Unsupervised Multi-Contrast MRI Super-Resolution with the Implicit Feature Sampling and Reciprocal Framework

Motivation: Multi-contrast MRI super-resolution (MCSR) can effectively shorten the MRI acquisition time. Yet, existing deep learning-based approaches requiring paired low-resolution (LR) and high-resolution (HR) images for training are impractical in clinical settings. Goal(s): We aim to propose an unsupervised model that can achieve MCSR without the need for ground-truth HR images. Approach: We construct a network to generate HR images from its LR counterparts and reference images. Meanwhile, a cycle-consistency network and a reciprocal network are proposed to constrain the outputs. Results: Experiments on two datasets demonstrate that our proposed model effectively restores HR images with clear anatomic details. Impact: Our model facilitates multi-contrast MRI super-resolution in the absence of ground-truth HR images, which not only substantially reduces MRI acquisition time, but enables the obtaining of certain HR sequences that are difficult to acquire in clinical settings.