🥃 k-GIN: Global k-Space Interpolation for Dynamic MRI Reconstruction using Masked Image Modeling

Jiazhen Pan1, Suprosanna Shit1, Özgün Turgut1, Wenqi Huang1, Hongwei Bran Li2, Nil Stolt-Ansó1, Thomas Küstner3, Kerstin Hammernik1,4, Daniel Rueckert1,4
MICCAI 2023
1 Klinikum Rechts der Isar, Technical University of Munich, Germany
2 Department of Quantitative Biomedicine, University of Zurich, Switzerland
3 Medical Image And Data Analysis, University Hospital of Tübingen, Germany
4 Department of Computing, Imperial College London, UK
Paper Supplementary Code arXiv
🥃 After GRAPPA, RAKI, Caipirinha, SAKE etc., We have one more alcoholic beverage in MR Reconstruction 😉

Abstract

In dynamic Magnetic Resonance Imaging (MRI), k-space is typically undersampled due to limited scan time, resulting in aliasing artifacts in the image domain. Hence, dynamic MR reconstruction requires not only modeling spatial frequency components in the x and y directions of k-space but also considering temporal redundancy. Most previous works rely on image-domain regularizers (priors) to conduct MR reconstruction. In contrast, we focus on interpolating the undersampled k-space before obtaining images with Fourier transform. In this work, we connect masked image modeling with k-space interpolation and propose a novel Transformer-based k-space Global Interpolation Network, termed k-GIN. Our k-GIN learns global dependencies among low- and high-frequency components of 2D+t k-space and uses it to interpolate unsampled data. Further, we propose a novel k-space Iterative Refinement Module (k-IRM) to enhance the high-frequency components learning. We evaluate our approach on 92 in-house 2D+t cardiac MR subjects and compare it to MR reconstruction methods with image-domain regularizers. Experiments show that our proposed k-space interpolation method quantitatively and qualitatively outperforms baseline methods. Importantly, the proposed approach achieves substantially higher robustness and generalizability in cases of highly-undersampled MR data.

Summary

  1. Instead of conducting MR reconstruction in the image-domain with denoising priors, the proposed work carried out an interpolation in k-space before the Fourier transform.
  2. In contrast to conventional k-space interpolation methods which utilized local kernels to estimate the missing data, the proposed work applied global dependencies modeling in k-space via Transformers to exploit the k-space redundancies completely.
  3. We leverage Masked Autoencoders (MAE) to accomplish the interpolation, in which the missing k-space data is estimated from the learned rich representation of the sampled k-space data.

Results

Qualitative comparison of the proposed method with TV-Optim, L+S, and DcCNN in the R = 4 and 8 undersampled data. It is to note that k-GIN and DcCNN are both trained only on R = 4 data, while the R = 8 data is unseen for them during the inference. k-GIN demonstrates higher robustness and generalizability because the task of interpolating k-space for R=4 and R=8 remains the same, i.e., to estimate missing data from sampled data, while there is a mismatch in artifact characteristics in image domain between R=4 and R=8 for DcCNN which conducts an image-domain denoising.

Video Presentation

Poster

Limitations and Outlook

There are some limitations in the current version of the work. First, this work is based on single-coil settings, we will study the multi-coils settings in our future work. Further, the input size to the network is fixed in this work, we will enable the flexible input size of the k-space in the future. Last but not least, this work can be regarded as a "pre-training" task since the image reconstruction itself is an "intermediate step" for downstream tasks e.g. cardiac segmentation and disease classification. In the future, one can reuse the learned encoder representation of k-GIN to directly solve downstream tasks without requiring image reconstruction.

BibTeX

@article{pan2023global,
        title={Global k-Space Interpolation for Dynamic MRI Reconstruction using Masked Image Modeling},
        author={Pan, Jiazhen and Shit, Suprosanna and Turgut, {\"O}zg{\"u}n and Huang, Wenqi and Li, Hongwei Bran and Stolt-Ans{\'o}, Nil and K{\"u}stner, Thomas and Hammernik, Kerstin and Rueckert, Daniel},
        journal={Medical Image Computing and Computer Assisted Intervention -- MICCAI 2023},
        year={2023}
      }