Link to paper

The full paper is available here.

You can also find the paper on PapersWithCode here.

Abstract

  • Motion artifacts can lead to incorrect diagnosis or characterization in imaging studies.
  • A deep network can reduce the joint image-motion parameter search to a search over rigid motion parameters.
  • The network is trained using simulated, motion-corrupted k-space data.
  • Experiments on simulated and realistic 2D fast spin echo brain MRI show high reconstruction fidelity.
  • Code is available online.

Paper Content

Introduction

  • Motion can corrupt MRI images and make them difficult to interpret
  • Prospective motion correction strategies can adjust the acquisition in real-time, but can be disruptive to clinical workflows
  • Retrospective strategies correct motion algorithmically after acquisition, without additional motion information
  • This paper develops a deep learning method for retrospective rigid motion correction
  • The method is trained on simulated pairs of corrupted and corrected examples and generalizes to a proof-of-concept acquired k-space example
  • The method produces reconstructions that are consistent with the acquired k-space measurements in the presence of inter-shot motion

Methods

  • Two stages to motion correction procedure: training a network and test-time optimization
  • MRI forward model described
  • Neural network training uses MRI forward model
  • Hypernetwork architecture used to generate weights of reconstruction subnetwork
  • Weights of reconstruction subnetwork never trained directly
  • Test-time optimization involves minimizing data consistency term and maximizing priors
  • Reconstruction network weights frozen at test-time
  • Loss monitored to determine success of motion correction procedure

Experiments

  • Data used is 2D T2 FLAIR FSE brain MRI k-space data from 3T GE Signa Premier
  • Dataset split into 553/197/100 training/validation/test 2D slices from 31/11/9 subjects
  • Motion simulated by applying ARC reconstruction, estimating coil sensitivity profiles, and synthesizing motion-corrupted measurements
  • Network architectures include hypernetwork and reconstruction subnetwork
  • Networks trained with SSIM loss function and Adam optimizer
  • Test-time optimization uses 4 trials of gradient descent
  • Baselines include ARC, Conv, Model-Based-GT, and HN-GT

Results

  • HN produces sharper, more accurate reconstructions than motion-naive methods
  • HN outperforms motion-naive methods across the test set
  • HN recovers accurate motion parameters for the first four high-energy shots
  • HN does not reliably recover motion parameters for the last two shots
  • HN successfully reconstructs motion-corrupted real k-space signals

Discussion and conclusions

  • Proposed deep rigid motion correction approach for multi-shot MRI
  • Evaluated on 2D FSE T2 FLAIR data
  • Learns mapping between corrupted k-space data, true motion parameters, and high-quality reconstructions
  • At test-time, only motion parameter estimates are optimized
  • Automatically rejects samples where optimization fails
  • Produces high-quality reconstructions in simulation and generalizes to realistic acquired k-space example