Deep Upscale U-Net for automatic tongue segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate tongue movement analysis is vital for diagnosing oral and oropharyngeal cancers.
• Automatic tongue segmentation is a prerequisite for measuring tongue movements.
• Existing U-Net models can experience feature loss in deeper network layers.

Purpose of the Study:

• To propose an improved U-Net architecture, termed Deep Upscale U-Net (DU-UNET), for enhanced tongue segmentation.
• To address the challenge of feature loss in deep convolutional layers within segmentation networks.
• To develop a robust tongue segmentation model applicable to real-world, far-distance imaging scenarios.

Main Methods:

• Implemented a novel Deep Upscale U-Net (DU-UNET) architecture.
• Incorporated additional up-sampling of feature maps from the contracting path to upper layers of the expansive path.
• Trained the DU-UNET model on publicly available datasets and a self-collected dataset of tongue images with five postures.

Main Results:

• The DU-UNET model achieved superior performance compared to existing methods.
• Achieved an accuracy of 99.2%.
• Obtained a mean Intersection over Union (IoU) of 97.8%, Dice score of 96.8%, and Jaccard score of 96.8%.

Conclusions:

• DU-UNET effectively overcomes feature loss issues in deep convolutional layers.
• The proposed model demonstrates high accuracy and robustness for tongue segmentation in real-world conditions.
• DU-UNET shows significant potential for improving diagnostic tools in oral cancer detection.