Felix Denzinger
Deep Learning-based Grading of Coronary Artery Disease from Coronary CT Angiography Scans
One of the deadliest disease types in modern societies is coronary artery disease (CAD). It is often related to coronary artery plaques aggregating within the vessel wall obstructing the lumen causing a stenosis. To support physicians this thesis aims to answer the following research question: Can we perform an automated CAD assessment from coronary CT angiography (CCTA) scans using machine learning (ML)? To answer this question, we tackle three main tasks: First, we develop deep learning (DL)-based methods capable of predicting a significant stenosis degree and whether a lesion leads to a revascularization procedure. Four approaches with different characteristics are developed on this task, striving for a prior knowledge induced data representation and ML method design. Depending on whether a prior segmentation step is performed or not we reach an area under the receiver operating characteristic curve (AUC) of 0.96/0.92 for significant stenosis detection and 0.88/0.90 for the revascularization decision target with/without the segmentation. As a second task, methods to automatically determine the coronary artery disease-reporting data system (CAD-RADS) score – a patient-level CAD severity score – is developed. We leverage the best performing approach from the first task and embed it in a taskspecific hierarchical architecture to aggregate single coronary subsegment features to allow a patient-level prediction. This approach is enhanced with a synergizing heuristic centerline labeling approach and auxiliary targets to reach an AUC of 0.942, 0.950 on the task of finding patients with CAD and on the task of detecting patients with an obstructive CAD respectively. With this strong performance, we tackle a third task of evaluating the clinical applicability of our CAD-RADS scoring approach. In a first step, we examine how changing some commonly altered CCTA image formation parameters influences the predictions of our approach. Here, we find that the overall performance stays on a high level, but predictions for individual patients changes. From this we conclude a need to create a more robust approach with respect to technical variation. In a second step, we develop an approach to automatically detect a norm variant of the coronaries, as out-of-domain samples may adversely impact ML-based CAD grading systems. On this task, we achieve a strong performance with an AUC of around 0.938. Additionally, we propose a quantile-based abstention approach, as an automated CAD grading system should know when a decision is better left to the human reader. Overall, this thesis concludes that – with limitations – its main research question can be answered with a “Yes”. A well-performing CAD grading system was developed, but future work on robustness with respect to technical variation, the handling of anatomical outliers and explainability of the method at hand remain on the horizon.