3D Super Resolution for Non-Isotropic Medical Image Through Multi-Input 3D ResUnet

Fluid-attenuated inversion recovery imaging (FLAIR) is a magnetic resonance (MR) method that is frequently utilized to diagnose brain lesions. However, it usually provides transverse section imaging in high resolution but with very low resolution in the other axis. This non-isotropy huddles its wide utilization in research including machine learning. In this study, we applied a deep-learning based super resolution (SR) technique to convert non-isotropic FLAIR images into isotropic one. We proposed a multi-input 3D ResUnet that uses both FLAIR and T1-weighted MR images as its inputs. As a result, our proposed model successfully reconstructed isotropic FLAIR images (SSIM: 0.9947
$\pm 0.0009)$
, and out-performed the FLAIR only single-input 3D ResUnet.