Feasibility study on the introduction of Micro-CT technology for the identification of Radix Bupleuri and its adulterants

Kehong Chen, Chen Gong, Zhelong Zhuang, Shouhua Luo, Jiaming Liu, Guorong LiuFrontiers in Pharmacology 2024 · DOI

This study, published in Frontiers in Pharmacology, explores a high-tech solution to a centuries-old problem: ensuring the authenticity of herbal medicines. By combining Micro-CT imaging with ResNext deep learning architectures, researchers have developed a non-destructive method to distinguish the medicinal herb Radix Bupleuri from its common (and sometimes toxic) look-alikes with 75% accuracy.

TL;DR

Researchers successfully applied Micro-CT imaging and ResNext deep learning to identify Radix Bupleuri and its adulterants non-destructively. By analyzing internal microstructures like wood fiber distribution, the system provides a digital "fingerprint" for herbal authentication, achieving a 75% classification accuracy.

What Problem Does This Paper Address?

Radix Bupleuri (known as "Chaihu" in China) is a cornerstone of Traditional Chinese Medicine (TCM), used for everything from fever reduction to liver protection. However, the market is plagued by adulterants—herbs that look nearly identical to the naked eye but lack medicinal efficacy or, worse, contain toxic compounds like those found in Bupleurum longiradiatum.

Current identification methods have significant friction points:

  • Morphological inspection: Requires highly specialized botanical knowledge.
  • Chemical analysis (HPLC/TLC): Accurate, but destroys the sample and is time-consuming.
  • DNA Barcoding: Highly technical and expensive.

The authors sought a "third way": a method that is non-destructive, provides high-resolution internal data, and can be automated via AI.

Methods

The researchers' workflow bridged the gap between traditional botany and modern computer vision:

  1. Micro-CT Scanning: Unlike clinical CT, Micro-CT offers resolution at the 1–100 μm level. Samples were scanned at 45kv to capture nuanced internal structures without slicing the herb.
  2. 3D Reconstruction: They used the FDK (Feldkamp-Davis-Kress) algorithm, a standard for cone-beam CT, to transform 2D X-ray projections into 3D volumetric data.
  3. Data Correction: To account for detector defects and pixel response distortion, they applied geometric and light-field corrections to the DICOM (Digital Imaging and Communications in Medicine) files.
  4. Deep Learning (ResNext-50): The core of the classification was a ResNext-50 model. ResNext improves upon the standard ResNet by introducing "cardinality"—splitting the network into multiple parallel paths within a single block. This allows the model to learn features across different dimensions more efficiently without a massive increase in complexity.
  5. Data Augmentation: Because physical herb samples are limited, they used random rotation, masking (occlusion), and Gaussian blurring to expand their dataset and prevent overfitting.

Datasets

The study utilized a targeted, high-quality dataset:

  • Samples: 9 physical samples (3 Radix Bupleuri, 3 Moyuanchaihu, and 3 Zangchaihu).
  • Image Count: 300 original DICOM images (100 per type) extracted from the 3D reconstructions.
  • Augmented Set: Expanded to 1,200 images (1024x1024 pixels) for training and testing.
  • Split: A standard 8:2 ratio for training and testing sets.

Key Insights and Ideas

  • Internal Morphological Signatures: The "aha" moment lies in the xylem. While the herbs look identical externally, their internal wood fiber distributions are distinct:
  • Radix Bupleuri: Fibers arranged in interrupted rings.
  • Moyuanchaihu: Fibers clustered near the formation layer.
  • Zangchaihu: Fibers clustered in the middle, forming a medullary-like structure.
  • Non-Destructive "Digital Slicing": Micro-CT allows researchers to observe oil chambers and cork layers—features usually only visible via thin-section microscopy—while keeping the physical sample intact for future use.
  • Stability of ResNext: The study found that while standard ResNet could eventually reach similar accuracy, ResNext was significantly more stable and reached its peak accuracy much earlier (70 epochs vs. 120).

Key Results

  • Classification Accuracy: The ResNext-50 model achieved 75% accuracy in identifying the three species.
  • Comparative Performance:
    • ResNext-50: 75% (High stability).
  • ResNet-50: 75% (Lower stability, slower convergence).
  • RegNet: 67% (Lower overall performance).
  • Visual Clarity: The FDK reconstruction successfully visualized microscopic features like oil chamber fissures and cambium rings that are critical for botanical validation.

Limitations and Future Work

While a successful proof-of-concept, the authors acknowledge several hurdles:

  • Sample Size: The study relied on a small number of physical samples. Future work needs to include a wider variety of growth regions and ages to ensure the model generalizes well.
  • Accuracy Ceiling: 75% is a strong start for a feasibility study, but industrial quality control typically requires higher precision. The authors suggest optimizing hyperparameters and exploring more advanced architectures (like Vision Transformers) in the future.
  • Scope: The current model only identifies three species. The researchers plan to expand the database to include a broader range of TCM tablets and their common forgeries.

Summary updated

Abstract

<b>Background:</b> <i>Radix Bupleuri</i>, a kind of Chinese herbal medicine with great clinical use, is often confused with its adulterants, and it is difficult to identify it without certain knowledge. The existing identification methods have their own drawbacks, so a new method is needed to realize the identification of <i>Radix Bupleuri</i> and its adulterants. <b>Methods:</b> We used Micro Computed Tomography (Micro-CT) to perform tomography scans on <i>Radix Bupleuri</i> and its adulterants, performed data screening and data correction on the obtained DICOM images, and then applied 3D reconstruction, data augmentation, and ResNext deep learning model for the classification study. <b>Results:</b> The DICOM images after data screening, data correction, and 3D reconstruction can observe the differences in the microstructure of <i>Radix Bupleuri</i> and its adulterants, thus enabling effective classification and analysis. Meanwhile, the accuracy of classification using the ResNext model reached 75%. <b>Conclusion:</b> The results of this study showed that Micro-CT technology is feasible for the authentication of <i>Radix Bupleuri</i>. The pre-processed and 3D reconstructed tomographic images clearly show the microstructure and the difference between <i>Radix Bupleuri</i> and its adulterants without damaging the internal structure of the samples. This study concludes that Micro-CT technology provides important technical support for the reliable identification of <i>Radix Bupleuri</i> and its adulterants, which is expected to play an important role in the quality control and clinical application of herbs.
BibTeX
@inproceedings{chen2024feasibility,
  author = {Kehong Chen and Chen Gong and Zhelong Zhuang and Shouhua Luo and Jiaming Liu and Guorong Liu},
  title = {{Feasibility study on the introduction of Micro-CT technology for the identification of Radix Bupleuri and its adulterants}},
  booktitle = {Frontiers in Pharmacology},
  year = {2024},
  doi = {10.3389/fphar.2024.1347316},
}
Figure