MC-RED: A deep learning network for motion correction in 3D CEST imaging

Chemical exchange saturation transfer (CEST) imaging is highly sensitive to patient motion, which can compromise the reliability of quantitative molecular analysis. This study aims to develop and validate a deep learning-based motion correction method for enhancing image quality in 3D CEST imaging.

We introduce MC-RED, a motion correction approach using a residual encoding-decoding network that incorporates frequency-specific information via a 2D Gaussian distribution with static reference images. This network generates motion-free reference frames for registering and correcting motion-corrupted CEST images. The method was evaluated on simulated data from healthy volunteers and clinical data from encephalitis patients, with performance measured by peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and mean squared error (MSE) from Lorentzian difference (LD) analysis.

MC-RED significantly enhances image quality by reducing motion artifacts, particularly in low-contrast regions near the water resonance. The corrected images show improved signal fidelity and spatial alignment, leading to more accurate quantitative maps.

The deep learning-based MC-RED effectively corrects motion artifacts in 3D CEST imaging, thereby holding potentials for increasing the reliability of quantitative CEST analysis.