RAFF-Net: An improved tongue segmentation algorithm based on residual attention network and multiscale feature fusion

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Traditional tongue segmentation methods struggle with accuracy due to complex facial features, uneven tongue textures, and interference from lips.
• Existing approaches often fail to achieve precise and automated tongue region segmentation.

Purpose of the Study:

• To propose RAFF-Net, an automated tongue region segmentation network designed for enhanced accuracy and end-to-end segmentation.
• To address the limitations of traditional methods in segmenting tongue regions from complex face images.

Main Methods:

• RAFF-Net utilizes a UNet backbone with ResBlocks and Squeeze-and-Excitation (SE) blocks for feature extraction.
• The network features a simplified decoder, reduced parameter count, and a multiscale feature fusion module.
• A custom loss function was employed instead of cross-entropy to optimize network parameters and improve detection accuracy.

Main Results:

• RAFF-Net achieved a Mean Intersection over Union (MIoU) of 97.85% and an F1-score of 97.73%.
• These scores represent improvements of 0.56% and 0.46% over the original UNet architecture.
• Ablation experiments confirmed the effectiveness of the proposed algorithmic enhancements in improving tongue segmentation.

Conclusions:

• The integration of residual attention networks and multiscale feature fusion in RAFF-Net significantly enhances tongue segmentation accuracy.
• Optimizations including ResBlocks, SE block, multiscale feature fusion, and a weighted loss function improved network stability and overall performance.
• RAFF-Net offers a robust solution for automated and accurate tongue region segmentation in medical imaging applications.