Tongue Images Classification Based on Constrained High Dispersal Network.

Area of Science:

• Computer Science
• Medical Imaging
• Traditional Chinese Medicine

Background:

• Traditional Chinese Medicine (TCM) diagnosis often relies on tongue observation, but current computer-aided methods use limited low-level features, hindering holistic analysis.
• Existing deep convolutional neural network (CNN) models face challenges with feature redundancy and imbalanced datasets in tongue image analysis.

Purpose of the Study:

• To develop a novel feature extraction framework, constrained high dispersal neural networks (CHDNet), for unbiased and high-level feature extraction in TCM tongue diagnosis.
• To address redundancy and improve handling of unbalanced sample distribution in CNN models for tongue image classification.

Main Methods:

• Proposed CHDNet incorporates high dispersal and local response normalization to mitigate redundancy.
• Multiscale feature analysis was integrated to enhance robustness against deformation.
• The framework focuses on learning high-level features for improved classification accuracy.

Main Results:

• CHDNet was evaluated on 267 gastritis patients and 48 healthy controls.
• The model demonstrated promising results in tongue image classification for TCM applications.
• The proposed method offers a more holistic view compared to low-level feature analyses.

Conclusions:

• CHDNet presents a significant advancement in computer-aided tongue diagnosis within TCM.
• The framework's ability to learn high-level, unbiased features enhances classification accuracy.
• CHDNet shows potential for reducing human labor and improving diagnostic efficiency in TCM.