SPGRE-to-bSSFP Sequence-Mapping with U-Net and GAN for Data Homogenization improves Comparability of PREFUL MRI

Motivation: Pulmonary proton MRI in free-breathing typically utilizes either spoiled or balanced steady-state free precession sequence types (SPGRE/bSSFP), which leads to sequence-dependent results variability. Goal(s): Since deep learning networks are known to be able to translate between different image types, image-mapping from SPGRE to bSSFP is hypothesized to yield more similar results. Approach: After training a U-Net, alone and part of a generative adversarial network pair (GAN), similarity of results was assessed, including SSIM, MSE, image sharpness and perfusion defect percentage. Results: Results showed less image sharpness and were significantly more similar after mapping. Both network architectures performed on par. Impact: As sequence homogenization is limited by vendor standards and hardware-limits the demonstrated sequence-mapping approach via deep learning is viable alternative. It could be specifically used to decrease the variability of perfusion-weighted maps acquired with bSSFP and SPGRE in multicenter settings.