Wave-Encoded Model-based Deep Learning for Highly Accelerated Imaging with Joint Reconstruction

Jaejin Cho, Borjan A. Gagoski, Tae Hyung Kim, Qiyuan Tian, Stephen Robert Frost, Itthi Chatnuntawech, Berkin BilgicarXiv 2022 · arXiv

Abstract

Magnetic Resonance Imaging (MRI) speed is a critical bottleneck in clinical practice. This work introduces wave-MoDL, a framework that integrates wave-encoding physics with an unrolled model-based deep learning (MoDL) network. By employing residual CNN constraints in both k-space and the image domain, and leveraging multi-contrast data similarity, wave-MoDL achieves up to 16-fold acceleration. Experimental results on whole-brain structural and quantitative imaging demonstrate high-fidelity reconstructions in under two minutes, significantly outperforming standard parallel imaging techniques.

Introduction

Magnetic Resonance Imaging (MRI) is inherently slow, which limits patient comfort, increases susceptibility to motion artifacts, and reduces clinical throughput. While techniques like parallel imaging (PI) and compressed sensing (CS) have significantly accelerated scan times, pushing these methods to extreme acceleration rates often results in severe noise amplification and aliasing artifacts.

Recently, deep learning has shown promise in image reconstruction, but pure data-driven models can hallucinate details or ignore underlying physics. On the acquisition side, "Wave-CAIPI" is a highly efficient physical scanning strategy that uses sinusoidal gradients to spread aliasing in three dimensions, making it easier to resolve. This paper tackles the challenge of achieving ultra-fast, high-fidelity 3D MRI by marrying this advanced physical acquisition strategy (wave-encoding) with a physics-constrained deep learning reconstruction framework.

Methods

The researchers propose wave-MoDL, which integrates three major innovations:

1. Wave-Encoding Physics

During the MRI readout, extra sinusoidal gradients are applied. This creates a "corkscrew" trajectory in k-space that better exploits the 3D spatial sensitivities of the receiver coils, spreading out aliasing artifacts so they are easier for algorithms to untangle.

2. Model-based Deep Learning (MoDL)

The framework uses an "unrolled" convolutional neural network (CNN) architecture. The network iteratively alternates between a CNN block (to denoise and extract features) and a data consistency (DC) block to ensure the image matches the actual physical measurements. Uniquely, residual CNN constraints are applied in both the image domain and k-space to capture diverse artifact features.

Overview of the wave-MoDL framework, illustrating the acquisition sequence (a), the unrolled network architecture with dual-domain denoisers (b), and the joint optimization objective.

3. Multi-Contrast Joint Reconstruction

For scans acquiring multiple image types simultaneously (e.g., multi-echo or quantitative mapping), a shifted sampling pattern (CAIPI) is applied across contrasts. The network is trained to jointly reconstruct these images, sharing complementary structural information to boost quality and reduce parameter count.

Results

The authors evaluated wave-MoDL using three 3D brain MRI datasets: MPRAGE (structural), Multi-echo MPRAGE, and 3D-QALAS (quantitative).

Performance and Noise Reduction

A key advantage of the wave-encoding strategy is the reduction of noise amplification, quantified by the g-factor. As shown in the comparison between standard SENSE and Wave-CAIPI, the wave-encoding approach significantly lowers both average and maximum g-factors across various sequences and acceleration rates.

Comparison of g-factor maps between SENSE and Wave-CAIPI across different sequences, demonstrating the physical advantage of wave-encoding in reducing noise amplification at high acceleration.

Key Metrics

  • 47-Second Structural MRI: On the MPRAGE dataset at 16-fold acceleration (R=4x4), wave-MoDL achieved a Normalized Root Mean Square Error (NRMSE) of 7.20%, a 1.92-fold improvement over standard wave-CAIPI.
  • 2-Minute Quantitative Mapping: For the 3D-QALAS dataset at 12-fold acceleration (R=4x3), wave-MoDL reduced NRMSE to 6.49%. This enables a 1:52-minute scan that accurately maps T1, T2, and PD values.
  • Multi-Contrast Synergy: Using the CAIPI sampling pattern across multi-echo data improved NRMSE by 0.81% while reducing network parameters by 5-fold compared to independent training.

Discussion and Conclusion

Wave-MoDL demonstrates that embedding physical forward models into deep learning architectures allows for extreme acceleration without sacrificing image fidelity. By leveraging dual-domain constraints and multi-contrast synergies, the method enables high-quality whole-brain imaging in under a minute.

Future work will address current memory constraints that require slice-group processing, which limits the exploitation of 3D structural information in the z-direction. Additionally, optimizing wave-gradient parameters for high-bandwidth sequences remains an area for further refinement to eliminate residual folding artifacts.

Summary updated

BibTeX
@misc{cho2022wave,
  author = {Jaejin Cho and Borjan A. Gagoski and Tae Hyung Kim and Qiyuan Tian and Stephen Robert Frost and Itthi Chatnuntawech and Berkin Bilgic},
  title = {{Wave-Encoded Model-based Deep Learning for Highly Accelerated Imaging with Joint Reconstruction}},
  year = {2022},
  eprint = {2202.02814},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2202.02814},
}
Figure