Accelerating High Resolution 3D EPI with Deep Learning Reconstruction

Motivation: To increase access to clinically relevant features disambiguated from partial volume based confounds by enabling ultra-high spatial resolution without clinically restrictive scan times. Goal(s): To investigate the feasibility of accelerating ultra-high resolution 3D EPI for rapid brain imaging, with DL based image reconstruction. Approach: Multi-shot and single-shot 3D EPI data were retrospectively undersampled with tiled variable-density Poisson-disc (VDPD) sampling patterns and then reconstructed by DL Speed, a DL image reconstruction method that we developed. Results: We demonstrated DL Speed can achieve an acceleration factor of 10 for 3D EPI while maintaining image quality compared to fully sampled data. Impact: Deep learning based sparse image reconstruction can accelerate ultra-high resolution 3D EPI scans for brain imaging with acceleration factors ranging from 3-10, enabling disambiguation of clinically relevant fine features in various neuro imaging applications such as SWI, DWI and fMRI.