Machine learning for cardiovascular disease analysis in chest CT

Cardiovascular disease is the leading cause of morbidity and mortality worldwide. Moreover, the largest lung screening trial—the National Lung Screening Trial—has shown that in a population of heavy smokers, cardiovascular disease is the major cause of mortality next to lung cancer. Cardiovascular disease can progress without symptoms for years, but suddenly reveal its presence with acute adverse events like stroke or myocardial infarction, which are sometimes even fatal. While smoking is sa well-known risk factor for cardiovascular disease, some heavy smokers develop cardiovascular disease and some do not. Early detection of lung-screening participants who are at increased cardiovascular risk might prove beneficial.

A direct sign of presence of cardiovascular disease are calcified arteries. The exact process of arterial calcification is unclear, but it is considered an advanced stage of atherosclerosis and therefore of cardiovascular disease. For instance, calcification in the coronary arteries is predictive for myocardial infarction and calcification of the thoracic aorta has been related to stroke. Quantification of calcium (i.e. calcium scoring) is standardly performed by experts in dedicated cardiac CT images, but it can also be quantified in Chest CTs obtained for lung screening. In fact, calcium scored in chest CT is proven useful to predict cardiovascular risk.

In this thesis machine learning based methods are proposed that automatically analyze lung screening chest CT for cardiovascular disease. To this end several novel deep learning based methods are introduced that can be used to automatically localize or precisely delineate organs in these medical images. The thesis also describes methods that use deep learning to increase speed of image registration, i.e. image alignment. This methodology is employed in a new automatic calcium scoring method that can achieve a calcium score in less than a second, several hundreds of times faster than previously proposed automatic methods. Subsequently, this method is used in a prediction model developed with chest CTs from more than 4 500 lung screening participants to predict 5 year mortality risk. The proposed prediction model provides insight in the locations of arterial calcium that were most predictive for mortality.