Enhancing SNR of CEST MRI Through Noise-to-Noise Deep Learning with K-space Data Consistency

Motivation: Endogenous CEST contrast is relatively small and vulnerable to imaging noise. Goal(s): To develop a retrospective denoising method to enhance SNR of acquired CEST images. Approach: A denoising neural network was trained using pairs of noisy CEST images through a noise-to-noise deep learning approach, distinct from conventional approaches that use clean or simulated CEST images. A data consistency layer was introduced to preserve center k-space of original CEST images to improve fidelity. A transformer module was used to exploit spatiotemporal correlations among different frequency offsets. Results: Multipool Lorentzian fitting was performed. Compared to clean images, our method achieved mean correlation coefficient of 0.90. Impact: Our method can considerably increase SNR of CEST images without sacrificing image fidelity. After denoising, the derived CEST maps could more reliably represent molecular changes in brain regions.