Joint time-frequency WaveNet architecture for signal processing: application to vibration artifact suppression in high resolution NMR relaxometry experiments

Fourier-domain methods are central to both signal processing in magnetic resonance and image processing in imaging techniques. Modern analysis tools, including deeplearning methods, rarely exploit both the time-domain and the frequency-domain representations simultaneously. Although the time-and frequency domains are generally 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 artefacts in spectra obtained by high-resolution relaxometry (HRR). In HRR, low-field relaxation decays are recorded on a high-field magnet by transferring the sample to a position in the stray field of the magnet and rapid transfers are needed to prevent polarization loss. However, these transfers lead to vibrations, which induce low-frequency modulations of the free-induction decay, thus generating anti-phase sidebands that compromise quantitative interpretation. To address this, we trained JTF-WaveNet, with physics-informed simulations, to suppress these artifacts. Critically, when characterising the interaction of small molecules with human serum albumin, applying this deep learning approach improves the precision and accuracy of site-specific low-field relaxation rates by ≈ 60%.