[Paper Review] k-t NEXT: Dynamic MR Image Reconstruction Exploiting Spatio-temporal Correlations

Chen Qin, Jo Schlemper+5 more|arXiv (Cornell University)|Jul 22, 2019

Advanced MRI Techniques and Applications18 references4 citations

TL;DR

This paper proposes k-t NEXT, a deep learning-based method for dynamic MR image reconstruction that iteratively alternates between image space and x-f (spatial-frequency) domain to exploit spatio-temporal correlations. By using an xf-CNN to reconstruct aliased k-t data and a CRNN to model temporal dynamics, the method achieves superior image quality over state-of-the-art approaches at high undersampling rates, with PSNR gains up to 34.23 dB and HFEN reductions to 0.196.

ABSTRACT

Dynamic magnetic resonance imaging (MRI) exhibits high correlations in k-space and time. In order to accelerate the dynamic MR imaging and to exploit k-t correlations from highly undersampled data, here we propose a novel deep learning based approach for dynamic MR image reconstruction, termed k-t NEXT (k-t NEtwork with X-f Transform). In particular, inspired by traditional methods such as k-t BLAST and k-t FOCUSS, we propose to reconstruct the true signals from aliased signals in x-f domain to exploit the spatio-temporal redundancies. Building on that, the proposed method then learns to recover the signals by alternating the reconstruction process between the x-f space and image space in an iterative fashion. This enables the network to effectively capture useful information and jointly exploit spatio-temporal correlations from both complementary domains. Experiments conducted on highly undersampled short-axis cardiac cine MRI scans demonstrate that our proposed method outperforms the current state-of-the-art dynamic MR reconstruction approaches both quantitatively and qualitatively.

Motivation & Objective

• Address the challenge of highly accelerated dynamic MRI reconstruction with limited k-t space sampling.
• Exploit spatio-temporal redundancies in both k-t space and image domains to improve reconstruction fidelity.
• Overcome limitations of existing deep learning methods that treat frames independently or rely only on image-domain modeling.
• Develop a joint learning framework that alternates between x-f and image spaces to capture complementary signal structures.
• Achieve state-of-the-art performance in dynamic cardiac cine MRI reconstruction under high undersampling factors.

Proposed method

• Propose a novel iterative reconstruction framework, k-t NEXT, that alternates between the x-f domain and image domain.
• Use an xf-CNN to reconstruct true signals from aliased k-t data in the x-f domain, leveraging sparsity and redundancy.
• Integrate a convolutional recurrent neural network (CRNN) to model temporal dependencies in the image domain.
• Formulate the reconstruction as an alternating optimization process: xf-CNN updates in x-f space, CRNN updates in image space.
• Train the network end-to-end with joint optimization, using residual learning to predict image updates.
• Apply a 3D convolutional layer with data sharing and residual connections to enhance feature learning in the CRNN module.

Experimental results

Research questions

• RQ1 Can joint modeling of spatio-temporal correlations in both x-f and image domains improve dynamic MRI reconstruction beyond single-domain methods?
• RQ2 How does iterative alternation between x-f and image space reconstruction enhance image quality compared to end-to-end single-domain learning?
• RQ3 To what extent does the xf-CNN component reduce aliasing artifacts by directly reconstructing in the x-f domain?
• RQ4 How does k-t NEXT compare quantitatively and qualitatively to state-of-the-art methods like k-t FOCUSS, CRNN-MRI, and DS+3DCNN at high undersampling rates?
• RQ5 Does the proposed method preserve fine temporal and spatial details better than methods relying on temporal averaging or interpolation?

Key findings

• k-t NEXT achieves a PSNR of 34.23 dB at 9× undersampling, significantly outperforming the next best method (DS+CRNN at 33.24 dB).
• At 12× undersampling, k-t NEXT achieves a PSNR of 33.18 dB, surpassing CRNN-MRI (31.30 dB) and DS+3DCNN (32.46 dB).
• The method achieves the lowest HFEN value of 0.196 at 9× undersampling, indicating superior preservation of high-frequency image details.
• SSIM values for k-t NEXT reach 0.979 at 9× undersampling, exceeding all baselines and indicating high structural similarity to ground truth.
• Visual comparisons show that k-t NEXT produces sharper images with fewer artifacts, especially in dynamic regions like the left ventricle.
• The x-f domain visualization confirms that aliasing artifacts are effectively suppressed, with the reconstructed x-f image closely matching the ground truth.