Abstract
Magnetic Resonance Imaging (MRI) reconstruction is an ill-posed inverse problem. It is complicated by undersampling, noise, motion, and uncertainty in coil sensitivities. Recently, physics-guided and learning-based methods have advanced the field, but current systems remain vulnerable to distribution shift, calibration errors, and manual method selection in clinical workflows. This dissertation develops physics-guided, data-efficient reconstruction and Vision Language model (VLM) based Agentic artificial intelligence systems for automated, robust MRI under multiple degradations. First, it introduces untrained neural network (UNN) priors for PROPELLER and Cartesian MRI acquisitions. UNN suppresses artifacts without external training data. Second, it proposes a novel synthetic blade augmentation technique to strengthen unrolled deep networks for PROPELLER MRI. This method improves generalization across scanners and protocol variations. Third, an ensemble learning-based approach for accelerated and noise-resilient parallel MRI is introduced. It stabilizes joint image-sensitivity estimation under limited or imperfect calibration. Fourth, a retrospective motion correction strategy is developed that ensembles three generative AI models, each developed to address distinct motion types. Finally, this dissertation introduces two agentic AI frameworks using VLM to automate MRI inverse problems end-to-end. The first, AgentMRI is a single VLM-driven controller that uses multi-query, confidence-weighted consensus to identify degradation and dispatch the appropriate correction tool. Additionally, a hierarchical multi-agent framework is developed, where agents debate and reach a reliability-weighted decision. The agents report decision confidence and reduce operator intervention via tool use and structured reasoning. Comprehensive experiments on diverse MRI datasets evaluate image quality, robustness to shift, motion severity, and computation. The agents are assessed by degradation-classification accuracy, confidence calibration, and reduction in manual steps. Results demonstrate improved reconstruction fidelity at matched acceleration, enhanced resilience to motion, and practical gains from automated method selection. Collectively, these contributions advance physics-guided, learning-based inverse problems and establish agentic AI as a viable controller for self-regulating MRI reconstruction.