Preview

PDF, English Download (36MB) | Terms of use

Abstract

Deep learning has revolutionized medical imaging by providing state-of-the-art solutions for a wide range of tasks. However, the use of deep learning in medical imaging also comes with its own set of challenges, three of which are addressed in the projects presented in this cumulative thesis. The first project focuses on the fair evaluation of deep learning-based low-dose computed tomography (LDCT) image denoising algorithms by introducing a novel benchmark framework. The second project addresses the interpretability and robustness of deep learning algorithms for LDCT image denoising by investigating their invariances (i.e., which features in the images they learned to represent and which to ignore). The third project tackles the scarcity of data for deep learning-based digital subtraction angiography (DSA) by simulating paired training data. The proposed methods are capable of overcoming the respective challenges associated with deep learning in medical imaging and through this could enable the development of better and safer algorithms for clinical practice.