Deep learning methods for accelerated MRI reconstruction: Optimization-inspired unrolled networks and beyond

Magnetic Resonance Imaging (MRI) provides superior soft-tissue contrast without ionizing radiation, but long acquisition times limit clinical throughput and increase motion artifacts. Acquiring fewer measurements reduces scan time at the cost of making image reconstruction an ill-posed inverse problem that requires strong prior knowledge. This dissertation develops optimization-inspired deep unrolled networks for MRI reconstruction. Starting from Half-Quadratic Splitting unrolled into a trainable network (HQS-Net), we introduce prompt-based conditioning that allows a single model to reconstruct across diverse acquisition settings (PromptMR). We then rethink the role of the CNN: instead of restricting it to a denoising role within the unrolled optimization, we let it learn the full reconstruction update, achieving 9 x faster convergence while reducing GPU memory to 55\% of the baseline (PromptMR+). Scaling to 32 cascades introduces instability inherent to the bilevel optimization structure of deep unrolling. We address this by adapting the optimizer to the cascade-wise structure of the network (Muon-Unroll). Beyond paired supervised training, we develop a diffusion-based generative prior for plug-and-play reconstruction and implicit neural representations for self-supervised per-scan reconstruction without any training data.