensitivity maps. It streamlines self-supervised reconstruction workflows and paves way for further advances in self-supervised reconstruction.