Extending LOUPE for K-Space Under-Sampling Pattern Optimization in Multi-coil MRI

Jinwei Zhang, Hang Zhang, Alan Wang, Qihao Zhang, Mert R. Sabuncu, Pascal Spincemaille, Thanh D. Nguyen, Yi Wang

Introduction

Magnetic Resonance Imaging (MRI) is a powerful diagnostic tool, but its utility is often limited by long acquisition times. Data is collected in "k-space" (the spatial frequency domain) line by line. To accelerate this process, clinicians employ under-sampling—skipping portions of k-space—and then use reconstruction algorithms to recover the full image. The challenge lies in determining the optimal under-sampling pattern. While traditional "variable density" (VD) patterns are common, they are not necessarily optimized for specific hardware or scan types.

This research introduces an upgraded version of the LOUPE (Learning-based Optimization of the Under-sampling Pattern) framework, specifically designed to handle real-world, multi-coil MRI data. By integrating "unrolled" optimization networks with a novel gradient estimation technique for binary sampling, the authors demonstrate that learned sampling patterns can significantly outperform traditional methods in both speed and image fidelity.

Methodology

The proposed framework addresses three primary limitations of the original LOUPE: the reliance on single-channel simulated data, the use of "fuzzy" probabilistic masks during training, and the use of physics-agnostic U-Net architectures.

Architecture of the proposed framework, featuring a K-rolled reconstruction network (a) and a sampling pattern learning network (b).

Key Innovations:

  1. Multi-coil Integration: The model utilizes raw, fully sampled multi-coil k-space data from a 3T scanner, accounting for the complex interactions in Parallel Imaging.
  2. Straight-Through (ST) Estimator: To enable the learning of binary (yes/no) sampling masks, the authors employed an ST estimator. This allows gradients to flow through the non-differentiable binary thresholding operation during backpropagation.
  3. Model-Based Deep Learning (MoDL): Instead of a standard U-Net, the framework uses an "unrolled" MoDL architecture. This ensures data consistency by embedding MRI physics (such as coil sensitivity maps) directly into the network, making it more robust and data-efficient.

Experimental Setup

The study utilized a high-quality in-house dataset consisting of T2-weighted 3D fast spin echo sagittal brain scans from 6 healthy volunteers. Data was acquired on a 3T GE scanner using a 32-channel head coil. The dataset comprised 600 total slices, split into training (300), validation (100), and testing (200) sets. Fully sampled k-space data served as the ground truth for performance evaluation.

Results and Discussion

The combination of MoDL and Binary Sampling (BS) yielded the highest performance across all metrics. Quantitatively, the MoDL + BS approach achieved a PSNR of 42.6 dB and an SSIM of 0.968, significantly outperforming U-Net-based alternatives which averaged around 32-33 dB PSNR.

Visual comparison and error maps for different reconstruction methods. MoDL + BS shows the lowest error and highest structural fidelity.

Visually, the error maps reveal that learned patterns produce sharper structural details and lower global error compared to standard VD patterns. Notably, the "learned" patterns proved universally beneficial, improving the performance of even conventional non-AI reconstruction methods like ESPIRiT and TGV. This suggests that the network learns fundamental properties of the k-space distribution that are broadly applicable.

Conclusion

By bridging the gap between deep learning theory and clinical MRI hardware, this work demonstrates the potential for AI-driven scan optimization. The shift to physics-informed, unrolled networks and binary mask learning provides a robust path toward "scanner-ready" AI. Future research will focus on prospective implementation—testing these patterns in real-time on scanners—and evaluating their generalizability across different anatomical regions such as the knee or heart.

Summary updated

BibTeX
@inproceedings{zhang2020extending,
  author = {Jinwei Zhang and Hang Zhang and Alan Wang and Qihao Zhang and Mert R. Sabuncu and Pascal Spincemaille and Thanh D. Nguyen and Yi Wang},
  title = {{Extending LOUPE for K-Space Under-Sampling Pattern Optimization in Multi-coil MRI}},
  booktitle = {MLMIR @ MICCAI},
  year = {2020},
  doi = {10.1007/978-3-030-61598-7\_9},
}
Figure