MAMOC: MRI Motion Correction via Masked Autoencoding

TL;DR

This work is the first to evaluate motion correction in MRI scans using real motion data on a public dataset, showing that MAMOC achieves improved performance over existing motion correction methods.

Abstract

The presence of motion artifacts in magnetic resonance imaging (MRI) scans poses a significant challenge, where even minor patient movements can lead to artifacts that may compromise the scan's utility.This paper introduces MAsked MOtion Correction (MAMOC), a novel method designed to address the issue of Retrospective Artifact Correction (RAC) in motion-affected MRI brain scans. MAMOC uses masked autoencoding self-supervision, transfer learning and test-time prediction to efficiently remove motion artifacts, producing high-fidelity, native-resolution scans. Until recently, realistic, openly available paired artifact presentations for training and evaluating retrospective motion correction methods did not exist, making it necessary to simulate motion artifacts. Leveraging the MR-ART dataset and bigger unlabeled datasets (ADNI, OASIS-3, IXI), this work is the first to evaluate motion correction in MRI scans using real motion data on a public dataset, showing that MAMOC achieves improved performance over existing motion correction methods.

Figures (5)

• Figure 1: Varying Levels of Head Motion. 3D & 2D renderings of the same subject from the MR-ART dataset with (a) no head motion, (b) moderate head motion and (c) heavy head motion, (d) motion-corrected with MAMOC
• Figure 2: Architecture.MAMOC arranges G2L transformer blocks in a U-Net-like structure, featuring a contracting encoder path and an expanding decoder path connected by skip connections and a bottleneck layer. During training and inference, inputs are randomly masked for reconstruction. A gating network learns which voxels from the masked input can be preserved and passed to the output.
• Figure 3: (a) We compare how well motion-corrected scans produced by each method are able to reconstruct motion-free scans on MR-ART. (b) Signal-to-noise ratios (SNR) and contrast-to-noise ratios (CNR) as estimated by MRIQC (Markers for statistical significance: '***' for $p < 0.01$, '**' for $p < 0.05$, '*' for $p < 0.1$, 'ns' for $p \geq 0.1$)
• Figure 4: Difference Maps. We visualize the reconstruction quality with absolute difference between the outputs of various models and the ground-truth (GT), i.e., motion-free scans, given the same motion-affected inputs. Each row of difference maps is normalized as a group w.r.t. the unit interval, with lower values indicating better reconstruction quality and greater similarity to the ground truth.
• Figure 5: Subcortical Brain Segmentation. Positive values indicate that segmentation with FASTSURFER benefits from applying an MC algorithm on motion-affected scans. The $y$-axis denotes to which extent the class-averaged (96 classes) Dice score improves if an motion-affected scan is motion corrected. See text for details.