Page 303 - 2025中醫藥與天然藥物聯合學術研討會-中醫藥與天然藥物的挑戰Ｘ機遇與未來大會手冊

CM-36

Deep learning-based image recognition of easily confused Traditional

Chinese Medicinal seeds: Identifying Chinese leek and spring onion

,1
1
Jin‑Xuan He, Meng‑Shiou Lee,* Chao‑Lin Kuo* ,1

1 Pharmacognosy Laboratory, Department of Chinese Pharmaceutical Sciences and Chinese
Medicine Resources, China Medical University, Taichung, Taiwan
* E-mail: leemengshiou@mail.cmu.edu.tw; clkuo@mail.cmu.edu.tw

Abstract
Accurate authentication of traditional Chinese medicinal (TCM) materials is essential for
quality control and patient safety. Seeds of Allium tuberosum (Chinese leek) and Allium
fistulosum (spring onion) are notoriously similar in appearance, leading to frequent
misidentification in markets. We established a curated image dataset consisting of 200 images
each for A. tuberosum and A. fistulosum and an independent validation set (20 images per class).
Images were preprocessed and augmented to enhance robustness. Three convolutional neural
networks—DenseNet201, InceptionV3, and VGG19—were trained and optimized. To support
practical deployment, we implemented a confidence‑thresholding scheme to abstain on
ambiguous cases and used explainable AI visualizations to verify that models relied on
botanically meaningful features. The best model was further evaluated for two downstream
tasks: (i) adulteration detection in mixed samples and (ii) rapid screening of commercial
samples. All three architectures achieved high classification performance on the held‑out
validation set, and confidence‑thresholding improved precision in field‑like settings by
rejecting uncertain predictions. Explainability analyses highlighted morphological cues
consistent with pharmacognostic assessment. In downstream tests, the pipeline correctly
flagged intentional adulteration and supported rapid triage of market samples for confirmatory
inspection. This study demonstrates a reproducible, explainable, and practically oriented
deep‑learning workflow for differentiating visually similar TCM seeds. The approach can be
extended to other easily confused crude drugs and may facilitate point‑of‑care quality control
in supply chains and production sites.

Keywords: Allium tuberosum; Allium fistulosum; Deep learning; Image recognition;
Pharmacognosy; Quality control