[Paper Review] k-Space Deep Learning for Parallel MRI: Application to Time-Resolved MR Angiography

Eunju Cha, Eung Yeop Kim+1 more|arXiv (Cornell University)|Jun 3, 2018

Advanced MRI Techniques and Applications21 references4 citations

TL;DR

This paper proposes a k-space deep learning framework for parallel MRI that enables real-time, flexible reconstruction of time-resolved MR angiography (tMRA) with improved temporal resolution. By learning multi-coil k-space interpolation kernels through a deep neural network inspired by Hankel matrix decomposition, the method achieves high-fidelity reconstructions across varying view-sharing levels without retraining, outperforming GRAPPA and ALOHA in speed and quality.

ABSTRACT

Time-resolved angiography with interleaved stochastic trajectories (TWIST) has been widely used for dynamic contrast enhanced MRI (DCE-MRI). To achieve highly accelerated acquisitions, TWIST combines the periphery of the k-space data from several adjacent frames to reconstruct one temporal frame. However, this view-sharing scheme limits the true temporal resolution of TWIST. Moreover, the k-space sampling patterns have been specially designed for a specific generalized autocalibrating partial parallel acquisition (GRAPPA) factor so that it is not possible to reduce the number of view-sharing once the k-data is acquired. To address these issues, this paper proposes a novel k-space deep learning approach for parallel MRI. In particular, we have designed our neural network so that accurate k-space interpolations are performed simultaneously for multiple coils by exploiting the redundancies along the coils and images. Reconstruction results using in vivo TWIST data set confirm that the proposed method can immediately generate high-quality reconstruction results with various choices of view- sharing, allowing us to exploit the trade-off between spatial and temporal resolution in time-resolved MR angiography.

Motivation & Objective

• To overcome the limited true temporal resolution in TWIST imaging caused by view-sharing across multiple frames.
• To address the inflexibility of existing GRAPPA-based TWIST protocols that require fixed acceleration factors and cannot be adjusted post-acquisition.
• To develop a deep learning-based k-space reconstruction method that enables immediate reconstruction at various spatial-temporal resolution trade-offs without retraining.
• To achieve computational efficiency and high image quality by leveraging deep neural networks to learn k-space interpolation kernels from multi-coil redundancy and sparsity.
• To provide a backward-compatible solution that works with existing TWIST acquisition protocols without pulse sequence modifications.

Proposed method

• The method employs a deep convolutional neural network designed as a multi-layer extension of data-driven Hankel matrix decomposition to model k-space interpolation.
• The network learns interpolation kernels directly in k-space, exploiting redundancy across coils and sparsity in the image domain.
• It uses an encoder-decoder architecture with recursive lifting and un-lifting operations to map low-rank k-space structures into higher-dimensional spaces for improved interpolation.
• The network is trained end-to-end on weighted k-space data, where low-rankness of the Hankel matrix is exploited as a prior for effective interpolation.
• During inference, the same trained network generates reconstructions for any number of view-sharing levels by simply adjusting the input window size.
• The approach avoids on-the-fly kernel computation, making it significantly faster than GRAPPA and ALOHA, with inference times of only 0.029 seconds per slice.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on k-space data generalize across different view-sharing configurations in TWIST without retraining?
• RQ2 How does k-space deep learning compare to GRAPPA and ALOHA in terms of reconstruction quality and computational efficiency for time-resolved angiography?
• RQ3 Can a single deep neural network learn interpolation kernels that are robust to varying acceleration factors and temporal resolution requirements?
• RQ4 Does the proposed method preserve spatial resolution at high acceleration factors where conventional GRAPPA fails?
• RQ5 Can the learned k-space interpolation be applied to unseen time frames with higher temporal resolution than the training data?

Key findings

• The proposed method achieves reconstruction times of only 0.029 seconds per slice, which is over 100 times faster than GRAPPA (6.09 seconds) and ALOHA (84.61 seconds).
• The method enables immediate reconstruction at any number of view-sharing levels using a single trained network, allowing flexible exploration of spatial-temporal resolution trade-offs.
• The network generalizes well to unseen temporal frames due to its learning of intrinsic k-space structure via low-rank Hankel matrix decomposition.
• At high acceleration factors, the proposed method significantly outperforms ALOHA in spatial resolution, especially when view-sharing is reduced.
• The deep k-space network achieves superior image quality compared to GRAPPA and ALOHA by exploiting deeper representational learning through multi-layered lifting and coding.