AI Learns From Three Teacher Networks to Spot Confusable Chinese Medicine Slices

A team of researchers in China has built a lightweight artificial intelligence model that can tell genuine Ligustici Rhizoma et Radix decoction pieces apart from their most confusable adulterants with 99.73 percent accuracy, using nothing more than a photograph taken with a smartphone. The study, published in Frontiers in Pharmacology, shows how a technique called knowledge distillation can compress the expertise of three large neural networks into a model small enough to run on a phone, potentially bringing laboratory-grade authentication to pharmacies, medicinal-material markets and procurement inspections.

Ligustici Rhizoma et Radix, known in Chinese medicine as Gaoben, is the dried rhizome and root of Ligusticum sinense or Ligusticum jeholense, both members of the carrot family Apiaceae. First documented in Shennong’s Classic of Materia Medica, it is traditionally used to dispel wind, disperse cold, eliminate dampness and relieve pain, and is regarded as an important medicine for vertex headache. Rising market demand has depleted wild resources, and cultivated material now dominates supply. Yet current quality standards do not adequately reflect the macroscopic characteristics of cultivated materials, and brief descriptions in ancient materia medica literature have left a long-standing confusion over identification, with multiple botanical origins and closely related species trading under similar names.

The identification problem is compounded by unresolved taxonomy. The genus Ligusticum comprises roughly 60 species worldwide, about 30 of them in China, and molecular phylogenetic evidence has shown the genus to be rampantly polyphyletic, with Chinese species scattered across multiple distantly related clades. Two groups traded as Xinjiang Gaoben and Huang Gaoben have been transferred to the genus Conioselinum but remain morphologically similar to Ligusticum pieces. Meanwhile Ligusticum chuanxiong, known exclusively from cultivation, is included under Ligusticum in the 2025 edition of the Chinese Pharmacopoeia, and its decoction pieces are macroscopically difficult to distinguish from those of L. sinense.

To sidestep the nomenclatural disputes, the researchers adopted an operational framework of six categories: L. sinense, L. jeholense, L. chuanxiong, Conioselinum vaginatum, Conioselinum tataricum and L. sinense cv. Jinxiong. Every batch was verified by dual authentication, combining expert morphological identification with ITS2 DNA barcoding. Over five years the team collected 120 batches from six provinces across China, then photographed the pieces in an enclosed light box under constant illumination using a smartphone equipped with a 50-megapixel Light Hunter 800 sensor, a 320 by 320 millimetre textured light tent and a fixed-color-temperature LED fill light. The result was a dataset of 7,363 images at 12.58 megapixels, split 60:20:20 into training, validation and test sets.

The images were standardized using ImageNet channel means and standard deviations, and the training and validation sets were expanded through data augmentation including mirror flipping, horizontal scaling, random rotation, Gaussian blurring and random brightness adjustment, bringing the augmented dataset to 36,818 images. The test set consisted exclusively of original, unaugmented images to ensure an objective evaluation. Because the smallest class, Jinxiong, contained only 572 original images, the team also applied weighted random sampling so that minority classes contributed more evenly to gradient updates during training.

The core innovation lies in the training strategy. The researchers selected three high-accuracy teacher models, ResNet50, DenseNet169 and VGG16, and distilled their knowledge into MobileNetV2, a lightweight student model with only 2.23 million parameters. Knowledge distillation, first proposed by Geoffrey Hinton and colleagues in 2015, transfers what a large, complex network has learned to a smaller one through soft-label supervision. In this multi-teacher setup, the output probability distributions of the three teachers were softened with a temperature parameter, fused with a weighted strategy, and compared against the student’s predictions using Kullback-Leibler divergence, while a conventional cross-entropy loss on the true labels kept the student anchored to ground truth.

The results were striking. The MobileNetV2 baseline, trained for 50 epochs on the dataset without distillation, achieved only 89.33 percent accuracy, far below the teachers, which reached 99.66 percent for DenseNet169, 99.18 percent for VGG16 and 98.78 percent for ResNet50. After three-teacher distillation, the student reached 99.73 percent accuracy with precision, recall and F1-score all at or above 99.70 percent, exceeding every teacher. It retained the student’s compact footprint of 2.23 million parameters and 8.75 megabytes of storage, required just 0.33 GFLOPs per inference, and ran at 294 frames per second. Compared with DenseNet169, the distilled model maintained comparable accuracy while shrinking model size to 6.1 percent and computational cost to 9.6 percent.

Ablation experiments revealed a subtle pattern. Single-teacher distillation improved all metrics, with ResNet50 the most effective individual teacher at 99.32 percent accuracy, suggesting its deep residual features were most compatible with the lightweight student. Surprisingly, dual-teacher combinations performed worse than ResNet50 alone, which the authors attribute to representational conflicts between heterogeneous feature distributions. Only the full three-teacher configuration surpassed all other setups, indicating that three heterogeneous backbones can integrate complementary knowledge and mitigate the conflicts seen in pairs. The confusion matrix showed just four misclassifications, clustered in two taxonomically meaningful pairs: reciprocal errors between C. vaginatum and L. jeholense, likely due to processing-mediated morphological convergence after slicing and drying, and bidirectional confusion between L. sinense and its cultivar Jinxiong, whose block-shaped rhizomes differ only in subtle cross-sectional coloration. L. chuanxiong and C. tataricum were classified without a single error.

Validation went beyond a single test split. Five-fold cross-validation showed the distilled model holding steady at 99.40 percent plus or minus 0.47 percent accuracy across folds, compared with 86.47 percent plus or minus 3.15 percent for the original MobileNetV2, demonstrating that distillation also improved robustness to data partitioning. Gradient-weighted class activation mapping, or Grad-CAM, confirmed that the model’s high-response regions overlapped with those of the teacher models and concentrated on discriminative features of the pieces themselves, such as texture and morphological edges, rather than on background regions. The team also packaged the model into an interactive identification application built with PyQt5, integrating model loading, image preprocessing, inference and result analysis behind a graphical interface.

The authors acknowledge limitations. The study covered only Ligustici Rhizoma et Radix and its adulterants, not other easily confused medicinal materials, and the standardized image acquisition conditions did not fully represent the complex environments of actual market circulation, so real-world robustness requires further validation. Still, the framework offers a potentially transferable approach for traditional Chinese medicine quality control. Because the distilled model inherits the accuracy of heavyweight networks while keeping the small size and fast inference of a lightweight architecture, it is suitable for smartphones, tablets and embedded terminals, pointing toward intelligent, on-site authentication in retail pharmacies, medicinal-material markets and sampling inspections where laboratory equipment cannot follow.

Subject of Research: Knowledge distillation for rapid image-based identification of Ligustici Rhizoma et Radix decoction piece varieties and adulterants

Article Title: Knowledge distillation-based rapid identification of Ligustici Rhizoma et Radix decoction piece varieties

Article References: Hu, H., Jiang, S., Liang, S., Zhong, W., Luo, S., Liu, S., Hao, R., & Li, G. (2026). Knowledge distillation-based rapid identification of Ligustici Rhizoma et Radix decoction piece varieties. Frontiers in Pharmacology, 17, Article 1923842. https://doi.org/10.3389/fphar.2026.1923842

Image Credits: AI Generated

DOI: 10.3389/fphar.2026.1923842

Keywords: knowledge distillation, deep learning, MobileNetV2, traditional Chinese medicine, Ligustici Rhizoma et Radix, decoction pieces, adulterant identification, convolutional neural network, ITS2 DNA barcoding, Grad-CAM, herbal medicine authentication, lightweight AI model