Rapid semi-automatic segmentation of real-time magnetic resonance images for parametric vocal tract analysis

Michael I. Proctor, Daniel Bone, Athanasios Katsamanis, Shrikanth S. NarayananInterspeech 2010 · DOI

Introduction

Studying the dynamics of the human vocal tract—including the movement of the tongue, lips, and pharynx—is essential for fields ranging from linguistics to speech pathology. Real-time MRI (rtMRI) has emerged as a powerful tool for this research, providing high-speed video of internal articulators. However, rtMRI generates a massive volume of data; a single short utterance can produce hundreds of image frames. Manually tracing the airway in every frame is labor-intensive and prone to inconsistency. Conversely, fully automated methods often fail to account for "invisible" structures like the hard palate and teeth, which do not appear clearly in MRI but are critical for defining speech sounds.

Methodology: A Semi-Automatic Pipeline

Researchers at the University of Southern California have developed a semi-automatic approach that balances speed with anatomical accuracy. The pipeline consists of several key stages designed to transform raw MRI frames into parametric data.

Analysis Grid Construction

The algorithm superimposes a custom "tract-normal" grid over the MRI frames. This grid uses a combination of linear and polar coordinates tailored to the anatomy of the vocal tract, extending from the glottis to the lips. This constrained search space prevents the algorithm from being misled by background noise or adjacent structures like the spine.

The analysis grid superimposed on a raw MRI frame, providing a structured search space for segmentation.

Centerline and Boundary Detection

To locate the airway, the system treats the image as a graph and employs Dijkstra’s algorithm to find the "optimal path" of lowest pixel intensity, representing the air-filled tract. Once this centerline is established, the algorithm searches outward along the grid lines to identify the steepest gradients in pixel intensity, marking the transition from air to tissue.

Incorporating Anatomical Constraints

To address the "invisible teeth" problem, the system utilizes manual reference frames. A researcher identifies the palate and dental boundaries in a few high-quality frames, which are then superimposed onto the rest of the sequence to guide the automatic segmentation. Finally, a Discrete Cosine Transform (DCT) is used to low-pass filter the tongue surface, removing jitter while preserving its natural shape.

Results and Validation

The method was evaluated using 50 image frames from a multilingual rtMRI database featuring five speakers (Tamil, English, German, Serbian, and Mandarin). The automatic segmentation was compared against manual traces performed by an expert phonetician.

Comparison of automatic (blue) and manual (red dashed) aperture measurements along the vocal tract, with the corresponding error shown at the bottom.

The system achieved a mean segmentation error of 0.988 mm per gridline. Performance was particularly strong in the pharynx (0.825 mm error) and mid-oral regions (0.703 mm error). The highest errors (~1.6 mm) occurred in the dental region, where accuracy depends on the registration of manual reference frames to the moving video. The utility of the tool was further demonstrated through the analysis of complex tongue shaping in Tamil "liquid" consonants (like 'l' and 'r' sounds).

Conclusion

This semi-automatic method provides a robust framework for processing large rtMRI datasets with sub-millimeter accuracy. By combining human expertise for static structures with computational efficiency for dynamic tracking, the system overcomes the primary bottlenecks of manual analysis. Future work will focus on improving performance in noisier data, incorporating temporal constraints between frames, and expanding the approach to 3D modeling.

Summary updated

BibTeX
@inproceedings{proctor2010rapid,
  author = {Michael I. Proctor and Daniel Bone and Athanasios Katsamanis and Shrikanth S. Narayanan},
  title = {{Rapid semi-automatic segmentation of real-time magnetic resonance images for parametric vocal tract analysis}},
  booktitle = {Interspeech},
  year = {2010},
  doi = {10.21437/interspeech.2010-458},
}
Figure