TongueTransUNet: toward effective tongue contour segmentation using well-managed dataset

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Speech Science

Background:

• Medical image analysis is crucial for telehealth and healthcare systems, but deep learning faces challenges like data size, labeling, and feature extraction.
• These challenges lead to complex, expensive, and opaque AI models, sometimes producing undesirable outputs.
• Lingual ultrasound analysis offers potential for understanding language behavior and providing biofeedback.

Purpose of the Study:

• To develop an effective AI design strategy for analyzing lingual ultrasound images.
• To extract tongue contours for language behavior analysis and biofeedback applications.
• To build a cumulative foundation model addressing deep learning challenges in medical image analysis.

Main Methods:

• A hybrid architecture combining UNet, Vision Transformer (ViT), and contrastive loss in latent space was employed.
• Human experts validated training data, and UNet/ViT encoders extracted feature representations.
• Contrastive loss compared new embeddings with reference embeddings, with a UNet decoder reconstructing the image.

Main Results:

• The proposed method demonstrated improved accuracy compared to traditional techniques.
• The AI model effectively extracts high-quality and relevant features from lingual ultrasound data.
• A quality control mechanism with human expert intervention was integrated for rejected segmentations.

Conclusions:

• The hybrid AI model provides a robust solution for lingual ultrasound image analysis.
• This approach enhances the reliability and interpretability of AI in medical image analysis.
• The developed foundation model can be cumulatively built for diverse biofeedback applications.