A novel neural network architecture named 'ETER-net' is proposed as a unified solution to reconstruct an MR image directly from k-space data acquired with various k-space trajectories.
The image reconstruction in MRI is performed by transforming the k-space data into image domain, where the domain transformation can be executed by Fourier transform when the k-space data satisfy the Shannon sampling theory. We propose an RNN-based architecture to achieve domain transformation and de-aliasing from undersampled k-space data in Cartesian and non-Cartesian coordinates. An additional CNN-based network and loss functions including adversarial, perceptual, and SSIM losses are proposed to refine and optimize the network performance.
This page is for validation of our method to a public dataset called 'FastMRI'. [link]
paper link :
trained weight, input data, and label: [link]
