Abstract
Coronary artery disease (CAD) is one of the main causes of premature mortality in the Western world, however with early and targeted treatment it is treatable. Calcium scoring multi-slice computed tomography (MSCT) imaging is used to visualize small calcified plaques located in the coronary arteries which represent a possible arterial narrowing, thus limiting or preventing the supply of oxygenated blood to the heart muscle. A clinician digitally quantifies the patient plaque burden to determine appropriate treatment, such as lipid lowering drugs, based on defined risk categories. MSCT is capable of imaging the entire heart within a single breath hold, thus exposing the patient to lower dose compared to coronary angiography, however this results in lower contrast images that may be affected by acquisition and motion artefacts. Clinically significant calcified plaques are located in the main coronary arteries, whereas other clinically insignificant calcified plaques can be found in close proximity but will not affect blood flow through the coronary arteries and thus does not concern the clinician for calcium scoring. This can lead to inter- and intra-observer variability when distinguishing between clinically significant and insignificant plaques, resulting in incorrect plaque burden quantification. Reducing the subjectivity and reproducibility in calcium scoring can lead to greater accuracy in patient treatment prescription as small score shifts will affect atherosclerosis risk categorisation. For a disease where treatment can include drug prescription, the reduction of non-required drug prescription is an important objective. A preliminary automation of the calcium scoring method was developed presenting good correlation between automated and observer generated risk assignment, however, variation in automated scores signified over or under inclusion of non clinically significant or clinically significant plaques respectively. Thus it was theorised that the extraction of the coronary artery tree would facilitate focussed calcified plaque assessment and quantification. To achieve this goal, a fully automated observer-less calcium scoring framework has been presented employing automated heart isolation and coronary artery tracking. Ground truth coronary artery trees are applied to aid vessel tracking, compensating for the poor image clarity and low axial resolution of MSCT calcium scoring images. Validation of this system employs ground truth clinical calcium scores obtained by an observer study using simulated CT images from the XCAT digital phantom. Assessment of atherosclerosis risk categorisation of the resulting scores from the framework presented in this thesis against ground truth shows a 93.8% system accuracy with a 85.4% sensitivity and 96.5% specificity when examined against the minimum cases required to represent agreement between all observers. However, combined observer performance of 0.729 from ROC analysis affirms observer variability in manual calcium scoring. Overall score difference analysis between framework and observer scores show an average of 10% overscore and underscore respectively, thus indicating compensation by the automated framework for observer variability based on the use of vessel tracking to identify actual coronary calcified plaques.