Motivation: MRI reconstruction models are typically trained for specific acceleration factors. Multiple model variants are required to accommodate varying acceleration factors in practice. Goal(s): We aim to develop a unified reconstruction model capable of handling arbitrarily accelerated MRI scans. Approach: We first use self-supervised learning to obtain feature representations of fully-sampled and arbitrarily sub-sampled data. Then, we employ latent diffusion model to map feature representations of sub-sampled data to those of fully-sampled data. Results: Experiments show that the use of feature transferring in our unified model brings an average performance gain of 3.86dB in PSNR for acceleration factors of 2x, 3x, and 4x. Impact: We adopt feature representation transfer in the field of MRI reconstruction to address a practical issue largely overlooked by existing studies. Our adaptive reconstruction model can significantly simplify the deployment of MR reconstruction model and reduce the development costs.
Jiang et al. (Tue,) studied this question.