DEEP LEARNING TECHNIQUES FOR KNEE MR IMAGES RECONSTRUCTION

MRI scanners acquire data samples in the spatial frequency domain (k-space) and classical or deep learning imaging techniques are applied to reconstruct the final image. A limitation of using classic reconstruction techniques is that it requires to acquire the full set of Fourier domain data. Recent deep learning techniques use subsampled Fourier data to produce diagnostic images. The aim of this project is to maximize the accuracy of the reconstructed images while minimizing the amount of data needed, the size of the model, and the training time. We used the Knee MR images from NYU fastMRI database to perform advance machine learning algorithms to reconstruct MR images. The quality assurance of these images was evaluated by NMSE, PSNR, SSIM and by expert opinions. A: We varied the size of fast MRI´s UNet B. We modified the UNet to use loss functions other than L1-loss. C: We implemented smaller variations of Resnet, CS-Net, DCCNN, CDCNN. D: We considered the expert evaluation for the feedback of the final outputs of the tuned neural networks. The CS-Net was the most performant. The CS-Net might have been judged the best bacause tempts to emulate a traditional MRI reconstruction method.