STELT: A spatiotemporal deep learning framework for diffusion-ordered NMR spectroscopy reconstruction with artifact suppression.

Abstract

Diffusion Ordered NMR Spectroscopy (DOSY) is a powerful technique for studying mixtures by probing the diffusion behavior of mixed compounds and enabling their identification and separation in mixture samples. The performance of DOSY generally relies on the adopted reconstruction algorithm to determine diffusion coefficients from diffusion-dependent signal decays, thus producing a 2D spectrum that resolves components by chemical shift and molecular diffusion. Although deep-learning provides an effective approach to DOSY reconstruction, existing deep-learning-based reconstruction methods generally face the limitation of inadequate feature extraction, which may lead to reconstruction artifacts in some scenarios. Here, we propose STELT (Spatiotemporal Extraction Laplace Transform), a lightweight deep-learning framework based on spatiotemporal feature extraction. STELT uses a dual-branch architecture that combines a temporal convolution module to capture dynamic patterns in decay signals with a self-attention module to extract spatial features along the chemical-shift dimension. Experimental results demonstrate the proposed method achieves superior reconstruction accuracy and noise suppression with significantly reduced computational overhead, thereby offering a practical and efficient solution for high-quality DOSY analysis.