Deep learning-based residual-subtractive denoising with latent mapping of noise: Application to Z-spectra

Motivation: Noise is an inherent limitation for all digital imaging and CEST MRI is no exception. With the already small effect-size some signals in the Z-spectra might be completely overshadowed. Available denoisers are limited in the aspect of lost-signal recovery. Goal(s): Proof-of-concept for a deep learning based denoising approach targeting the latent representation of noise. Approach: A DL-approach for latent mapping of noise and subsequent residual-subtractive removal was developed to map the latent noise and a subsequent residual-subtractive approach was applied. Results: The developed approach showed great promise with elevated performance compared to state-of-the-art in the field in terms of several quantitative measures. Impact: Denoising is crucial for usability of Z-spectra from CEST-images. Our developed DL-based solution is designed to map noise in latent space and subsequently remove it in a residual fashion allowing for an increased potential to recover overshadowed signal in Z-spectra.