Tongue model construction based on ultrasound images with image processing and deep learning method

Area of Science:

• Biomedical Engineering
• Speech Science
• Medical Imaging

Background:

• Lateral articulation (LA) is a speech disorder affecting articulation.
• Accurate 3D tongue models are crucial for understanding and treating speech disorders.
• Ultrasound (US) imaging offers a non-invasive method for visualizing tongue movement.

Purpose of the Study:

• To construct a 3D tongue model.
• To generate a dynamic animation of tongue movement.
• To develop a tool for speech therapy in patients with lateral articulation.

Main Methods:

• Generating a 3D tongue model using ultrasound (US) images.
• Employing image processing and deep learning for tongue surface extraction.
• Creating a reference model from a normal speaker and adapting it for an LA patient.
• Generating tongue movement animation through temporal deformation of the model.

Main Results:

• The deep learning method achieved 49% accuracy for normal speakers and 64% for LA patients in estimating tongue surfaces.
• Maximum vertical errors were 1.01 mm (normal) and 1.03 mm (LA).
• The constructed 3D tongue model demonstrated high fidelity to actual tongue shapes.

Conclusions:

• A functional 3D tongue model and animation were successfully created for an LA patient using US images.
• The low maximum vertical error (1.03 mm) indicates the model's precision.
• The developed tongue model is a valuable tool for effective speech therapy in LA patients.