Attention Incorporated Network for Sharing Low-rank, Image and K-space Information during MR Image Reconstruction to Achieve Single Breath-hold Cardiac Cine Imaging

TL;DR

This work addresses the challenge of accelerating cardiac Cine MRI by integrating information from multiple domains. It introduces A-LIKNet, a physics-based unrolled network with parallel image and k-space branches connected by a learnable information sharing layer, augmented with time-wise and coil-wise attention and a patch-based local low-rank prior. The approach yields superior reconstructions over conventional and state-of-the-art DL methods across retrospective and prospective data, maintaining high image quality up to $24\times$ undersampling and demonstrating potential for single breath-hold imaging. The findings highlight the value of multi-domain learning, tailored priors, and attention mechanisms in dynamic MRI reconstruction, with implications for faster, more comfortable clinical cardiac imaging.

Abstract

Cardiac Cine Magnetic Resonance Imaging (MRI) provides an accurate assessment of heart morphology and function in clinical practice. However, MRI requires long acquisition times, with recent deep learning-based methods showing great promise to accelerate imaging and enhance reconstruction quality. Existing networks exhibit some common limitations that constrain further acceleration possibilities, including single-domain learning, reliance on a single regularization term, and equal feature contribution. To address these limitations, we propose to embed information from multiple domains, including low-rank, image, and k-space, in a novel deep learning network for MRI reconstruction, which we denote as A-LIKNet. A-LIKNet adopts a parallel-branch structure, enabling independent learning in the k-space and image domain. Coupled information sharing layers realize the information exchange between domains. Furthermore, we introduce attention mechanisms into the network to assign greater weights to more critical coils or important temporal frames. Training and testing were conducted on an in-house dataset, including 91 cardiovascular patients and 38 healthy subjects scanned with 2D cardiac Cine using retrospective undersampling. Additionally, we evaluated A-LIKNet on the real-time 8x prospectively undersampled data from the OCMR dataset. The results demonstrate that our proposed A-LIKNet outperforms existing methods and provides high-quality reconstructions. The network can effectively reconstruct highly retrospectively undersampled dynamic MR images up to 24x accelerations, indicating its potential for single breath-hold imaging.

Figures (17)

• Figure 1: Proposed A-LIKNet: Attention-incorporated physics-based unrolled network for sharing low-rank, image, and k-space information during MR image reconstruction. The network consists of two branches: the bottom branch, featuring an image subnetwork, a learnable local low-rank operator, and an image data consistency layer for solving Eq. \ref{['eq:img_solution']}, and the upper branch, comprising a k-space subnetwork and a k-space data consistency layer to address Eq. \ref{['eq:kspace_solution']}. The information sharing layer facilitates communication between branches, enforcing consistency as in Eq. \ref{['eq:ISL_constrain']}.
• Figure 2: Image subnetwork with time-wise attention block. A residual 2D+t UNet with attention blocks in the decoder. The number of filters is denoted beside each layer. The attention block squeezes and excites spatial-channel features to generate a temporal attention map, assigning different weights to frames.
• Figure 3: K-space subnetwork with coil-wise attention blocks. A 3D three-layer convolutional neural network with attention blocks after each layer except the last one. The attention block squeezes and excites spatial-temporal features to generate an attention map along the MR coil dimension, assigning different weights to MR coils.
• Figure 4: Reconstructions in spatial (x-y) and spatial-temporal (y-t) plane of the proposed A-LIKNet for a retrospectively undersampled (VISTA) patient with transposition of the great vessels. Results for the acceleration factors R=6, 9, 12, 15, 18, 21, and 24 are shown in each column. The dynamic performance in the y-t plane corresponds to the blue line in the reference x-y plane image. The second row shows the enlarged views of the cardiac region (yellow box region). The third row presents the corresponding 5-times scaled absolute error maps.
• Figure 5: Reconstructions in spatial (x-y) plane of the proposed A-LIKNet for a retrospectively undersampled (VISTA) healthy subject. Results for every 3rd frame over one cardiac cycle are shown in each column. Both R=12 and R=24 reconstructions are shown alongside the corresponding absolute error maps. The enlarged views of the cardiac region are shown in Fig. \ref{['fig:Recon_result_cycle_cardiac']} in the supplementary material.