Deep learning-based derivation of physiological information from cardiac CT angiography

Coronary artery disease (CAD) is a leading cause of death worldwide. It is characterized by the accumulation of atherosclerotic plaque in the coronary arteries. This buildup can lead to stenosis, a narrowing of the arteries, which is functionally significant if it results in myocardial ischemia. The current clinical standard for determining the functional significance of stenosis is given by invasive fractional flow reserve (FFR) measurement. However, this procedure is invasive, costly and burdensome for patients.
While coronary CT angiography (CCTA) allows for the visual identification of most functionally significant stenoses, it suffers from low specificity. This causes a significant number of unnecessary invasive FFR measurements, emphasizing the need for improved patient selection schemes.
This thesis addresses the non-invasive selection of patients who may require invasive treatment by using machine learning (ML), specifically deep learning, to derive FFR from CCTA. Deep learning enables the automatic extraction of robust features from high-dimensional input data, which facilitates accurate and fast FFR prediction without the need for manual intervention. The ultimate goal of this work is to enable automatic and non-invasive FFR prediction, providing clinicians with a tool that can assist in assessing coronary artery disease and making informed decisions about whether invasive procedures are necessary. To achieve this goal, several steps were taken, including the design of a multi-step deep learning model that localizes and characterized the arteries into supervised and unsupervised features, to balance model interpretability and performance.