Towards vocal tract MRI synthesis from facial signals using external to internal correlation modelling
Publication Date
July 30, 2025
Creator
Abstract
Oral health underpins everyday functions such as speech, mastication and swallowing, yet acquiring detailed kinematic data on the vocal tract remains technically and financially demanding. Ultrasound and electromagnetic articulography offer only partial coverage, while Real Time Magnetic Resonance Imaging (RtMRI) data delivers richer information but requires expensive scanners and bespoke acquisition protocols. These constraints limit large-scale studies and the routine use of dynamic vocal-tract models in both research and clinical practice.
Motivated by the need for an affordable, non-invasive alternative, this thesis introduces External to Internal Correlation Modelling (E2ICM), a novel framework that learns correlations between external facial signals and internal articulator motion, enabling vocal-tract modelling without direct imaging. The work pursues four objectives: (i) advanced segmentation of RtMRI sequences, (ii) quantification of articulator interdependencies, (iii) prediction of internal motion from purely external observations, and (iv) ethical evaluation of AI-driven approaches in oral healthcare.
Both static and temporal segmentation pipelines are developed for RtMRI data. Generative adversarial networks and diffusion models are then employed to synthesise internal views from facial video, addressing data scarcity through tailored augmentation strategies. A thematic analysis of professional interviews highlights concerns around privacy, security and algorithmic bias, informing an ethical framework for clinical deployment.
A key contribution is a dual-view dataset comprising synchronised high-resolution RtMRI and external video captured during controlled speech and chewing tasks. Experimental results demonstrate that (E2ICM can predict vocal-tract configurations with promising accuracy while reducing reliance on costly imaging. Improved segmentation techniques and a deeper understanding of articulator dynamics further advance the state of the art in non-invasive oral-movement modelling.
Item Type
ethesis
Thesis Type
PhD
Supervisors
Subjects (LC)
Associated Schools / Departments
School of Computer Science (UK)
eprints ID
81139
UoN Repository URI
Except where otherwise noted, this item's license is described as
File(s)![Thumbnail Image]()
Name
University_of_Nottingham_PhD_Thesis_Suhaib__Final_correction.pdf
Type
Full-text
Description
Examined
Size
28.46 MB
Format
Adobe PDF
Checksum (MD5)
845f5694bf9e8453b05b9e834ade0611