This thesis introduces novel deep learning techniques to improve signal acquisition and image reconstruction in MRI. MRI acquires data in the inverse image space (a.k.a k-space), and the first part addresses the limitations of scan-specific artificial neural networks for k-space interpolation when training data are scarce, a common issue in clinical routines. A novel iterative, self-consistent training data augmentation technique for complex-valued neural networks achieves a reduction in normalized mean squared error of 31.5% on average compared to standard approaches. This allows for the use of scan-specific deep learning without requiring changes to standard clinical protocols in 2D imaging. Furthermore, this study enhances the interpretability of these neural networks through a new image-space formalism. By introducing the concept of activation masks, the work translates nonlinear activations in k-space into a human-readable counterpart in the image space. This framework enables the analytical quantification of noise propagation and explains the origin of specific reconstruction artifacts (image blurring and “autocorrelation” center artifact). The study reveals that modulating the degree of nonlinearity in a model can act as a form of regularization, balancing reconstruction error against noise enhancement. Finally, the thesis presents an end-to-end learning approach to optimize MR sequences in self-learning MRI. Optimized variable flip angle schemes for 3D Fast Spin Echo imaging are identified that balance signal-to-noise ratio (SNR) and the point-spread function. In vivo evaluations at ultra-high field demonstrate that these optimized schemes substantially reduce image blurring, enhance visibility of anatomical structures with low baseline SNR, and eliminate “pseudo-lesions” in FLAIR imaging. It is shown that this physics-guided MR sequence learning is complementary to state-of-the-art image reconstruction using artificial neural networks.
Peter Dawood (Thu,) studied this question.