Micro Transformer Cascades for Accelerated Multi-Coil MRI Reconstruction with Overlapped SwinV2 Attention

Motivation: Shifted Window (Swin) Vision Transformer cascades with hybrid attention have been shown to be effective at undersampled MRI reconstruction, but their large memory requirement makes long cascades computationally expensive. Goal(s): In this work, we aim to reduce the memory footprint of the SwinV2 Transformer to improve its suitability for cascaded MRI reconstruction models. Approach: We introduce the new SwinV2-Micro Transformer which allows the formation of longer cascades with overlapped attention and test our model against previously developed hybrid SwinV2-Tiny and other state-of-the-art methods. Results: Our proposed SwinV2-Micro cascade outperforms the competing methods in multi-channel MRI reconstruction at multiple acceleration factors. Impact: A smaller than conventional SwinV2-Micro Transformer architecture is proposed to facilitate the incorporation of powerful overlapped attention in Transformer cascades for MRI reconstruction. This allows the formation of longer cascades enabling higher reconstruction quality at different acceleration factors.