Improving the magnetic resonance images super-resolution with a dual-channel enhancement model incorporating complementary information

Abstract

Although significant progress has been made in image super-resolution using artificial intelligence, achieving high-quality super-resolution for magnetic resonance (MR) images remains challenging due to their unique imaging principles and processes. Inspired by MR parallel imaging, we propose a novel concept to improve the MR image super-resolution quality by leveraging the complementary spatial information inherently contained in multi-channel receive coils. A new MR image degradation model was developed to generate the training dataset that complies with the parallel imaging and Sensitivity Encoding (SENSE) reconstruction. A dual-channel enhancement model, named sensitivity encoding based super-resolution (SenseSR), is then devised with the main channel processing the single low-resolution image and the enhancement channel processing the multiple images from each coil channel. SenseSR is mainly featured with cascaded double enhancement blocks that can extract deeper features of the multiple coil-channel images and fuse them together into the main channel. Experiments were performed to test the performance and compare it with other benchmark models. The results demonstrate a significant improvement in MR image super-resolution quality, with an enhancement of peak signal-to-noise ratio ranging from 0.5 to 6.5 decibels (dB). Further experiments with different testing datasets and MR images collected in-situ demonstrated that SenseSR also has good generalization capability and robustness, indicating its potential for clinical applications. The code is available at https://github.com/MISR-Lab/SenseSR.