Deep learning approach to super-resolution correction of brain MRI motion artifacts for accurate hippocampal volumetry.

To evaluate whether enhanced deep super-resolution network (EDSR) can reduce motion artifacts in brain MRI and improve hippocampal volumetry. We enrolled 24 healthy adults. Each participant underwent one acquisition without intentional movement (OI) and two acquisitions with instructed head movements to generate motion-corrupted images (MI). EDSR was applied to MI to generate motion-corrected images (EI). Hippocampal volumes were automatically measured using the brain anatomical analysis using diffeomorphic deformation software, and absolute percentage errors were calculated relative to the VOI. Participants were classified into low-motion (MI absolute percentage error (APE) < 10% for both hippocampi) and high-motion (MI APE ≥ 10% for either hippocampus) groups. Bland–Altman analysis was performed using absolute differences. Image quality was evaluated using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and mean absolute error (MAE) computed against registered OI in the MI space. Median APE did not differ significantly between MI and EI overall. In the high-motion sneeze subgroup, left-hippocampal APE decreased after EDSR, whereas other subgroup comparisons were not significant. EI showed higher PSNR and SSIM and lower MAE than MI (all P < 0.001). EDSR improves image quality metrics and can reduce hippocampal volumetry error in selected high-motion cases.