Dissertation
Towards accessible deep learning for upper-airway segmentation: open-source tool development and application in voice and aleep apnea studies
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Spring 2025
DOI: 10.25820/etd.007972
Abstract
The human vocal tract is an incredible instrument in our body, playing a vital role in physiological processes such as speech, singing, and breathing. The modulation of various articulators, like the tongue, velum, and lips, enables the smooth production of sound and breathing. Any disturbances in the shape or dynamics of these articulators can result in conditions such as impaired speech, cleft palate, or sleep apnea. Researchers and clinicians have increasingly utilized modalities including Electro-magnetic Articulography (EMA), Endoscopy, Projection X-ray, X-ray Computed Tomography, Ultrasound, and Magnetic Resonance Imaging (MRI). Amongst these, MRI has emerged as a powerful tool due to its non-invasive nature, flexibility in imaging orientation, ability to image deep vocal structures, and suitability for longitudinal studies. However, MRI is relatively slow compared to other modalities, potentially prolonging data acquisition. Recent advancements in accelerated MRI techniques, such as parallel imaging with custom neck coils and compressed sensing, have enabled rapid imaging of the entire vocal tract during sustained speech sounds.
Quantitative assessment of vocal tract posture modulation, such as through 3D vocal tract area functions and 2D mid-sagittal vocal-tract area change analysis, has informed numerous speech applications. These include understanding speaker-to-speaker variability, developing and optimizing vocal tract models for voice synthesis, and conducting acoustic and aerodynamic studies of 3D-printed vocal tracts derived from MRI data and research on sleep apnea. Accurate segmentation of the vocal tract from MRI data is crucial for these analyses. Typically, they are segmented via manual annotation. However, this method is time and labor-intensive as well as error prone. Accurate segmentations in 3D can take up to 90 minutes of manual editing for 64 slices, and studies of manually segmented vocal tracts typically have small sample sizes and weak statistical power. In addition to that lack of open-source datasets in this domain has slowed down the development of Machine learning based segmentation methods.
This thesis addresses these challenges by presenting various segmentation methods for multimodal MRI data that are both data-efficient and robust to different MRI protocols. It also involves creating open-source datasets with expert segmentations for 2D and 3D MRI scans. The neural networks developed in this work are specifically modified to overcome the scarcity of large datasets in the domain of upper airway MRI. The immediate applications of these studies include understanding changes in the dynamics of the upper airway during speech and in sleep apnea.
Details
- Title: Subtitle
- Towards accessible deep learning for upper-airway segmentation: open-source tool development and application in voice and aleep apnea studies
- Creators
- Subin Erattakulangara
- Contributors
- Sajan Lingala (Advisor)Joseph Reinhardt (Committee Member)Sarah Gerard (Committee Member)David Meyer (Committee Member)Junjie Liu (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Biomedical Engineering
- Date degree season
- Spring 2025
- DOI
- 10.25820/etd.007972
- Publisher
- University of Iowa
- Number of pages
- xv, 111 pages
- Copyright
- Copyright 2025 Subin Erattakulangara
- Grant note
- The NIH, through NHLBI grant R01 HL173483, has been crucial in supporting this work. We utilized an MRI instrument funded by the NIH under grant 1S10OD025025-0. Furthermore, financial backing came from the Roy J. Carver Department of Biomedical Engineering, the Department of Radiology, and the Department of Otolaryngology
- Language
- English
- Date submitted
- 04/19/2025
- Description illustrations
- illustrations (some color)
- Description bibliographic
- Includes bibliographical references (page 94-111).
- Public Abstract (ETD)
The human vocal tract is an integral part in our human anatomy that enables essential functions like speaking, singing, and breathing. Its smooth operation relies on precise coordination of various parts, such as the tongue, velum, and lips. Disruptions to the shape or movement of these structures can lead to conditions like speech impairments, cleft palate, or sleep apnea. To study and understand these conditions, researchers rely on imaging techniques like magnetic resonance imaging (MRI), which provides detailed, non-invasive views of the vocal tract.
While MRI offers incredible insights, it comes with challenges, particularly its slow data acquisition speed and the time-consuming process of manually labeling images for analysis. My research addresses these issues by developing advanced machine learning techniques to automatically and accurately segment MRI images of the vocal tract. These methods are designed to work efficiently with limited data, a common hurdle in this field. I’ve also created open-source datasets and introduced neural networks capable of analyzing 3D, 2D dynamic, and 2D static MRI data, providing new tools to study how the vocal tract changes during speech and breathing. The outcomes of this research have immediate applications in areas like speech synthesis, acoustic modeling, and sleep apnea studies, offering a path toward better diagnosis, treatment, and understanding of vocal tract dynamics.
- Academic Unit
- Roy J. Carver Department of Biomedical Engineering
- Record Identifier
- 9984830923902771
Metrics
3 Record Views