Automatic vocal tract landmark localization from midsagittal MRI data.

Area of Science:

• Medical Imaging
• Speech Science
• Artificial Intelligence

Background:

• Accurate analysis of speech production requires identifying anatomical landmarks on medical images.
• Current manual methods are time-consuming and struggle with data variability and image quality.

Purpose of the Study:

• To develop an automated deep learning solution for landmark identification in speech-related medical images.
• To address challenges of inter- and intra-speaker variability and moderate image quality.

Main Methods:

• A novel deep learning network architecture, Flat-net, was proposed.
• The network was trained and evaluated on midsagittal Magnetic Resonance Images of 9 speakers with 62 articulations and 21 annotated landmarks per image.
• Performance was compared against eleven state-of-the-art methods.

Main Results:

• The Flat-net approach demonstrated superior performance compared to existing methods.
• An overall Root Mean Square Error of 3.6 pixels (0.36 cm) was achieved using a leave-one-out cross-validation strategy.
• Implementation code was made publicly available.

Conclusions:

• Deep learning, specifically the Flat-net architecture, offers a robust and accurate solution for automated anatomical landmark detection in speech imaging.
• This advancement facilitates quantitative analysis for speech production research and clinical applications.
• The public release of the code promotes further research and development in the field.