Registration and analysis of dynamic magnetic resonance image series

Cystic fibrosis (CF) is an autosomal-recessive inherited metabolic disorder that affects all organs in the human body. Patients affected with CF suffer particularly from chronic inflammation and obstruction of the airways. Through early detection, continuous monitoring methods, and new treatments, the life expectancy of patients with CF has been increased drastically in the last decades. However, continuous monitoring of the disease progression is essential for a successful treatment. The current state-of-the-art method for lung disease detection and monitoring is computed tomography (CT) or X-ray. These techniques are ill-suited for the monitoring of disease progressions because of the ionizing radiation the patient is exposed during the examination. Through the development of new magnetic resonance imaging (MRI) sequences and evaluation methods, MRI is able to measure physiological changes in the lungs. The process to create physiological maps, i.e. ventilation and perfusion maps, of the lungs using MRI can be split up into three parts: MR-acquisition, image registration, and image analysis. In this work, we present different methods for the image registration part and the image analysis part. We developed a graph-based registration method for 2D dynamic MR image series of the lungs in order to overcome the problem of sliding motion at organ boundaries. Furthermore, we developed a human-inspired learning-based registration method. Here, the registration is defined as a sequence of local transformations. The sequence-based approach combines the advantage of dense transformation models, i.e. large space of transformations, and the advantage of interpolating transformation models, i.e. smooth local transformations. We also developed a general registration framework called Autograd Image Registration Laboratory (AIRLab), which performs automatic calculation of the gradients for the registration process. This allows rapid prototyping and an easy implementation of existing registration algorithms. For the image analysis part, we developed a deep-learning approach based on gated recurrent units that are able to calculate ventilation maps with less than a third of the number of images of the current method. Automatic defect detection in the estimated MRI ventilation and perfusion maps is essential for the clinical routine to automatically evaluate the treatment progression. We developed a weakly supervised method that is able to infer a pixel-wise defect segmentation by using only a continuous global label during training. In this case, we directly use the lung clearance index (LCI) as a global weak label, without any further manual annotations. The LCI is a global measure to describe ventilation inhomogeneities of the lungs and is obtained by a multiple breath washout test.