Optimized K-space under-sampling for brain MRI reconstruction with reinforcement learning

Accelerating MRI acquisition while preserving diagnostic quality remains a fundamental challenge in medical imaging. In this work, we introduce a novel dual-domain reinforcement learning (RL) framework for optimizing k-space sampling patterns in accelerated brain MRI. Unlike conventional approaches, our method formulates the acquisition process as a sequential decision-making problem, leveraging reward signals from both the frequency (k-space) and spatial (image) domains to learn adaptive, modality-specific sampling strategies.We comprehensively evaluate our framework on the BraTS 2021 dataset across four MRI modalities: T1-weighted (T1), T2-weighted (T2), Fluid-Attenuated Inversion Recovery (FLAIR), and contrast-enhanced T1 (T1ce). Experimental results demonstrate that our approach consistently outperforms state-of-the-art learning-based and conventional sampling strategies in terms of both reconstruction quality and performance on downstream clinical tasks.Furthermore, task-based evaluation on brain tumor segmentation reveals that our optimized sampling patterns maintain segmentation accuracy close to that of fully-sampled acquisitions, with substantially reduced performance degradation compared to baseline methods. These findings demonstrate that dual-domain reinforcement learning offers a promising pathway toward clinically viable accelerated MRI. Code is available at https://github.com/Ruru-Xu/RL-Brain-MRI-Reconstruction