Knowledge distillation-based rapid identification of Ligustici Rhizoma et Radix decoction piece varieties
- 1State Key Laboratory for Quality Ensurance and Sustainable Use of Dao-Di Herbs, Institute of Chinese Materia Medica, China Academy of Chinese Medical Sciences
- 2School of Pharmacy, Jiangxi University of Chinese Medicine
- 3Modern Research Center for Traditional Chinese Medicine, Shanxi University
Abstract
To address the commingling of different varieties of Ligusticum decoction pieces in the market and the difficulty of achieving rapid, efficient, and accurate on-site identification, this study aimed to develop a lightweight deep learning model that combines high identification accuracy with fast inference speed. Methods: Images of six categories of Ligusticum decoction pieces and their adulterants were acquired in an enclosed photographic light box to construct a standardized image dataset. Using ImageNet-pretrained weights, several high-accuracy models were selected as teacher models through comparative model evaluation, and knowledge distillation was subsequently employed to optimize the lightweight MobileNetV2 model as the student model. Finally, an interactive identification application was developed using PyQt5. Results: After knowledge distillation, the MobileNetV2 model achieved an identification accuracy of 99.73%, approaching the accuracy of the teacher models. Meanwhile, the distilled model contained only 2.23 million parameters, substantially fewer than the teacher models, thereby achieving both high identification accuracy and ease of deployment. Conclusion: The lightweight identification model developed in this study efficiently extracted high-level discriminative features from images of Ligusticum decoction pieces and their adulterants, enabling rapid and accurate identification and providing a feasible technical approach for the rapid on-site authentication of traditional Chinese medicine decoction pieces.
Related
- Knowledge distillation-based lightweight generative large language models for standardized quality control of Chinese radiology reports
- Ultrasound-based breast nodule malignancy diagnosis using CNN enhanced by knowledge distillation
- Prediction of lncRNA–cervical cancer associations using a multi-view collaborative variational autoencoder with knowledge distillation transfer learning
- Hybrid Attention Model Using Feature Decomposition and Knowledge Distillation for Blood Glucose Forecasting
- Critical risk node identification in enterprise resource planning heterogeneous knowledge graphs
BibTeX
@article{hu2026knowledge,
title = {{Knowledge distillation-based rapid identification of Ligustici Rhizoma et Radix decoction piece varieties}},
author = {Haoyu Hu and Shangzhou Jiang and Shengting Liang and Wentao Zhong and Shenglin Luo and Sijie Liu and Ruifen Hao and Geng Li},
journal = {Frontiers in Pharmacology},
year = {2026},
month = oct,
doi = {10.3389/fphar.2026.1923842},
url = {https://www.frontiersin.org/articles/10.3389/fphar.2026.1923842/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Knowledge distillation-based rapid identification of Ligustici Rhizoma et Radix decoction piece varieties
AU - Haoyu Hu
AU - Shangzhou Jiang
AU - Shengting Liang
AU - Wentao Zhong
AU - Shenglin Luo
AU - Sijie Liu
AU - Ruifen Hao
AU - Geng Li
PY - 2026
DA - 2026/10/22
JO - Frontiers in Pharmacology
DO - 10.3389/fphar.2026.1923842
UR - https://www.frontiersin.org/articles/10.3389/fphar.2026.1923842/full
AB - To address the commingling of different varieties of Ligusticum decoction pieces in the market and the difficulty of achieving rapid, efficient, and accurate on-site identification, this study aimed to develop a lightweight deep learning model that combines high identification accuracy with fast inference speed. Methods: Images of six categories of Ligusticum decoction pieces and their adulterants were acquired in an enclosed photographic light box to construct a standardized image dataset. Using ImageNet-pretrained weights, several high-accuracy models were selected as teacher models through comparative model evaluation, and knowledge distillation was subsequently employed to optimize the lightweight MobileNetV2 model as the student model. Finally, an interactive identification application was developed using PyQt5. Results: After knowledge distillation, the MobileNetV2 model achieved an identification accuracy of 99.73%, approaching the accuracy of the teacher models. Meanwhile, the distilled model contained only 2.23 million parameters, substantially fewer than the teacher models, thereby achieving both high identification accuracy and ease of deployment. Conclusion: The lightweight identification model developed in this study efficiently extracted high-level discriminative features from images of Ligusticum decoction pieces and their adulterants, enabling rapid and accurate identification and providing a feasible technical approach for the rapid on-site authentication of traditional Chinese medicine decoction pieces.
SN - 1663-9812
ER -