Deep learning in cardiovascular imaging:Using A1 to improve risk predictions and optimize clinical workflows

Cardiovascular disease is the most common preventable cause of death, accounting for up to 45% of mortality in Europe and 31% in the United States. This PhD research focused on developing robust and efficient deep learning systems applied to radiological data to improve cardiovascular risk predictions. This research was conducted in cooperation with experts from the Harvard Medical School, Dana-Farber Cancer Institute, Massachusetts General Brigham and Maastricht University. This deep learning system was able to automatically predict cardiac risk from computed tomography scans as good as medical experts and in some scenarios even surpassing human performance. The analyses were focused on real world applicability, generalization and robustness. Therefore, very large, distinct and well established datasets to validate the performances of the developed systems were used. Furthermore, all code and trained deep learning models were made publicly available without restrictions. In summary, the presented research showed the potential of deep learning to automate and improve medical research and clinical treatment and is on the verge to be applied in daily clinical routines