On Deep Learning in Cardiovascular Magnetic Resonance Imaging

Cardiovascular disease is the leading global cause of death. Cardiovascular magnetic resonance (CMR) imaging is the gold standard method for evaluating cardiac function and structure and is thus an important tool for diagnostics in cardiovascular disease. Recently, deep learning (DL) methods have achieved notable developments in various aspects of CMR. The aim of this thesis was to further assess the benefits of incorporating DL-based methods in CMR, specifically in the process of evaluating cardiac function. This was done in four studies that explored different applications and aspects of DL in CMR. In these, DL was applied alongside other new developments in image acquisition, reconstruction, and analysis. Assessments were conducted on the usefulness of the specific DL applications, and on how randomness in DL can affect the reporting of results.In Study I, a DL-based right ventricular segmentation pipeline was developed, and the aspect of clinical time reduction was assessed, indicating that it could substantially reduce the time required to produce segmentations of clinically useful quality. Study II assessed how randomness affects the reliability of standard methods for comparing the performance of DL-based image segmentation models. Randomness caused statistically significant differences between models from the same learning algorithm, indicating that the standard methods cannot be used to reliably compare different learning algorithms. In Study III, a method was developed for synchronizing time-resolved ventricular cine images from 2D real-time exercise CMR. The results indicated that the method is feasible for deriving time-resolved measures of cardiac function and showed that DL could help accelerate the method. In Study IV, a rapid simultaneous multi-slice (SMS) bSSFP sequence and a DL-based image artifact suppression reconstruction method were developed for real-time CMR, achieving a reasonable agreement with standard CMR in volume-based clinical measures, but needing further improvements in image quality.To conclude, DL is a powerful tool for accelerating existing workflows and facilitating new developments in CMR, but the process of developing DL methods involves randomness that can affect the reliability of comparisons. The usefulness of DL methods needs to be assessed on a case-by-case basis, but DL methods will likely continue to be a big part of future developments in CMR.