Visualization and interpretation of origin identification of Panax notoginseng samples using neural networks

作者：Shuwen Wang, Hong Gu, Yan Peng · 年份：2024 · DOI：10.1109/icops58192.2024.10627637 · 研究领域：Spectroscopy and Chemometric Analyses

The combination of deep learning with terahertz spectroscopy technology has been widely used in the analysis and identification of traditional Chinese medicinal materials. However, the decision-making process inside neural networks is often seen as a black box, making it difficult to understand and explain. This ’black box’ problem limits its application in some traditional Chinese medicinal material analysis and identification areas. Here, we propose an interpretable neural network algorithm for identifying the origin of Panax notoginseng samples and visualizing its process. We first use Convolutional Neural Networks (CNN) to classify the origin of the terahertz spectral data of Panax notoginseng samples, with an accuracy of up to $95 \%$. Furthermore, the model uses t-distributed stochastic neighborhood embedding (t-SNE) to visualize high-dimensional spectral data and analyzes the neural network to investigate how it processes input spectral data to achieve origin classification. Finally, applying explainable artificial intelligence (XAI) technology to explain the decision-making process of CNN, we found that the frequency, amplitude, relative proportion, peak area, and full width at half maximum (FWHM) of terahertz absorption peaks are all related to their respective characteristics of origin. These features play an important role in the origin recognition of neural networks, explaining the basis of neural networks in identifying the source of Panax notoginseng samples. Ou...