Multi-Domain Neumann Network with Sensitivity Maps for Parallel MRI Reconstruction

Jun-Hyeok Lee, Junghwa Kang, Se-Hong Oh, Dong Hye Ye2022 · DOI

TL;DR

This paper introduces the Multi-Domain Neumann Network with Sensitivity Maps (MDNNSM), a deep learning approach for faster MRI reconstruction. By jointly learning coil sensitivity maps and reconstructing images in both image and frequency domains, MDNNSM significantly improves image quality and reduces artifacts compared to existing methods, especially at high acceleration factors.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI) is a powerful, non-invasive diagnostic tool, but its long scan times lead to patient discomfort and motion artifacts. Parallel MRI (pMRI) techniques aim to reduce scan times by undersampling k-space data, but this undersampling introduces aliasing artifacts. Reconstructing high-quality images from this undersampled data, especially at high acceleration factors (meaning more undersampling), remains a significant challenge. Traditional pMRI methods struggle to accurately remove these artifacts, and while deep learning has shown promise, accurately estimating coil sensitivity maps—crucial for the forward model in pMRI—is difficult with conventional methods, particularly at high acceleration factors.

Methods

The core of this work is the Multi-Domain Neumann Network with Sensitivity Maps (MDNNSM). It builds upon the Neumann network architecture, which is known for its ability to solve inverse problems by unrolling an optimization process. The MDNNSM introduces two key innovations:

  • CNN-based Sensitivity Map Estimation: Instead of relying on traditional methods like ESPIRIT (which can struggle with accuracy at high acceleration factors) or requiring additional scan time for full-coil sensitivity maps, MDNNSM uses a Convolutional Neural Network (CNN) to directly estimate coil sensitivity maps from the undersampled multi-coil k-space data. This integrated approach improves accuracy and efficiency.
  • Multi-Domain Regularization: The network reconstructs the MR image not only in the image domain but also in the frequency (k-space) domain. This "multi-domain" approach allows for more effective removal of aliasing artifacts and interpolation of missing data. The regularization block combines image-domain and frequency-domain CNNs.

The overall architecture includes a sensitivity map estimation block, a data consistency block (which enforces the forward model), and a regularization block. Skip connections are used to accumulate outputs from each iterative step, leading to a more refined final reconstruction.

Datasets

The study utilizes the NYU fastMRI Initiative dataset. This dataset contains T1-weighted brain MR images. The authors used 498 scans (7782 slices) for training and 1542 slices for validation, excluding data with ringing artifacts. They generated undersampled k-space data by multiplying fully sampled k-space data with equi-spaced sampling masks at acceleration factors of 2, 4, and 8.

Key Insights and Ideas

  • Joint Estimation is Key: The paper strongly suggests that jointly learning coil sensitivity maps and the MR image is superior to relying on traditional, separate estimation methods, especially when dealing with high acceleration factors.
  • Frequency Domain Matters: Reconstructing in both the image and frequency domains provides complementary information that helps resolve aliasing artifacts more effectively than image-domain-only reconstruction.
  • Model-Based Deep Learning Advantage: Incorporating the physics-based forward model directly into the deep learning architecture (as done in Neumann networks and MDNNSM) leads to more accurate and detailed reconstructions compared to purely data-driven approaches.
  • Robustness to Data Quantity: The MDNNSM shows robust performance even with fewer training subjects, indicating that its model-based approach is less dependent on massive datasets compared to some other deep learning methods.

Key Results

  • Quantitative Improvement: MDNNSM consistently achieved lower Normalized Mean Squared Error (NMSE) and higher Structural Similarity Index Measure (SSIM) scores across acceleration factors of 2, 4, and 8, compared to zero-filled, GRAPPA, U-Net, and the original Neumann network. For example, at an acceleration factor of 8, MDNNSM achieved an NMSE of 0.0051 and an SSIM of 0.9441, outperforming the original Neumann network (NMSE 0.0069, SSIM 0.9362).
  • Qualitative Improvement: Visual inspection of the reconstructed images showed that MDNNSM effectively removed aliasing artifacts and preserved fine details, producing images comparable to the fully sampled reference. This was particularly evident at higher acceleration factors where other methods struggled.
  • Memory Efficiency: By mapping multi-coil data into single data using estimated sensitivity maps, the method also improved memory efficiency.

Limitations and Future Work

  • Dataset Specificity: The study was conducted using only T1-weighted brain images from the fastMRI dataset. The authors acknowledge the need to evaluate MDNNSM on other MRI modalities (e.g., T2-weighted, FLAIR) and anatomical regions.
  • Sensitivity Map Accuracy: While the CNN-based estimation improved accuracy, the authors note that the accuracy of estimated sensitivity maps significantly impacts reconstruction performance. Further research into methods for estimating sensitivity maps with even higher accuracy is suggested.
  • Generalizability: While the paper shows robustness to the amount of training data, further investigation into the generalizability of the model across different scanners and acquisition parameters would be beneficial.

Summary updated

BibTeX
@misc{lee2022multi,
  author = {Jun-Hyeok Lee and Junghwa Kang and Se-Hong Oh and Dong Hye Ye},
  title = {{Multi-Domain Neumann Network with Sensitivity Maps for Parallel MRI Reconstruction}},
  year = {2022},
  doi = {10.3390/s22103943},
}
Figure