RTC_TongueNet: An improved tongue image segmentation model based on DeepLabV3

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Artificial Intelligence in Medicine

Background:

• Automated tongue recognition in traditional Chinese medicine (TCM) is crucial but hindered by segmentation defects.
• Existing methods struggle with network degradation and fail to capture global features, impacting segmentation accuracy.

Purpose of the Study:

• To develop an improved tongue image segmentation model, RTC_TongueNet, addressing limitations of current approaches.
• To enhance the extraction of both local and global features for more effective tongue segmentation.

Main Methods:

• An improved DeepLabV3 architecture (RTC_TongueNet) was developed, integrating transformer and enhanced residual structures.
• The Efficient Channel Attention (ECA) module was incorporated into the Atrous Spatial Pyramid Pooling (ASPP) structure to improve feature fusion.
• The model was evaluated against FCN, UNet, LRASPP, and DeepLabV3 on two datasets.

Main Results:

• RTC_TongueNet demonstrated superior performance compared to baseline models on both datasets.
• The model achieved a 0.9-1.0% increase in Mean Intersection over Union (MIOU) and a 0.3-1.1% increase in Mean Pixel Accuracy (MPA) over DeepLabV3.
• The RTC_TongueNet model exhibited the best segmentation results on both evaluated datasets.

Conclusions:

• RTC_TongueNet effectively segments tongue images by leveraging improved feature extraction and attention mechanisms.
• The proposed model offers practical application and reference value for tongue image segmentation in TCM and related fields.