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
Josh Lewis
Abstract
Cardiovascular Diseases (CVDs) are among the leading causes of death worldwide, with Abdominal Aortic Calcification (AAC) as a stable marker of CVDs. AAC appears in the walls of the abdominal aorta near the lumbar region of the spine and can be detected through different imaging modalities like Computed Tomography (CT), Digital X-ray imaging, or Dual-energy X-ray Absorptiometry (DXA). Although CT is the gold standard, DXA imaging is the preferred modality due to its comparatively lesser radiation exposure, and cost-effectiveness. Conventionally, it is performed for Bone Mineral Density (BMD) measurement, osteoporosis analysis, and vertebral fracture assessment; however, it can also be used for AAC detection and quantification. Although this imaging modality has low radiation exposure and cost, DXA image analysis for AAC detection brings challenges due to low contrast, vague vertebral boundaries, and artifacts. The research mentioned in this thesis addresses various challenges associated with AAC detection and vertebral landmark localization using lateral spine DXA images. First, it introduces a deep learning-based framework, ‘GuideNet,’ which localizes the vertebral corners of the lumbar region and generates Intervertebral Guides (IVGs), showing encouraging results as an assistive tool for image readers. Next, this thesis presents ‘AACLiteNet,’ a deep learning model that predicts granular AAC scores from lateral spine DXA images, with hazard ratios for Major Acute Cardiovascular Event (MACE) outcome predictions aligning closely with those from trained human assessors. Next, a novel architecture, ‘Hybrid-FPN-AACNet,’ is introduced, achieving SOTA performance in granular AAC score prediction from lateral spine DXA images. An extension of Hybrid-FPN-AACNet termed ‘VerteNet’, is also presented, showing improved performance and the ability to handle DXA images from various sources for IVGs generation and abdominal aorta crop detection. Lastly, this thesis presents ’GLMHA,’ a lightweight and efficient Channel-wise Self-Attention (CSA) mechanism that can be effortlessly incorporated into advanced hybrid CNN-Transformer models for image restoration and spectral reconstruction. It offers a substantial reduction in computational cost with minimal performance impact when replacing conventional attention mechanisms.
DOI
10.25958/xmfb-9986
Access Note
Access to this thesis is embargoed until 21st March 2027
Recommended Citation
Ilyas, Z. (2025). Deep learning for medical image interpretation. Edith Cowan University. https://doi.org/10.25958/xmfb-9986