An identification method of herbal medicines superior to traditional spectroscopy: Two-dimensional correlation spectral images combined with deep learning

Traditional spectral analysis is prone to integration and signals overlap. It makes the extracted information unavailable or inaccurate, limiting its application in the detection of complex components. Therefore, two-dimensional correlation spectroscopy (2D-COS) that was superior to traditional spectral analysis has been established, by extracting useful information under certain chemical or physical stimulus from a series of spectra to improve the spectral resolution. In this study, with the identification of the different origins and parts of Panax notoginseng as an example. First, as a comparison, the attenuated total reflection infrared spectroscopy (ATR-FTIR) of P. notoginseng samples combined with chemometric methods was applied to identify it. Then, we generated synchronous and asynchronous 2D-COS spectral images from the fingerprint regions in the spectrum, and established a Residual convolutional neural network (ResNet) to identify the different origins and parts. Finally, the externally verified was applied to evaluate the accuracy of the model. By comparison, the 2D-COS was more suitable for the identification of the origins and parts of P. notoginseng than traditional spectros-copy. The identification accuracy of these samples of synchronous 2D-COS spectral images in the training set and test set were both greater than 99%. And the results of the externally verified indicated that synchronous 2D-COS spectral images were significantly outperforming asynchronous 2D-COS spectral images, and the sensitivity, specificity, and accuracy of the samples are both better than 0.99. Furthermore, when the sample number was relatively small or large differences in sample number were large, the synchronous 2D-COS spectral images could also successfully distinguish the different origins and parts well of P. notoginseng. In general, the ResNet model could provide a better discriminant model, and perform high-resolution processing.