Abstract
Purpose: To introduce a combined machine learning (ML) and physics-basedimage reconstruction framework that enables navigator-free, highly acceleratedmultishot echo planar imaging (msEPI), and demonstrate its application inhigh-resolution structural and diffusion imaging. Methods: Singleshot EPI is an efficient encoding technique, but does not lenditself well to high-resolution imaging due to severe distortion artifacts andblurring. While msEPI can mitigate these artifacts, high-quality msEPI has beenelusive because of phase mismatch arising from shot-to-shot variations whichpreclude the combination of the multiple-shot data into a single image. Weemploy deep learning to obtain an interim image with minimal artifacts, whichpermits estimation of image phase variations due to shot-to-shot changes. Thesevariations are then included in a Joint Virtual Coil Sensitivity Encoding(JVC-SENSE) reconstruction to utilize data from all shots and improve upon theML solution. Results: Our combined ML + physics approach enabled Rinplane x MultiBand (MB)= 8x2-fold acceleration using 2 EPI-shots for multi-echo imaging, so thatwhole-brain T2 and T2* parameter maps could be derived from an 8.3 secacquisition at 1x1x3mm3 resolution. This has also allowed high-resolutiondiffusion imaging with high geometric fidelity using 5-shots at Rinplane x MB =9x2-fold acceleration. To make these possible, we extended the state-of-the-artMUSSELS reconstruction technique to Simultaneous MultiSlice (SMS) encoding andused it as an input to our ML network. Conclusion: Combination of ML and JVC-SENSE enabled navigator-free msEPI athigher accelerations than previously possible while using fewer shots, withreduced vulnerability to poor generalizability and poor acceptance ofend-to-end ML approaches.