Accelerated MRI sampling and reconstruction with reinforcement learning

Abstract

Magnetic Resonance Imaging stands as a cornerstone of modern clinical diagnostics due to its superior soft tissue visualization achieved without radiation exposure. However, the technology faces a persistent challenge. Scan durations remain prohibitively long, leading to patient discomfort, motion-induced image degradation, and reduced throughput in clinical facilities. This work tackles MRI acceleration through a dual-pronged approach, developing deep learning solutions for both reliable image reconstruction and intelligent data acquisition strategies. Conventional MRI workflows rely on predetermined sampling schemes that remain static regardless of individual patient anatomy or specific diagnostic requirements. Moreover, existing reconstruction algorithms frequently lack safeguards against generating artifacts that appear anatomically plausible but are fictitious. We propose an alternative paradigm. Our framework integrates reconstruction methods designed for robustness and generalization with acquisition policies that adapt sampling decisions based on evolving information during the scan process. The thesis develops solutions across three major research thrusts. The reconstruction component addresses fundamental reliability concerns through multiple complementary techniques. We begin with HyperCMR, which employs frequency-aware optimization to handle diverse cardiac imaging protocols. HierAdaptMR follows, introducing hierarchical adaptation mechanisms that enable deployment across different clinical sites without requiring institution-specific retraining. We then present a temporal memory architecture that treats reconstruction as a progressive refinement task, transitioning from artifact removal in early stages to detail recovery in later iterations. The Hallucination-Aware Network represents our approach to reconstruction reliability, utilizing k-space consistency analysis and progressive confidence scheduling to detect and suppress spurious structures while generating spatial uncertainty maps. These methods demonstrate effectiveness across multiple anatomies including cardiac, brain, and knee imaging, with strong performance in the MICCAI CMRxRecon Challenge series, achieving 3rd place in multi-disease assessment and 4th place in pediatric evaluation tasks. Building on these reconstruction advances, we address the acquisition problem through reinforcement learning applied to Cartesian sampling patterns. The framework incorporates several innovations. An automatic stopping mechanism learns when sufficient data has been collected for individual patients. Segmentation-guided policies direct sampling toward anatomically relevant regions through attention-based weighting. We balance reconstruction fidelity across spatial and frequency domains via adaptive metric combination. Finally, dual-domain policy architectures leverage both k-space measurements and image-space features to generate modality-appropriate sampling patterns. Experiments span cardiac datasets such as OCMR, ACDC, and CMRxRecon, alongside knee data from fastMRI and brain acquisitions from BraTS2021, demonstrating improved reconstruction and segmentation outcomes across acceleration rates ranging from 4× through 10×. The framework extends beyond Cartesian trajectories to radial acquisition geometries, which provide inherent advantages for cardiac imaging through motion resilience. A dual-branch architecture processes measurements and images through cross-attention mechanisms. Golden-ratio angular encoding ensures comprehensive k-space sampling. The reward formulation emphasizes both global reconstruction metrics and cardiac-specific anatomical accuracy. This represents an early demonstration of reinforcement learning applied to non-Cartesian MRI acquisition optimization. This research establishes methodologies spanning reconstruction trustworthiness, adaptive data collection under temporal constraints, and optimization aligned with clinical priorities. The progression moves from foundational reconstruction with uncertainty characterization, through decisions about acquisition termination, to selection of informative measurements using multi-domain representations, culminating in geometric extensions beyond standard Cartesian patterns. These contributions provide building blocks for next-generation MRI systems capable of substantial scan time reduction while maintaining reconstruction fidelity and diagnostic confidence.