(An) MR image reconstruction method using recurrent neural network for domain transformation

Abstract

Image reconstruction in MRI is a key process that generates images from the measured MR signal. In the case of parallel imaging, image reconstruction is an ill-posed problem and it has traditionally been performed by deterministic algorithms. Recently due to the rapid development of the deep learning methods, it has become possible to solve the problem by using example-based deep learning methods. The proposed novel neural network architecture named ‘ETER-net’ is a unified solution to reconstruct MR image directly from k-space data in various k-space trajectories. In the proposed method, the proposed neural network consists of two parts, such as a domain transform network and a refinement network, to reconstruct the final image with high image quality. The domain transform network of the proposed method should conduct domain transforming from sensor domain to image domain. As examine the output features of recurrent layers, it is verified that the domain transform network is working as expected. To improve the network performance, additional SSIM loss function is applied. To demonstrate the practical application of the proposed method, k-space data acquired at a 3T MRI scanner with Cartesian and radial trajectories are used. As the proposed network is a unified solution for image reconstruction, it can be applied to different types of k-space trajectories, including the Cartesian, non-Cartesian, fully-sampled and under-sampled datasets. And it is more appropriate for reconstruction of MR images from multiple sets of k-space data acquired from multi-channel RF coil. The main advantage of the proposed method in comparison with the conventional method (which is consists of fully connected layers and convolutional layers) is that it dramatically decreases the number of parameters in the network. The tendency of blurry results was found again when only the Euclidean distance was used as the loss function, and it was verified again that it could be overcome through additional losses. The proposed method is an end-to-end MR image reconstruction method based on recurrent neural networks. It performs direct mapping of the input k-space data and the reconstructed images, operating as a unified solution that is applicable to various scanning trajectories.