Abstract
<jats:p>Fourier-domain methods are central to both signal processing in magnetic resonance and image processing in imaging techniques. Modern analysis tools, including deep-learning methods, rarely exploit simultaneously the time-domain and the frequency-domain representations. Although the time and frequency domains are unitarily equivalent, a combined analysis of the two representations of the data allows for a highly improved convergence of the analysis. Here, we introduce the Joint Time–Frequency WaveNet (JTF-WaveNet) deep-learning architecture that processes NMR signal in both domains, producing the desired output along with point-wise uncertainty estimates. We use this architecture to suppress artifacts in spectra obtained by high-resolution relaxometry (HRR). In HRR, low-field relaxation decays are recorded on a high-field magnet to determine dynamics and interactions in complex molecular systems. The sample is transferred through the stray field of the magnet. Transfers need to be fast to prevent polarization loss. However, faster transfers lead to increased vibrations, which induce low-frequency modulations of the free-induction decay, thus generating anti-phase sidebands that compromise quantitative interpretation. We trained JTF-WaveNet with physics-informed simulations to suppress these artifacts while perserving intensities to quantify relaxation rates. We characterised the interactions of small molecules with human serum albumin, with this deep learning approach improving the precision and accuracy of site-specific low-field relaxation rates by about ~60 %.</jats:p>