Continuous K-space Recovery Network with Image Guidance for Fast MRI Reconstruction

TL;DR

This work introduces IGKR-Net, a continuous k-space recovery network guided by image-domain information to accelerate MRI. By casting k-space recovery as an implicit neural representation problem and integrating a transformer-based encoder-decoder with an image-domain guidance module and a tri-attention refinement, the method progressively recovers dense k-space through a four-stage training strategy. The architecture combines Low- and High-Resolution Implicit Transformers (LRIT/HRIT), an Image Domain Guidance Module (IDGM), and a Tri-Attention Refinement Module (TARM) to achieve superior quantitative and qualitative results across single- and multi-coil datasets and varied undersampling masks. The approach demonstrates robust k-space reconstruction, preserves fine details, and offers improved efficiency, highlighting its potential for practical fast MRI with higher fidelity.

Abstract

Magnetic resonance imaging (MRI) is a crucial tool for clinical diagnosis while facing the challenge of long scanning time. To reduce the acquisition time, fast MRI reconstruction aims to restore high-quality images from the undersampled k-space. Existing methods typically train deep learning models to map the undersampled data to artifact-free MRI images. However, these studies often overlook the unique properties of k-space and directly apply general networks designed for image processing to k-space recovery, leaving the precise learning of k-space largely underexplored. In this work, we propose a continuous k-space recovery network from a new perspective of implicit neural representation with image domain guidance, which boosts the performance of MRI reconstruction. Specifically, (1) an implicit neural representation based encoder-decoder structure is customized to continuously query unsampled k-values. (2) an image guidance module is designed to mine the semantic information from the low-quality MRI images to further guide the k-space recovery. (3) a multi-stage training strategy is proposed to recover dense k-space progressively. Extensive experiments conducted on CC359, fastMRI, and IXI datasets demonstrate the effectiveness of our method and its superiority over other competitors.

Figures (8)

• Figure 1: Our main idea. (a) Previous studies typically employ general networks to reconstruct the undersampled data both in the k-space and image domain. (b) We customize a continuous k-space recovery network from a new perspective of implicit neural representation with image domain guidance, thereby enhancing the performance of MRI reconstruction.
• Figure 2: The architecture of the our IGKR-Net. Given the undersampled k-space, we first encode the coordinates $C_{s}$ and k-values $K_{s}$ of the sampled points, and the resulting features are added and then fed into LRIT. In LRIT, we use the LR coordinates to query and get the recovered results, i.e., $\widehat{{K}}_{1}$ and $\widehat{{I}}_{1}$. Next, we send $\widehat{{I}}_{1}$ into IDGM and get $\widehat{{K}}_{2}$ and $\widehat{{I}}_{2}$ with the guidance of the low-quality image $I_{s}$. Then, the k-values $K_{s}^{'}$ and coordinates $C_{s}^{'}$ of $\widehat{{K}}_{2}$ are encoded and fed into HRIT, where $\widehat{{K}}_{3}$ and $\widehat{{I}}_{3}$ are obtained by querying HR coordinates. Finally, we design TARM to refine $\widehat{{I}}_{3}$, yielding the final output $\widehat{{K}}_{4}$ and $\widehat{{I}}_{4}$.
• Figure 3: Our transformer based Encoder (a) and Decoder (b).
• Figure 4: (a) The proposed IDGM consists of two stages, i.e., shallow fusion stage and deep fusion stage. (b) The proposed TARM consists of three branches, i.e., PA, CA, and SA.
• Figure 5: Visual comparison of different methods under various undersampling masks. Red boxes illustrate the enlarged views in detail. Yellow ellipses highlight the performance differences in the results of anatomical structures among various methods.