Successful generalization for data with higher or lower resolution than training data resolution in deep learning powered QSM reconstruction

A pipeline to reconstruct multiple resolution QSM data using a QSM network trained at a single resolution is proposed. The local field map is re-sampled multiple times in different spatial locations, and the re-sampled local field maps are used to reconstruct QSM maps at training data resolution. The reconstructed maps are then combined, and corrected for using a procedure named “dipole compensation”. When compared to two scenarios to reconstruct different resolution data using network trained at a single resolution, the proposed pipeline demonstrated the best performance both qualitatively and quantitatively.