Enhancing SNR of CEST MRI Through Noise‐to‐Noise Deep Learning With K‐Space Data Consistency

Tha aim of this study is to develop a deep learning-based retrospective denoising method, via noise-to-noise (N2N) CEST, to enhance the SNR of acquired CEST images. The N2NCEST model was trained using pairs of noisy CEST images through a N2N deep learning approach, distinct from conventional approaches that use clean or simulated CEST images. A transformer module was used to exploit spatiotemporal correlations embedded in CEST images among different frequency offsets. A data consistency layer was introduced to preserve the center k-space of input CEST images that were acquired through centric-encoding order. The performance of this model was evaluated using in vivo data from wild-type mouse brains. Multipool Lorentzian fitting (MPLF) was applied to obtain creatine, amide, and rNOE CEST maps to demonstrate the applicability of N2NCEST. Reference CEST images were acquired by setting signal average as four during scanning, which quadrupled the total scan time. The denoising results were evaluated by peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). Pixel-wise correlation coefficient was also calculated for fitted CEST maps. N2NCEST surpassed existing CEST denoising methods and exhibited better image fidelity.N2NCEST was able to restore high-SNR images from their noisy acquisitions without sacrificing high-frequency structural information. The unique N2N scheme can avoid complicated parameter tuning and data simulation, achieving good compatibility and availability.