Hybrid Learning: A Novel Combination of Self-Supervised and Supervised Learning for MRI Reconstruction without High Quality Training Reference

Motivation: High-quality training references are not always available for deep learning reconstruction methods that rely on supervised learning. While self-supervised learning can be beneficial, but it can suffer from compromised performance at high acceleration rates. Goal(s): We introduce a novel concept called hybrid learning for MRI reconstruction in cases where only low-quality reference images are available. Approach: This was implemented in two training phases. Self-supervised learning is first employed to generate high-quality images from low-quality reference data. The obtained high-quality images in the first stage are subsquenetly used for supervised training. Results: This enables high acceleration rates beyond the capabilities of standard self-supervised learning. Impact: This study proposes a novel hybrid learning strategy to address challenges when obtaining high-quality reference data is difficult, which enables more accurate reconstruction at higher acceleration rates, which is beneficial in various applications where only low-quality reference images are available .