Generalize or Specialize? The Effect of Coil Combining in Deep Learning Based MR Image Reconstruction

With the advent of deep-learning based MR image reconstruction, new questions are raised regarding best design choices. In particular, when using recent reconstruction models that have been developed for processing multi-channel MR data. A fundamental question is whether to first combine channels to improve model generalizability (i.e., a “coil-combined” approach) or to keep channel processing separate to fully utilize multi-channel information (i.e., an “all-coil” approach). In this work, we compare three popular architectures using coil-combined and all-coil styles on brain MR images. All-coil styles improved in-distribution performance, such as when reconstructing only presumed healthy individu-als. Coil-combined designs better generalized to unseen data from patients with pathology.