Moderní techniky rychlé rekonstrukce MR obrazů s využitím hlubokých neuronových sítí

Abstract

This master’s thesis deals with accelerated reconstruction of knee MR images from undersampled k-space using deep learning and reinforcement learning. The theoretical part summarizes the physical principles of magnetic resonance imaging, the role of k-space, and current approaches to MRI acceleration, especially parallel imaging, compressed sensing, and reconstruction methods based on neural networks. The practical part proposes a two-stage system in which a U-Net-based reconstruction network with a data consistency step generates an image from incomplete data, while a reinforcement learning agent adaptively selects previously unmeasured k-space columns. The agent state is formed by the current reconstructed image, the sampling mask, and the remaining acquisition budget, while the reward is defined by the increment of the SSIM metric. The system was trained and evaluated on the public fastMRI dataset, specifically on single-coil proton-density knee sequences acquired in sagittal and coronal planes.

On the test set, the proposed system achieved, for an acceleration factor of $R = 4$, a median SSIM of 0.8092, NMSE of 0.00640, and PSNR of 27.98\,dB; for $R = 6$, values of 0.7562, 0.00795, and 27.06\,dB; and for $R = 8$, values of 0.6878, 0.01320, and 24.81\,dB. A pilot clinical evaluation performed by three radiologists confirmed the diagnostic usability of all assessed reconstructions. In an additional comparison with conventional parallel imaging methods, the proposed approach showed higher image quality and fewer artifacts at comparable acceleration factors. The results indicate that the proposed system is most promising at low and moderate acceleration factors, whereas more aggressive undersampling leads to more pronounced image smoothing and loss of fine details. The thesis therefore confirms the potential of adaptive k-space sampling for further shortening MRI examinations while preserving the diagnostic usability of the reconstructed images.