“…To facilitate a wide range of further analyses, the signal representation must be model-free, whilst also offering rank-reduction properties needed to make the data self-consistent. Prior work in dMRI slice-to-volume reconstruction has used representations based on the diffusion tensor ( Fogtmann, Seshamani, Kroenke, Cheng, Chapman, Wilm, Rousseau, Studholme, 2014 , Jiang, Xue, Counsell, Anjari, Allsop, Rutherford, Rueckert, Hajnal, 2009 , Marami, Salehi, Afacan, Scherrer, Rollins, Yang, Estroff, Warfield, Gholipour, 2017 , Marami, Scherrer, Afacan, Erem, Warfield, Gholipour, 2016 ), multi-tensor models ( Marami et al., 2018 ), single-shell spherical harmonics ( Deprez et al., 2019 ), or Gaussian processes ( Andersson, Graham, Drobnjak, Zhang, Filippini, Bastiani, 2017 , Scherrer, Gholipour, Warfield, 2012 ), which can limit the supported input data or downstream analysis. Here, we use a data-driven signal representation for multi-shell dMRI based on spherical harmonics and a radial decomposition (SHARD) that was shown to have compelling low-rank properties highly suitable for motion correction applications ( Christiaens et al., 2019a ).…”