Encoder-decoder network with RMP for tongue segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate tongue segmentation is crucial for medical applications like disease identification and rehabilitation tracking.
• Existing methods may lack precision in isolating the tongue region for detailed analysis.

Purpose of the Study:

• To propose an encoder-decoder Convolutional Neural Network (CNN) architecture for precise tongue segmentation in images.
• To enhance the feature extraction capabilities for improved segmentation accuracy.

Main Methods:

• Developed an encoder-decoder CNN architecture for image-based tongue segmentation.
• Incorporated residual multi-kernel pooling (RMP) for multi-scale feature encoding.
• Evaluated the method on public datasets and a new dataset with varying tongue postures.

Main Results:

• The proposed CNN-based method achieved superior performance compared to existing techniques.
• Demonstrated effectiveness across different datasets and tongue postures.
• Showcased the benefit of re-training for adapting the model to unseen data.

Conclusions:

• The novel CNN architecture provides a robust solution for tongue segmentation.
• The method shows significant potential for real-world medical applications.
• Re-training is essential for successful deployment in clinical settings.