[Paper Review] Data Consistent Deep Rigid MRI Motion Correction

Nalini Singh, Neel Dey+6 more|PubMed|Jan 25, 2023

Advanced MRI Techniques and ApplicationsMedicine27 references11 citations

TL;DR

A deep learning method that learns a motion-parameter dependent reconstruction network to produce data-consistent, motion-corrected MRI images, with test-time optimization over motion parameters only.

ABSTRACT

Motion artifacts are a pervasive problem in MRI, leading to misdiagnosis or mischaracterization in population-level imaging studies. Current retrospective rigid intra-slice motion correction techniques jointly optimize estimates of the image and the motion parameters. In this paper, we use a deep network to reduce the joint image-motion parameter search to a search over rigid motion parameters alone. Our network produces a reconstruction as a function of two inputs: corrupted k-space data and motion parameters. We train the network using simulated, motion-corrupted k-space data generated with known motion parameters. At test-time, we estimate unknown motion parameters by minimizing a data consistency loss between the motion parameters, the network-based image reconstruction given those parameters, and the acquired measurements. Intra-slice motion correction experiments on simulated and realistic 2D fast spin echo brain MRI achieve high reconstruction fidelity while providing the benefits of explicit data consistency optimization. Our code is publicly available at https://www.github.com/nalinimsingh/neuroMoCo.

Motivation & Objective

• Motivate and address motion artifacts in brain MRI, especially intra-slice motion in multi-shot acquisitions.
• Develop a retrospective motion correction method that avoids joint image-motion optimization.
• Introduce a hypernetwork-based reconstruction that maps motion parameters to image reconstructions.
• Enable test-time optimization over motion parameters with data-consistency monitoring to detect failures.

Proposed method

• Forward model: y = A(m)x + ε with A(m) varying by motion m across shots.
• Train f(y, m; θ) to map corrupted k-space data and motion parameters to a reconstruction, using simulated data with known motion.
• Use a hypernetwork h(·; θh) to generate reconstruction network weights θg(m) conditioning on motion m.
• Train by minimizing reconstruction loss L over simulated data, e.g., negative SSIM.
• At test time, freeze θh and optimize m by minimizing ||y − A(m) f(y, m; θ*)||² to obtain reconstruction f(y, m̂; θ*).
• Monitor the data-consistency loss to reject poor reconstructions.

Experimental results

Research questions

• RQ1 Can a neural network learn a motion-parameter dependent reconstruction to reduce joint optimization over image and motion parameters to optimization over motion parameters alone?
• RQ2 Does test-time optimization over motion parameters with data consistency yield reconstructions consistent with acquired k-space measurements in multi-shot MRI?
• RQ3 How well does the approach generalize from simulated data to realistic acquired k-space data?
• RQ4 Can automatic rejection criteria identify failed reconstructions based on data consistency?

Key findings

• The method yields reconstructions that are consistent with acquired measurements in simulated data with inter-shot motion.
• Reconstructions are sharper and closer in quality to motion-aware baselines that have access to ground-truth motion parameters.
• Motion parameter estimates are accurate for high-energy shots and the automated rejection identifies optimization failures.
• The approach generalizes to a realistic acquired k-space example despite training on simulated data.