Encoder-decoder CNN models for automatic tracking of tongue contours in real-time ultrasound data

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Speech Science

Background:

• Ultrasound imaging visualizes tongue dynamics for speech production studies.
• Low-contrast and noisy ultrasound images challenge accurate tongue shape and motion analysis.
• Manual tongue contour extraction is time-consuming, subjective, and unsuitable for real-time applications.

Purpose of the Study:

• To develop advanced deep learning models for precise and efficient real-time tongue contour extraction from ultrasound data.
• To address the limitations of manual contour extraction and improve quantitative analysis of tongue motion.

Main Methods:

• Development of two novel deep neural networks, termed BowNet models.
• Utilizing encoding-decoding fully convolutional networks for global prediction.
• Incorporating dilated convolutions for full-resolution feature extraction.

Main Results:

• The BowNet models demonstrated outstanding performance in speed and robustness across datasets from two ultrasound machines.
• Significant improvements in the accuracy of prediction maps were observed.
• The models effectively addressed challenges posed by low-contrast and noisy ultrasound images.

Conclusions:

• The proposed BowNet models offer a robust and efficient solution for real-time ultrasound tongue contour extraction.
• These deep learning advancements enhance the study of speech production, aiding in the analysis of both healthy and impaired speech.
• The models provide a foundation for more accurate quantitative analysis of tongue motion in clinical and research settings.