TL;DR
This paper introduces a novel system for capturing high-speed, 3D images of the vocal folds (glottis) in vivo. By combining stereo-endoscopy with advanced image processing, the researchers can now estimate the glottal area function, a crucial metric for understanding voice production, with unprecedented detail and accuracy. This work opens doors for more precise diagnostics and research into voice disorders.
What Problem Does This Paper Address?
Understanding how our voices are produced hinges on accurately observing the intricate movements of the vocal folds. The glottis, the space between these folds, changes shape rapidly during speech. Quantifying these changes, specifically the "glottal area function" (how the glottal opening changes over time), is vital for understanding vocal fold vibration, voice quality, and the mechanisms behind voice disorders.
Previous methods often relied on in vitro (lab-based, non-living) measurements or provided only relative estimates of glottal area. The challenge has been to capture high-resolution, dynamic, 3D information of the glottis in vivo (in a living person) in a way that allows for precise quantitative analysis. This paper aims to bridge that gap by developing a system capable of capturing these dynamic 3D movements.
Methods
The core of this research lies in a stereo-endoscopic high-speed digital imaging system. Here's a breakdown of their approach:
- Stereo-Endoscope: They utilized a specialized endoscope with two independent optical systems. These systems are designed with prisms to capture images from slightly different angles (70° oblique view), mimicking human binocular vision. The distance between the optical axes of these two tips is a critical parameter (10 mm).
- High-Speed Digital Camera: A Photron Fastcam 1024PCI camera was employed, capable of capturing images at an impressive 3750 frames per second (fps) at a reduced resolution. This high frame rate is essential for capturing the rapid oscillations of the vocal folds.
- 3D Coordinate Calculation: The system captures pairs of images (one from each endoscope lens) side-by-side. Using geometric principles and calibration constants derived from photographing a Cartesian graph paper, they calculate the 3D coordinates (Xp, Yp, Zp) of points on the glottis. This process involves accounting for optical distortions and calibrating the system to ensure accuracy.
- Glottal Edge Detection: For each frame, the system automatically detects the edges of the glottis in both the left and right images. This is done by identifying points with maximal brightness derivative among those with minimum brightness.
- Glottal Area Function Estimation: To represent the glottis as a 3D surface, they perform several steps:
- Smoothing: The detected edges are smoothed to account for the fact that glottal edges are continuous curves in 3D space.
- Planar Approximation: A regression line is fitted to the smoothed edges to represent the glottis as a plane. This involves calculating lateral inclinations between the left and right glottal edges.
- Glottal Hyperplane: A hyperplane (a 3D plane) is defined that best represents the glottis, considering the calculated inclinations.
- Projection: Glottal edge points are then projected onto this hyperplane to obtain the final 3D representation.
Datasets
The paper describes experiments conducted on one female participant without vocal problems. They recorded sustained vowels at three different fundamental frequencies (F₀):
- Middle F₀ (modal voice): 230 Hz
- High F₀ (modal voice): 450 Hz
- Low F₀ (vocal fry): 100 Hz
The data consists of stereo-endoscopic high-speed video recordings of the larynx during these phonations. No pre-existing datasets were used for training or evaluation; the data was generated specifically for this study.
Key Insights and Ideas
- In Vivo 3D Glottal Dynamics: The most significant insight is the successful development of a system capable of capturing dynamic, 3D information of the glottis in vivo. This moves beyond static or 2D analyses.
- Impact of Fundamental Frequency: The study demonstrates how glottal characteristics (area, length, width) vary significantly with fundamental frequency, particularly highlighting the difference between modal voice and vocal fry.
- Closing Phase Dynamics: Observations suggest that the glottis closing phase can be slightly shorter than the opening phase in each vibratory cycle, a subtle but important detail for understanding vocal fold mechanics.
- Lateral Inclination Matters (but can be simplified): The paper shows that accounting for the lateral inclination of the glottis improves accuracy. However, they also propose using a time-averaged inclination as a practical simplification, which still yields reasonably accurate results and is less sensitive to endoscope positioning difficulties.
Key Results
- Accuracy of 3D Reconstruction: The distribution of errors between real and estimated 3D coordinates had a median of 0.55 mm, with 95% of errors within 2.96 mm. Errors in Xp and Yp were less than 15% of the error in Zp.
- Glottal Dimensions at Different Frequencies:
- At F₀ = 230 Hz: Maximum glottal area was 11 mm², maximum width was 2 mm, and maximum length was 7 mm.
- At F₀ = 450 Hz: Maximum glottal area was ~5 mm², maximum length was ~8 mm, and maximum width was 1.0 mm.
- At F₀ = 100 Hz (vocal fry): Maximum glottal area was 1.3 mm², and maximum length was ~2 mm.
- Verification with a Slit: The method was verified using a rectangular slit, estimating its area as 29.5 mm² (compared to the actual 29.04 mm²), demonstrating the method's validity.
Limitations and Future Work
- Single Participant: The study was conducted on only one female participant, limiting the generalizability of the findings. Future work should involve a larger and more diverse participant pool.
- Definition of Glottal Area: The paper acknowledges that the glottal area is a section of a 3D space, and its definition can vary depending on the plane chosen. Further theoretical and practical improvements are needed for more robust definitions.
- Endoscope Positioning: While they propose a method to handle lateral inclination, precisely positioning the endoscope to minimize this inclination remains a challenge in the confined pharyngeal space.
- Future Directions: The authors explicitly state the need for further improvement in the estimation method and 3D reconstruction of laryngeal views. They also suggest exploring more advanced theoretical and instrumental approaches.
Summary updated