RaGS-SR for High-Fidelity Super-Resolution Reconstruction of Raman Spectral Images.

Raman imaging provides label-free and molecule-specific spatial profiling of biological samples and has been widely applied in cellular analysis and pathological research. However, the intrinsically weak Raman scattering signal imposes a fundamental trade-off between spatial resolution and imaging speed, limiting its use in high-throughput applications. Computational reconstruction, particularly super-resolution, offers a promising strategy for improving Raman imaging quality from sparsely sampled data. Existing reconstruction methods often struggle to preserve intrinsic spatial-spectral coupling under limited computational resources, which may compromise reconstruction accuracy and the reliability of subsequent analysis. Here, we propose Raman groupwise spectral super-resolution (RaGS-SR), a three-dimensional spatial-spectral super-resolution network with groupwise spectral modeling for Raman spectral images. RaGS-SR employs a spectral encoder to compress high-dimensional Raman data into compact latent representations while preserving essential spectral signatures and reducing computational overhead. The subsequent 3D reconstruction network jointly models spatial structures and interband spectral correlations to achieve accurate and faithful reconstruction. Experimental results demonstrate that RaGS-SR achieves 4-fold spatial super-resolution, corresponding to a 16-fold reduction in acquisition time under sparse sampling. Visual evaluation and vertex component analysis (VCA)-based unsupervised spectral unmixing further confirm its superior performance in spatial detail restoration and spectral fidelity preservation. Cross-data set transfer learning on breast cancer Raman data also demonstrates its strong generalization capability. Overall, RaGS-SR provides an efficient and practical computational tool for high-throughput Raman imaging, with promising potential for molecular characterization and pathological diagnosis.