Unsupervised Transformer Learning for Rapid and High-Quality MRI Data Acquisition

Background: Magnetic resonance imaging (MRI) is of considerable importance due to its wide range of applications in both scientific research and clinical diagnostics. Acquiring high-quality MRI data is of paramount importance. Super-resolution reconstruction serves as a post-acquisition method capable of improving MRI data quality. Current methods predominantly utilize convolutional neural networks in super-resolution reconstruction. However, convolutional layers have inherent limitations in capturing extensive spatial dependencies due to their localized nature. Methods: We developed a new methodology that enables rapid and high-quality MRI data acquisition through a novel super-resolution approach. We proposed an innovative architecture using transformers to exploit long-range spatial dependencies present in images, allowing for an unsupervised learning framework specifically designed for super-resolution tasks tailored to individual subject. We validated our approach using both simulated data and clinical data comprising 40 scans acquired with a 3-T MRI system. Results: We obtained images with T2 contrast at an isotropic spatial resolution of 500 μm in just 4 min of imaging time, and simultaneously, the signal-to-noise ratio and contrast-to-noise ratio were improved by 13.23% and 18.45%, respectively, in comparison to current leading super-resolution techniques. Conclusions: The results demonstrated that incorporating long-range spatial dependencies substantially improved super-resolution reconstruction, thereby allowing for the acquisition of high-quality MRI data with reduced imaging time.