Beyond traditional Magnetic Resonance processing with Artificial Intelligence

TL;DR

The paper demonstrates that AI-guided Magnetic Resonance processing (MR-Ai) can extend NMR data processing beyond traditional methods by reconstructing spectra from incomplete phase-modulated quadrature data and providing per-point uncertainty estimates. It introduces a 2D WaveNet-based architecture with sequential correction steps to recover Echo/Anti-Echo spectra via data completion in the VE time domain, and it couples this with a Gaussian-layer uncertainty model trained on synthetic data to create a reference-free quality score (pSQ). The results show MR-Ai achieving higher-quality reconstructions than compressed sensing approaches, capturing twisted phase-line shapes as absorptive spectra, and delivering robust, per-point uncertainty measures that align with traditional metrics. Together, these advances enable faster acquisitions, quantitative uncertainty assessment, and objective spectrum quality evaluation, with potential to reshape NMR signal processing and analysis.

Abstract

Smart signal processing approaches using Artificial Intelligence are gaining momentum in NMR applications. In this study, we demonstrate that AI offers new opportunities beyond tasks addressed by traditional techniques. We developed and trained several artificial neural networks in our new toolbox Magnetic Resonance with Artificial intelligence (MR-Ai) to solve three "impossible" problems: quadrature detection using only Echo (or Anti-Echo) modulation from the traditional Echo/Anti-Echo scheme; accessing uncertainty of signal intensity at each point in a spectrum processed by any given method; and defining a reference-free score for quantitative access of NMR spectrum quality. Our findings highlight the potential of AI techniques to revolutionize NMR processing and analysis.

Figures (16)

• Figure 1: Illustration of the Echo and normal spectrum. (a) Echo spectrum with a phase-twist lineshape, (b) a normal spectrum with a pure, absorptive phase in the frequency domain, and (c and d) their corresponding Virtual Echo presentations in the time domain respectively. In the figures, $P$ and $N$ represent the P-type and N-type data sets, while $\widetilde{P}$ and $\widetilde{N}$ indicate the time reverse and conjugation of P-type and N-type data sets respectively -- procedures for transition between the presentations are indicated by arrows.
• Figure 2: Performance of Echo and Anti-Echo reconstruction by using MR-Ai and CS on real data. (a) 2D ^1H-^15N — TROSY spectra of MALT1han2022assignment Echo reconstruction using MR-Ai with predicted uncertainty in pink color - The insets show zooming part and corresponding reference with the actual absolute error between reference and reconstruction in red color. Bar graphs (b) represent $RMSD$ as a traditional reference-based evaluation metric and Boxplots (c) represent normalized uncertainty as the intelligent reference-free evaluation metric for comparison reconstructed spectra using MR-Ai and CS for Malt as described in the Appendix \ref{['sec:Materials and methods']} section for additional details.
• Figure 3: Illustration of training and predicting MR-Ai for estimation of the uncertainty of the reconstruction generated by any method
• Figure A.1: Scheme of Magnetic Resonance processing with Artificial Intelligence (MR-Ai) for Echo (or Anti-Echo) reconstruction. (a) MR-Ai network architecture with five 2D WNNs and five correction steps (b) Scheme of the WNN module used in MR-Ai (c) Training- and cross-validation losses for the WNNs during training
• Figure A.2: Training- and cross-validation losses for the WNNs during training MR-Ai models for uncertainty estimation