MGAM-FusionNet: A masked reconstruction and multi-scale attention fusion network for small-sample near-infrared spectral quantification of echinacoside in Cistanche deserticola.

Echinacoside is a major bioactive constituent and an important quality marker of Cistanche deserticola Y. C. Ma. Conventional high-performance liquid chromatography (HPLC) provides accurate reference measurements but is time-consuming and destructive, limiting its use for rapid quality assessment. Near-infrared (NIR) spectroscopy offers a rapid, non-destructive alternative; however, severe band overlap and matrix-related variation can compromise the generalization of chemometric and deep learning models when only small datasets are available. To address these challenges, we developed MGAM-FusionNet, a quantitative regression framework that integrates masked self-supervised reconstruction with multi-scale attention. The dataset comprised NIR spectra and HPLC-determined echinacoside concentrations for 130 Cistanche samples representing multiple geographical origins and species. Twenty samples were randomly withheld as an external independent test set, and the remaining samples were partitioned by the SPXY algorithm into a training set and an internal validation set. MGAM-FusionNet achieved an R2 of 0.9197 and an RMSE of 0.3291 on the internal validation set. On the external independent test set, it achieved an R2 of 0.8800, an RMSE of 0.4352, and an RPD of 2.8872. These results represented the best overall test-set performance among the evaluated models, although 1D-ViT yielded a slightly lower MAE. Ablation experiments supported the contributions of the principal model components. Overall, MGAM-FusionNet provides a promising approach for rapid, non-destructive echinacoside quantification in Cistanche samples.