University of British Columbia
Automated thyroid ultrasound assessment : clinically aligned segmentation, TI-RADS risk stratification, and reader variability analysis
Abstract
dc:descriptionThyroid ultrasound is the first-line imaging modality for assessing thyroid nodules and guiding clinical decisions such as biopsy, follow-up, or active surveillance. In routine clinical practice, thyroid nodule assessment involves multiple coupled tasks, including nodule localization or segmentation, measurement, and risk stratification using systems such as the Thyroid Imaging Reporting and Data System (TI-RADS). However, automated thyroid ultrasound assessment remains challenging due to dataset heterogeneity, limited annotated data, variable ultrasound image quality, and inter-observer variability in image acquisition, contouri,ng and interpretation. This thesis investigates automated thyroid ultrasound assessment from a clinically aligned perspective. First, public thyroid ultrasound datasets are benchmarked for multiclass TI-RADS classification, highlighting the challenges introduced by inconsistent labels, class imbalance, and differences in dataset design. Second, binary classification experiments are conducted to compare the proposed methods with the broader thyroid ultrasound literature, where benign–malignant or clinically actionable binary endpoints are commonly reported. Third, a clinically guided multitask framework is studied for joint thyroid nodule segmentation and TI-RADS classification, reflecting the coupled nature of clinical reporting. Finally, a reader study is conducted using thyroid ultrasound data collected under institutional ethics approval to quantify inter-observer variability in thyroid nodule segmentation. Together, these studies show that automated thyroid assessment should not be treated as an isolated classification task. Instead, robust thyroid ultrasound artificial intelligence requires workflow-aware evaluation that accounts for segmentation, multiclass risk stratification, binary clinical endpoints, and variability among human readers.
Degree
thesis:*- Name thesis:degree_name
- Master of Applied Science - MASc
- Level thesis:degree_level
- master's
- Discipline thesis:degree_discipline
- Biomedical Engineering
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Salama, Dina
- Contributors dc:contributor
-
- Liu, David M.
- Hacihaliloglu, Ilker
Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-NoDerivatives 4.0 International
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2429/95457