UNMR-Net: A Unified Deep Reconstruction Network for Multitask Compressed Sensing NMR Spectroscopy

NMR spectroscopy has become an indispensable tool in chemistry and life sciences, yet its high resolution is generally constrained by the cost of increased acquisition times. Recently, sparse data sampling combined by deep learning (DL) reconstruction serves as an effective manner to highly speed up NMR acquisition while maintaining satisfactory spectral quality. However, existing end-to-end DL models are generally designed specifically for a single spectroscopy type and operate as "black boxes", hindering their generalizable applications and trustworthy interpretation in practical deployments. Here, we introduce a unified deep reconstruction network architecture for multitask compressed sensing NMR spectroscopy by unrolling the classical iterative shrinkage-thresholding algorithm. First, this framework fully exploits the theoretical foundations and multitask adaptability of a compressed sensing solver to break the task-customized network design and enable a single network to be simultaneously applied to two NMR sparse reconstruction tasks, in which a common network architecture is trained separately for different applications. Second, benefiting from leveraging both model-driven reliability and data-driven expressivity, the proposed model provides reconstruction performance, outperforming specialized state-of-the-art models tailored to individual tasks, and implements improved transparency of the reconstruction process via stage-wise visualization. Therefore, our findings somewhat respond to critical concerns of adaptability and interpretability inherent in DL models and support more reliable DL-assisted spectroscopic interpretation, thus boosting the potential of artificial intelligence to extend the scope of NMR methodologies.