| dc.description.abstracteng | Magnetic resonance imaging (MRI) is an essential imaging modality in today's medical field
and allows the visualization of soft tissue without the use of ionizing radiation.
For cardiac magnetic resonance imaging (CMR), the inherent movement of the beating heart and the respiratory motion of the lungs,
which influences the spatial position of the heart, complicate the acquisition and reconstruction process.
Therefore, the clinically established method of cine CMR uses extended periods of breatholds and an ECG
to retrospectively reconstruct a synthetic cardiac cycle based on the monitored heartbeat.
For non-cooperative patients or those with arrhythmia, the imaging of the heart in real-time without the requirement of breath holding,
so called real-time free-breathing CMR, produces superior results.
However, real time CMR usually has a decreased image quality when compared to cine CMR of cooperative subjects with a regular heartbeat,
so that the segmentation of the heart, which is mandatory to assess its function, is more challenging.
In this thesis, the problem of the cardiac function analysis in real-time CMR is addressed by the use of deep learning methods
to increase examination speed, reproducibility, and precision.
It covers three studies, each focussing on the analysis and improvement of a different aspect of the problem:
the derivation of cardiac function parameters from neural network segmentations,
the generation of training data for supervised learning,
and the joint reconstruction and segmentation of CMR data.
The current status of deep learning methods for cardiac segmentation is evaluated by assessing the performance of two deep learning methods for
cardiac segmentation.
For this purpose, a curated dataset is created based on CMR measurements of healthy volunteers, which features cine images and real-time images
at rest and under exercise stress, manually corrected contours,
and intra- and inter-observer variability for cardiac function parameters derived from the segmentations.
Although both deep learning methods were designed for the segmentation of cine CMR, their performance on real-time CMR is comparable
to intra-observer variability and reported inter-observer variability of cine CMR.
Additionally, a novel method based on the self-gated MRI method SSA-FARY for the generation of training data for real-time free-breathing CMR is proposed and evaluated.
The method successfully creates a training dataset for supervised learning consisting of real-time images and corresponding segmentations.
Multi-task training has been shown to be a beneficial strategy for deep-learning methods, thus,
the question is investigated whether a neural network for the joint reconstruction and segmentation of real-time CMR
can lead to improvements in reconstruction or segmentation.
A joint network for the reconstruction and segmentation of real-time CMR is successfully implemented, trained, and applied,
and compared to different pipelines featuring the reconstruction and segmentation problem as two separate, subsequent tasks.
The network does not show clear benefits with further undersampled data as established in previous work
but has potential for the integration of a reconstruction network
which utilizes self-supervised learning and the temporal correlation of consecutive frames.
The thesis introduces several new datasets, tools, and methods as components for future work for the analysis of cardiac function in real-time free-breathing CMR. | de |