Abstract
PURPOSE: Low-power (LP) chemical exchange saturation transfer (CEST) Z-spectra have significantly reduced confounding effects and enhanced peak resolvability, thereby improving the observation and quantification of various CEST effects. However, LP Z-spectra suffer greatly from reduced contrast-to-noise ratio (CNR). This study aims to develop a dual-power feature preparation for an autoencoder-based deep learning approach (DPDL), for denoising LP Z-spectra. This method leverages the high CNR of higher saturation power and the enhanced peak resolvability of low saturation power. METHODS: The DPDL model was trained on simulated CEST data, validated on both simulated and BSA phantoms, and then applied to measured data from rat brain and leg muscles at 4.7T. A Lorentzian difference (LD) analysis was used to quantify various CEST effects. Several evaluation metrics, including peak signal-to-noise ratio (PSNR), were used to assess the denoising performance. To demonstrate the advantage of the DPDL method, it was compared with the autoencoder-based method without feature preparation and various existing denoising methods that utilized a single LP Z-spectrum with two averages as input, thereby ensuring equivalent acquisition times. RESULTS: In phantom experiments, the DPDL method demonstrated higher PSNR than existing denoising techniques, validating our approach. In animal experiments, the DPDL method showed improved image quality, outperforming existing denoising techniques. Additionally, several peaks from major tissue components in both brain and muscle were revealed on the LP Z-spectrum. CONCLUSION: The superior denoising performance using the DPDL for LP CEST imaging can enhance the isolation of various pools, thereby improving CEST applications, particularly at low fields.