ACE-Net: AutofoCus-Enhanced Convolutional Network for Field Imperfection Estimation with application to high b-value spiral Diffusion MRI

Mengze Gao, Zachary Shah, Xiaozhi Cao, Nan Wang, Daniel Abraham, Kawin SetsompoparXiv 2024 · DOI · arXiv

Abstract

ACE-Net is a novel deep-learning framework designed to correct image artifacts in rapid MRI scans caused by magnetic field imperfections. By integrating traditional autofocus physics with unrolled neural networks, it accurately estimates complex spatiotemporal field variations without requiring lengthy calibration scans or specialized hardware. This approach significantly improves the clarity of high b-value spiral diffusion MRI.

Introduction

In magnetic resonance imaging (MRI), speed is often achieved at the cost of image quality. Rapid imaging techniques, such as spiral and Echo Planar Imaging (EPI), are highly sensitive to field imperfections—specifically $B_0$ inhomogeneity and eddy currents. These variations lead to significant blurring and geometric distortions, which are particularly problematic in high-resolution Diffusion MRI.

Traditionally, these issues are addressed through long calibration scans or expensive external field probes. However, these methods are brittle; patient movement or scanner heating can render calibrations obsolete. ACE-Net provides a data-driven solution to estimate and correct these imperfections on-the-fly using only the data acquired during the scan.

Methodology

Static $B_0$ Estimation

The first stage of ACE-Net focuses on estimating the static $B_0$ field from a non-diffusion-weighted image ($b=0$). Rather than relying solely on raw image data, the network is provided with "autofocus metrics"—a stack of images reconstructed at different frequency offsets. This physics-informed "hint" helps the network identify the frequencies that produce the sharpest images, preventing the AI from falling into incorrect local minima.

ACE-Net architecture for static B0 estimation using autofocus metrics.

Spatiotemporal Correction and Unrolled Architecture

For diffusion-weighted scans, where the magnetic field changes across both space and time, ACE-Net models these variations using 3rd-order spatial spherical harmonics and 3rd-order temporal polynomials.

To ensure the network's predictions remain physically grounded, the authors developed an unrolled network architecture. This framework includes Data Consistency (DC) blocks that utilize Conjugate Gradient Descent. By iteratively checking its work against the raw MRI physics, the network avoids "hallucinating" features and maintains mathematical consistency with the acquired signal.

Overview of the ACE-Net spatiotemporal field parameter estimation and unrolled reconstruction framework.

Datasets and Validation

The model was trained on synthetic data generated from 15 subjects, where clean images were corrupted with simulated field imperfections to provide a ground truth. Testing was conducted on real-world data from volunteers using a 3T scanner. To validate the results, the network’s field estimates were compared against measurements from a "Skope" external field probe, an industry-standard hardware solution for magnetic field mapping.

Key Results

  • Visual Sharpness: ACE-Net significantly reduced blurring in high b-value (b=3000) spiral diffusion images, recovering details that were previously smeared.
  • DTI Clarity: In Diffusion Tensor Imaging (DTI), the framework successfully recovered white matter fiber tracts that were nearly invisible in uncorrected images.
  • Hardware-Level Fidelity: The software-based field estimates closely matched the measurements from physical hardware probes, suggesting that ACE-Net can effectively replace expensive external sensors.
  • Robustness of Unrolling: The unrolled version of the network consistently outperformed standard CNN architectures in field parameter estimation accuracy.

Discussion and Future Work

ACE-Net demonstrates the power of combining deep learning with classical signal processing. By using autofocus metrics and unrolled architectures, the system provides a robust, calibration-free method for high-fidelity MRI reconstruction.

Future research will focus on validating the network across different scanner manufacturers and field strengths (1.5T and 7T). Additionally, the authors plan to extend the framework to EPI sequences and incorporate motion-robust features to better serve clinical populations who may have difficulty remaining still during scans.

Summary updated

BibTeX
@misc{gao2024ace,
  author = {Mengze Gao and Zachary Shah and Xiaozhi Cao and Nan Wang and Daniel Abraham and Kawin Setsompop},
  title = {{ACE-Net: AutofoCus-Enhanced Convolutional Network for Field Imperfection Estimation with application to high b-value spiral Diffusion MRI}},
  year = {2024},
  eprint = {2411.14630},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.14630},
}
Figure