Thesis defense of Moritz Rempe

Magnetic Resonance Imaging (MRI) is a well-established imaging modality in modern medicine, yet the common practice of discarding raw k-Space data after image reconstruction represents significant underutilization of valuable information. This k-Space inherently contains richer information than the final magnitude image alone, including crucial phase data. The focus on image-domain processing means this additional information is frequently lost. We hypothesize, that with this loss of information, the potential for more efficient or deeper insights into diagnostic tasks is often not realized.
This thesis confronts this challenge by investigating hypotheses centered on the untapped potential of k-space. First, we hypothesize that it is advantageous to perform clinical tasks like segmentation and classification directly on complex-valued k-space data. Second, we propose that the bottleneck of data scarcity in raw MRI deep learning can be addressed by developing generative models that synthesize realistic MRI raw data.
To test these hypotheses, this thesis presents a cycle of interconnected projects. The first hypothesis is validated through "k-Strip", a novel complex-valued U-Net for direct k-Space segmentation. Applied to skull-stripping, it achieves competitive performance while enabling direct anonymization on non-human-readable data, while preserving valuable phase information. This is extended to clinical classification, where a model leveraging k-Space data for prostate cancer likelihood estimation demonstrates improved prediction accuracy, particularly under accelerated acquisition scenarios, promising faster and more robust diagnostic pathways.
To address data scarcity in raw MRI deep learning, a novel complex-valued generative diffusion model, "PhaseGen," was developed. PhaseGen synthesizes realistic MRI raw data, specifically generating phase information from available magnitude images, which enables the creation of large synthetic datasets for pretraining. Its effectiveness is shown by improving the generalization of the "k-Strip" model on real-world data and by enabling MRI reconstruction models to achieve performance comparable to full real-data training, even when augmented with synthetic data and having access to only 10-15% of the real data.
Finally, these methodologies are synergistically integrated into a publicly available de-identification tool. This software uniquely extends anonymization protocols to raw MRI data, encompassing both metadata cleansing and k-Space domain feature removal using the PhaseGen-trained k-Strip model, thus facilitating safer data sharing and collaborative research.
Collectively, by confirming these hypotheses, this work provides a foundational framework and practical tools for harnessing the full potential of MRI k-space through complex-valued deep learning, paving the way for more efficient and insightful medical imaging.