University of Glasgow
Deep learning for ultrasound tongue imaging: towards robust, interpretable, and deployable assessment of speech sound disorders
Abstract
dc:description.abstractSpeech sound disorders (SSDs) are among the most common developmental communication difficulties in childhood, with long-term consequences for intelligibility, literacy, and psychosocial well-being. Clinical assessment relies primarily on auditory–perceptual judgement, which, although effective, is subjective and provides limited insight into underlying articulatory mechanisms. Ultrasound tongue imaging (UTI) offers a safe, non-invasive method for visualising tongue movement during speech; however, its interpretation remains challenging due to image noise, speaker and acquisition variability, and the high cost of expert annotation. This thesis examines the systematic adaptation of deep learning (DL) to address three core challenges in automated UTI analysis: (C1) data variability and generalisability limitations, (C2) data scarcity and annotation inefficiency, and (C3) lack of interpretability and clinical usability. Reproducible baseline deep neural network (DNN) models are first established for phonetic classification from raw UTI, quantifying generalisation limitations under speaker-independent evaluation. A novel multi-input FusionNet architecture is then introduced, combining raw ultrasound frames with texture-based representations to improve cross-speaker robustness. A two-stage conditional generative adversarial framework is proposed for field-of-view (FoV) standardisation and tongue region enhancement, improving image consistency and classification performance across domains. To address data scarcity, a cost-focused framework integrates statistical power-curve modelling with active learning to optimise annotation effort, achieving substantial reductions in required labelled data while maintaining clinically meaningful accuracy. Model interpretability is examined using an explainable AI (XAI) technique to assess how image representation and standardisation influence network attention and anatomical relevance. Finally, the feasibility of clinical translation is demonstrated through a prototype web-based deployment system for real-time inference and visualisation. Collectively, this work presents an integrated DL framework that advances robustness, data efficiency, interpretability, and deployment for ultrasound-based speech assessment, contributing toward objective and scalable clinical decision-support tools for SSDs.
Degree
thesis:*- Level dc:type.qualificationlevel
- PhD
- Grantor dc:publisher.institution
- University of Glasgow
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Al Ani, Saja
Rights
- Language dc:language
- en
Identifiers
dc:identifier.*- DOI dc:identifier.doi
- https://doi.org/10.5525/gla.thesis.86062