This is the coding part of the project ELEC70121 Trustworthy Artificial Intelligence in Medical Imaging. Includes the final report as well.
Title: Deep Q-Networks in Adaptive k-space Sampling for Precision MRI
This work presents a novel Deep Reinforcement Learning approach for optimizing k-space vertical line selection in undersampled Magnetic Resonance Imaging (MRI). By training a Deep Q-Network (DQN) agent to directly learn a sampling policy from reconstruction quality feedback (PSNR and SSIM), we demonstrate its ability to discover strategies superior to common heuristics, especially in high uncertainty regimes. Experimental results demonstrate significant quantitative gains, suggesting that actively tailoring sampling patterns to specific image content and the reconstruction process through RL can substantially improve accelerated MRI performance within a fixed acquisition budget. This work highlights the potential of reinforcement learning for precise image reconstruction within the MRI workflow, opening promising avenues for further improvement of clinical MRI technologies.
If you use code or ideas, please cite this repository
@misc{kaliutau2025mri,
title={Deep {Q-Networks} in Adaptive k-space Sampling for Precision {MRI}},
author={Kaliutau, Aliaksei},
publisher = {GitHub},
journal = {GitHub repository},
url = {https://github.com/akaliutau/taimi-dqn},
year={2025}
}