Deep learning-based MR image reconstruction: coil-by-coil reconstruction versus direct mapping

Motivation: Deep learning has been increasingly applied in image reconstruction. However, existing methods have limitations. Goal(s): We propose DL-based multi-coil image reconstruction methods that exploit data redundancy in the coil dimension. Approach: New coil-by-coil reconstruction approaches are proposed, where individual coil images are reconstructed and combined with or without incorporating coil sensitivity. Additionally, a direct mapping model predicts coil-combined images directly from undersampled coil images. Results: Using the proposed methods, high-quality knee images are derived with 6× and 10× acceleration. Incorporation of conventionally calculated coil sensitivity improves coil-by-coil image reconstruction. The direct mapping approach demonstrates slightly better performance even without including coil sensitivity. Impact: Comparing DL-based direct mapping with coil-by-coil reconstruction and incorporating coil sensitivity in different ways are two fundamental questions in multi-coil MRI reconstruction. This work provides evidence that implicit estimation and integrated use of coil sensitivities may provide improved reconstruction.