Date of Award
2025
Document Type
Thesis
Publisher
Edith Cowan University
Degree Name
Doctor of Philosophy
School
School of Science
First Supervisor
Syed Zulqarnain Gilani
Second Supervisor
David Suter
Third Supervisor
Joshua Lewis
Fourth Supervisor
Ghulam Mubashar Hassan
Abstract
Machine learning has significantly transformed medical image analysis in the current age of artificial intelligence offering vast potential in improving disease diagnosis and management. Cardiovascular diseases (CVDs) are among the leading cause of global mortality, emphasizing the need for early detection for effective intervention and prevention. Abdominal Aortic Calcification (AAC) is an early indicator and contributor to Atherosclerotic Cardiovascular Diseases (ASCVDs) and is commonly assessed through imaging modalities such as computed tomography (CT), X-rays, and Dual-energy X-ray Absorptiometry (DXA). Among these, lateral spine DXA scans, commonly used for osteoporosis screening, offer a cost-effective and low-radiation opportunity for opportunistic CVD risk assessment. Despite advancements in medical imaging technologies, AAC evaluation still relies on manual interpretation by trained clinicians, a process that is labor-intensive, subjective, and prone to variability. Automating the process of AAC quantification can address these challenges and enable consistent, early screening for CVD risk.
This research presents robust machine-learning frameworks for the automated and accurate prediction of the AAC-24 score and its classification into relevant risk classes (low, moderate, and high). First, we explore deep feature ensembling methods to develop a deep feature fusion network for AAC-4 scoring using regression loss. However, its performance was limited by class ambiguities from inter-class similarities, intra-class variations, and low resolution VFA DXA artifacts. To mitigate this problem, we formulate AAC-24 scoring as an ordinal regression problem and propose a novel supervised contrastive ordinal learning (SCOL) framework. SCOL leverages a label-dependent distance metric to capture the ordinal nature of AAC labels. Using SCOL, we develop a Dual-encoder Contrastive Ordinal Learning (DCOL) framework to learn contrastive ordinal representation at global and local levels, improving feature separability and class diversity in the latent space among the AAC-24 categories. Clinical validation demonstrated a strong association between ML-AAC-24 scores and ASCVD risk, with substantial agreement between ML predictions and expert assessments. To enhance generalizability across different imaging modalities, SCOL framework is further explored for lateral spine X-rays via cross-domain fine-tuning, enhancing its utility in diverse clinical settings.
To strengthen this work on highly imbalanced disease grading medical datasets, a prototype-based learning approach is incorporated within the SCOL framework to develop a generic disease grading system. The framework is evaluated on public datasets for diabetic retinopathy grading and breast cancer staging, demonstrating its ability to learn robust, ordinal-aware prototypes that generalize across diverse medical imaging tasks. Additionally, to enhance the interpretability and reliability of automated systems in clinical diagnosis, we develop a context-aware ordinal learning framework for granular-level AAC-24 scoring. We address the challenges of SCOl in handling class imbalance for ordinal regression tasks and introduce SCOL+. We explore SCOL+ in a multi-label setting to determine the extent of calcification in each section of the aorta to aid clinicians in making detailed and interpretable diagnoses.
In this thesis, the AAC algorithms are developed using five large clinical datasets obtained from machines with different manufactures, including patients from Australia, Canada, and the United States, spanning both male and female patients. In conclusion, as DXA scans are commonly captured in various clinical scenarios, this research offers a novel and opportunistic approach to cardiovascular disease detection and monitoring in clinical practice, potentially revolutionizing the way we approach CVD risk screening.
DOI
10.25958/m2sy-xm36
Access Note
Access to this thesis is embargoed until 24th May 2026
Recommended Citation
Saleem, A. (2025). Machine learning for computer-aided diagnostics from complex medical images. Edith Cowan University. https://doi.org/10.25958/m2sy-xm36