What we do sample, we must learn to reconstruct From missing k-space data to meaningful images: Deep learning in MRI reconstruction and beyond

Magnetic Resonance Imaging (MRI) is a powerful imaging modality, yet its inherently long acquisition times remain a fundamental limitation. Accelerating MRI requires subsampling k-space measurements, but this violates the Nyquist–Shannon sampling criterion and renders image reconstruction an ill-posed inverse problem. This thesis addresses this challenge by developing deep learning–based methods for accelerated multi-coil MRI reconstruction that integrate data-driven priors with the structure of classical optimization algorithms. Building on principles from Parallel Imaging and Compressed Sensing, the work introduces recurrent unrolled variational networks operating directly in k-space, ADMM-based variable-splitting reconstruction frameworks adaptable to static and dynamic imaging, and hybrid supervised and self-supervised training strategies that leverage proxy fully sampled datasets when target-domain ground truth is unavailable. Beyond reconstruction alone, the thesis investigates the interaction between sampling trajectories and deep models, proposes end-to-end adaptive dynamic subsampling strategies jointly optimized with reconstruction and deformable registration, and develops unified reconstruction models capable of generalizing across anatomies, contrasts, and acceleration factors, including zero-shot adaptation to unseen domains. To facilitate reproducible research and translation, an open-source reconstruction toolkit is presented. Finally, the methodological advances are evaluated in a multi-center, multi-reader clinical study in prostate MRI, demonstrating that AI-driven reconstruction can enable substantial acceleration while maintaining diagnostic performance. Collectively, this work presents a coherent framework for AI-accelerated MRI that spans acquisition design, inverse problem modeling, learning under limited supervision, cross-domain generalization, and clinical validation.