Master's Thesis work on 'An Enhanced Architecture for Accelerating Magnetic Resonance Imaging Based on Res-U-Net'
Abstract: This thesis proposes novel architectures to address the challenges associated with the reconstruction of undersampled MRI data using neural networks. The proposed approach involves designing and implementing a neural network, the Residual U-Net, along with two state-of-the-art preprocessing methods, Compressed Sensing and Parallel Imaging architectures, included to further improve the results obtained. Specifically, the GRAPPA (Generalized Autocalibrating Partially Parallel Acquisitions) and ESPIRiT (Eigenvector-based SPIRiT) algorithms will be applied to the MRI data to enhance the performance of the new architecture. The effectiveness of the proposed approach will be evaluated by training the new architecture on the NYU Health dataset in two distinct training instances, using two undersampling techniques, where the amount of data acquired is reduced by a factor of 4 (4x) and 8 (8x), respectively. The results obtained will be compared with those achieved by the teams that participated in the FastMRI challenge, using well-known metrics, such as SSIM, PSNR and NMSE. The purpose of this thesis is to investigate and analyze the potential of neural networks in enhancing the computational efficiency of MR imaging. The primary objective is to explore how neural networks can be leveraged to accelerate the entire scan process by a factor of four, thereby significantly reducing the total duration of the scan. The thesis contributes to the field of MRI reconstruction by providing insights into the effectiveness of the Res-U-Net and two specific preprocessing methods, GRAPPA and ESPIRiT, to improve the duration of Magnetic Resonance Imaging.