Magnetic Resonance Imaging (MRI) workflows typically discards raw k-Space data, wasting valuable phase information. This thesis challenges the "reconstruct-then-analyze" paradigm, demonstrating that processing complex-valued k-Space data directly significantly enhances diagnostic efficiency and accuracy. This thesis validates this approach through k-Strip, a neural network for direct k-Space segmentation and anonymization, and a prostate cancer classification model that improves predictive performance, particularly in accelerated scans. To address the scarcity of raw MRI data, the thesis introduces PhaseGen, a generative diffusion model. PhaseGen synthesizes realistic raw data, enabling robust training of algorithms even with minimal real-world datasets. These architectures are integrated into an open-source de-identification tool. By establishing the superiority of complex-valued neural networks for MRI signal analysis, this work lays the foundation for a new, data-efficient framework in medical imaging.
Moritz Rempe (Wed,) studied this question.