Blind Adaptive Local Denoising for CEST Imaging

TL;DR

This work addresses noise and heteroscedasticity in chemical exchange saturation transfer (CEST) MRI by introducing BALD, a two-stage denoising framework that combines Adaptive Variance Stabilization (AVST) with local SVD-based denoising. AVST estimates a data-specific noise curve $g(u)$ along saturation frequency and defines $f(u)=\sigma \int_{0}^{u} 1/g(t)\,dt$ to transform the data into a variance-stabilized domain, enabling effective patch-wise SVD denoising that uses hard thresholding and Wiener refinement. The BALD pipeline maps back to the original intensity via the inverse transform and demonstrates superior performance in both phantom and in vivo studies, improving PSNR and preserving molecular contrast maps such as APTw and NOE, while enhancing downstream analyses like MPF fitting and tumor detection. A key contribution is showing that AVST not only improves BALD but also enables pretrained natural-image denoisers to generalize to CEST data, potentially accelerating clinical adoption. The results suggest BALD as a robust, adaptable denoising framework that improves quantitative CEST imaging across varied acquisition protocols and could facilitate more reliable disease characterization.

Abstract

Chemical Exchange Saturation Transfer (CEST) MRI enables molecular-level visualization of low-concentration metabolites by leveraging proton exchange dynamics. However, its clinical translation is hindered by inherent challenges: spatially varying noise arising from hardware limitations, and complex imaging protocols introduce heteroscedasticity in CEST data, perturbing the accuracy of quantitative contrast mapping such as amide proton transfer (APT) imaging. Traditional denoising methods are not designed for this complex noise and often alter the underlying information that is critical for biomedical analysis. To overcome these limitations, we propose a new Blind Adaptive Local Denoising (BALD) method. BALD exploits the self-similar nature of CEST data to derive an adaptive variance-stabilizing transform that equalizes the noise distributions across CEST pixels without prior knowledge of noise characteristics. Then, BALD performs two-stage denoising on a linear transformation of data to disentangle molecular signals from noise. A local SVD decomposition is used as a linear transform to prevent spatial and spectral denoising artifacts. We conducted extensive validation experiments on multiple phantoms and \textit{in vivo} CEST scans. In these experiments, BALD consistently outperformed state-of-the-art CEST denoisers in both denoising metrics and downstream tasks such as molecular concentration maps estimation and cancer detection.

Figures (7)

• Figure 1: Schematic illustration of the Blind Adaptive Local Denoising (BALD) Framework.
• Figure 2: Evaluation on synthetic phantoms corrupted by multi-level Rician noise. Quantitative comparisons are based on the (a) CEST sequence and (b) extracted contrasts by multi-pool Lorentzian fitting, respectively. The visualization of CEST contrasts (d) is based on the synthetic phantom under the noise level of 0.05.
• Figure 3: Visualization of dopamine maps generated from BALD processed scan.
• Figure 4: APTw images acquired under $B_1=2\mu T$ and obtained after denoising, with tumor region labeled on the T2-weighted image on the left and the $\textit{M}_0$ image for reference.
• Figure 5: CEST Contrasts intensity distributions in the regions of interest (ROIs), which are labeled on the $\textit{M}_0$ images on the left.