Fourier Transformer for Joint Super-Resolution and Reconstruction of MR Image

Super resolution (SR) and fast magnetic resonance (MR) imaging play a paramount role in medical diagnosis community. Though significant progress has been achieved, existing methods are mostly trapped at the local feature dependencies. Transformer holds the potential to tackle this issue, yet it is notoriously known with heavy computational burden. As a remedy, we propose a Fourier Transformer Network (FTN), which leverages 2D Fast Fourier Transform (FFT) to harvest the global relationships. Specifically, multiple Fourier Transformer Blocks (FTBs) are stacked as the backbone, comprehensively extracting the sematic representations. Due to the appealing efficiency of FFT, the overall model complexity is linear to tokens, as opposed to the quadratic complexity of previous transformer-based models. Besides, an edge-enhancement branch is incorporated into FTB in a flexible manner, which further sharpens the contour information of the organs and tissues. Extensive experiments on IXI dataset validate the superiority of FTN over most state-of-the-art methods.