Self-supervised k-Space Regularization for Motion-Resolved Abdominal MRI Using Neural Implicit k-Space Representations

Veronika Spieker, Hannah Eichhorn, Jonathan K. Stelter, Wenqi Huang, Rickmer F. Braren, Daniel Rueckert, Francisco Sahli Costabal, Kerstin Hammernik, Claudia Prieto, Dimitrios C. Karampinos, Julia A. SchnabelMICCAI 2024 · DOI

Introduction

Abdominal MRI is a critical diagnostic tool, yet it remains one of the most challenging applications of magnetic resonance imaging due to respiratory motion. To mitigate motion artifacts, researchers typically employ "motion-resolved" techniques that bin data into different respiratory states. However, this partitioning of data leads to severe undersampling, where each bin lacks sufficient data to satisfy the Nyquist theorem, resulting in significant aliasing and noise.

Recently, Neural Implicit k-Space Representations (NIK) have emerged as a powerful alternative to traditional grid-based reconstructions. NIK models the MRI signal as a continuous mapping of coordinates $(k_x, k_y, t)$ to signal values using a neural network. While NIK offers high temporal resolution and flexibility, it faces a fundamental hurdle: regularization. Standard MRI regularization techniques, such as Total Variation, are defined in the image domain. Because NIK operates entirely in the frequency domain (k-space), transforming between domains during training is computationally prohibitive. Consequently, vanilla NIK models often suffer from overfitting or excessive noise.

The PISCO Method

To address these limitations, this paper introduces PISCO (Parallel Imaging-inspired Self-Consistency), a novel self-supervised regularization framework. PISCO bridges the gap between classical signal processing and implicit neural representations by reformulating the principles of GRAPPA (GeneRalized Autocalibrating Partially Parallel Acquisitions).

From GRAPPA to Self-Consistency

Traditional GRAPPA assumes that any given point in k-space can be reconstructed as a linear combination of its neighbors across multiple receiver coils. This relationship is typically "calibrated" using a small region of fully sampled data known as the Auto-Calibration Signal (ACS).

PISCO evolves this concept into a self-supervised loss function. The authors hypothesize that if the linear relationship between k-space neighbors is a global property of the signal, it must remain consistent across different random subsets of the data. During the NIK training process, PISCO samples multiple patches of k-space and solves for the weights ($W$) that relate points to their neighborhoods. It then enforces a Self-Consistency Loss ($L_{PISCO}$), which penalizes discrepancies between the weight matrices derived from different data subsets.

Advantages of the Implicit Approach

Because NIK can be queried at any arbitrary coordinate, PISCO can enforce consistency even at locations that were not physically sampled during the scan. This is a significant advantage for radial trajectories, where large gaps often exist in the periphery of k-space. Furthermore, by sorting coordinates into temporally adjacent subsets, PISCO effectively regularizes the reconstruction across both spatial and temporal dimensions, ensuring smooth transitions between respiratory phases.

Experimental Validation

The researchers evaluated the PISCO-NIK framework using three distinct datasets:

  1. XCAT Phantom: A digital simulation used to provide "ground truth" for quantitative metrics like Peak Signal-to-Noise Ratio (PSNR).
  2. In-vivo Abdominal Data: Real-world scans acquired at 3T using a golden-angle stack-of-stars trajectory.
  3. Static Thigh Data: A control dataset used to verify the regularization's efficacy in the absence of motion.

Key Results and Discussion

The integration of PISCO into the NIK pipeline yielded substantial improvements over both "vanilla" NIK and the current state-of-the-art method, XD-GRASP:

  • Quantitative Performance: In XCAT simulations, PISCO-NIK achieved a PSNR increase of up to 1.1 dB and demonstrated superior spatial feature similarity (FSIM).
  • Image Sharpness: In-vivo results showed that PISCO-NIK effectively denoised images while preserving the sharpness of small vascular structures that were often blurred or lost in competing reconstructions.
  • Artifact Suppression: The method successfully mitigated the "ringing" and streaking artifacts that typically plague highly accelerated radial MRI acquisitions.

Crucially, PISCO achieves these results without requiring a dedicated ACS or performing expensive Fourier Transforms during the optimization loop, making it a highly efficient regularization strategy for implicit models.

Limitations and Future Outlook

Despite its success, the authors identify several areas for future development. Like most implicit neural representations, NIK requires a lengthy per-scan training time, which currently limits its use in real-time clinical settings. Additionally, the reconstruction quality remains dependent on the accuracy of the respiratory navigator signal used for binning. Future work will explore adaptive kernel designs to better match specific sampling patterns and investigate the application of PISCO to other dynamic contexts, such as cardiac imaging.

Summary updated

BibTeX
@inproceedings{spieker2024self,
  author = {Veronika Spieker and Hannah Eichhorn and Jonathan K. Stelter and Wenqi Huang and Rickmer F. Braren and Daniel Rueckert and Francisco Sahli Costabal and Kerstin Hammernik and Claudia Prieto and Dimitrios C. Karampinos and Julia A. Schnabel},
  title = {{Self-supervised k-Space Regularization for Motion-Resolved Abdominal MRI Using Neural Implicit k-Space Representations}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2024},
  doi = {10.1007/978-3-031-72104-5\_59},
}
Figure