An Automatic Method for Sublingual Image Segmentation and Color Analysis.

Area of Science:

• Computer Vision
• Medical Imaging
• Traditional Chinese Medicine Diagnostics

Background:

• Accurate sublingual image analysis is crucial for Traditional Chinese Medicine (TCM) disease diagnosis.
• Current segmentation and color analysis methods for sublingual veins lack precision and are subject to inter-observer variability.
• Automated processing offers a non-invasive, convenient approach for tongue observation.

Purpose of the Study:

• To develop an improved method for sublingual image segmentation using a modified UNet++ network.
• To implement a robust color classification for sublingual veins via a triplet network.
• To introduce a linear discriminant analysis-based color quantization for objective, one-dimensional results.

Main Methods:

• Modified UNet++ for enhanced sublingual image and tongue dorsum segmentation.
• Triplet network for sublingual vein color classification.
• Linear Discriminant Analysis (LDA) for sublingual vein color quantization.

Main Results:

• Achieved 88.2% mIoU and 94.1% pixel accuracy for tongue dorsum segmentation.
• Attained 69.8% mIoU and 82.7% pixel accuracy for sublingual vein segmentation, outperforming state-of-the-art by 5.8% and 5.3% respectively.
• Sublingual vein color classification reached 81.2% overall accuracy (77.5% minority class recall); color quantization achieved 90.5% accuracy.

Conclusions:

• The proposed methods significantly enhance the accuracy of sublingual image segmentation and vein color analysis.
• These advancements provide objective, quantified data to aid TCM practitioners in disease diagnosis.
• The automated approach offers a reliable, non-invasive tool for clinical decision-making.