Abhinav Saksena1, Makarand Parigi1, Nicole Seiberlich2, and Yun Jiang2
1EECS, University of Michigan, Ann Arbor, MI, United States, 2Department of Radiology, University of Michigan, Ann Arbor, MI, United States
1EECS, University of Michigan, Ann Arbor, MI, United States, 2Department of Radiology, University of Michigan, Ann Arbor, MI, United States
We investigate both per-pixel (L1) and perceptual based (SSIM) loss functions for deep learning based MRI reconstruction.
A loss function that linearly combines MSSSIM and L1 (with a higher weightage towards MSSSIM) produces superior image reconstructions and achieves higher SSIM scores.
Figure 1: Sample reconstructions for the different loss functions with 4-fold acceleration. Each row represents a different image contrast.
Table 1: 4-fold acceleration results averaged across the 16-coil only FastMRI validation dataset. For PSNR and SSIM, the higher the score the better, and vice-versa for NMSE.