Yuling Deng,Meifang Jiang,Qing Liu,Jun Wang,Yi Tao
Ginkgo ester (GBE50) is a well-known botanical drug derived from crude material Ginkgo Folium . Nonetheless, a swift and nondestructive method for full quality inspection on Ginkgo Folium is still lacking. Herein, an interpretable deep learning embedded hyperspectral imaging methodology was established to control the quality of Ginkgo Folium . Hyperspectral images were collected from 102 batches of Ginkgo Folium ranging from 389.8 nm to 1048.2 nm. The contents of three flavonol glycosides and four terpene lactones in Ginkgo Folium were unambiguously determined by well-validated high-performance liquid chromatography (HPLC) and ultra-high performance liquid chromatography-triple quadrupole-mass spectrometry (UHPLC-TQ-MS) methods, whilst the moisture content measured through a moisture meter. In order to realize the real-time full quality inspection, competitive adapative reweighted sampling-partial least squares regression (CARS-PLSR) and seven deep learning algorithms (ResNet50, AlexNet, GoogLeNet, DenseNet121, ResNeXt50, Inception-V4 and Inception-ResNet-V2) were attempted to establish a robust quantitative calibration model embedded hyperspectral imager. As a result, DenseNet121 algorithm showed the most excellent predictive performance. The R p 2 values of quercetin, kaempferol, isorhamnetin, bilobalide, ginkgolide A, ginkgolide B, ginkgolide C, and moisture were determined to be 0.9969, 0.9913, 0.9965, 0.9988, 0.9951, 0.9952, 0.9988 and 0.9981, respectively, and the RMSEP values were 0.0084, 0.0094, 0.0103, 0.0053, 0.0056, 0.0099, 0.0084 and 0.0050, respectively. Grad-CAM was used to unlock the blackbox of DenseNet121model and interpret the model. It’s believed that the established DenseNet121 embedded hyperspectral imager can rapidly and nondestructive determine the eight quality attributes and provide a tool for full quality inspection on Ginkgo Folium .
Xujin Yang,Meifang Jiang,Yuling Deng,Jun Wang,Yi Tao
Fluid-bed granulation (FBG) is a critical unit in the production of Ginkgo ester , a widely recognized therapeutic agent in cardiovascular medicine. However, persistent challenges such as moisture-induced bed collapse and heterogeneous binder distribution necessitate the need for advanced real-time monitoring solutions. This study presents a novel integration of hyperspectral imaging (HSI) with convolutional neural networks (CNNs) to establish a smart process analytical technology (PAT) platform for multi-objective quality assessment during Ginkgo ester granulation. A comprehensive dataset was obtained through analysis of 99 samples from ten batches of granules, where seven active pharmaceutical ingredients were quantitatively determined using ultra-performance liquid chromatography-tandem mass spectrometry (UPLC-TQ-MS). Critical quality attributes, including particle size distribution, content uniformity, specific surface area, moisture content, Hausner ratio, and angle of repose, were systematically characterized. Through comparative evaluation of seven CNN architectures and a Competitive Adaptive Reweighted Sampling-Partial Least Squares Regression (CARS-PLSR) model, DenseNet121 emerged as the optimal architecture, demonstrating superior predictive performance across 15 quality attributes ( R p 2 : 0.9557–0.9933; RMSEP: 0.0149–0.0284). The proposed HSI-CNN framework represents a significant advancement in real-time process monitoring capabilities, offering enhanced predictive accuracy and addressing critical limitations in current FBG operations. This research establishes a robust technological foundation for implementing advanced process analytical technologies in pharmaceutical manufacturing.
危险品化学品经营许可证（带存储）