Quantitative multi-echo MRI stops correcting every scan: MDCC gates repair on a motion check and cuts R2* error 16.1% versus corrupted input
Synopsis
Motion artifacts in multi-echo gradient-recalled-echo quantitative MRI propagate along the echo train into R2* and other parametric maps, while existing deep-learning corrections often discard phase, treat echoes independently, and correct clean data indiscriminately. MDCC couples a complex-input motion detector, ME-MoD, with a complex-valued corrector, ME-MoC, in a detect–correct–re-evaluate loop that invokes correction only on volumes flagged as motion corrupted. On a 58-volume held-out set of real subject motion, ME-MoD reached an AUC of 0.947 and sensitivity of 0.91; on 20 synthetic moderate-motion volumes, ME-MoC reduced R2* RMSE by 16.1% relative to the corrupted input and by 30.5% relative to a 2D U-Net, with the difference from LEARN not statistically significant.
Interpretation
Motion correction for multi-echo gradient-recalled-echo MRI can now decide before it acts: only scans flagged as motion corrupted by a volume-level detector enter the correction network, and clean scans are returned unchanged. Earlier deep retrospective correction typically applied the correction model to every input volume, yet a correction network is not an identity operation on clean data and can introduce smoothing, hallucinated high-frequency detail, or amplitude bias. On a 58-volume held-out detection set, all acquired in vivo with real subject motion, ME-MoD at its validation-frozen threshold reached 0.90 accuracy, 0.91 sensitivity, 0.87 specificity and 0.91 F1; against the architecture-matched magnitude-input detector Mag-MoD, sensitivity rose by 0.343 (Holm-adjusted p = 0.004), while specificity fell from 1.00 to 0.87 without a significant difference.
The corrector consumes complex multi-echo volumes directly, treating the ten echoes as complementary views of the same anatomy and steering the output with a T2* decay constraint. Most published methods operate on magnitude images, discard phase, and treat echoes independently, even though phase is needed for susceptibility-based applications and the echoes share anatomy. ME-MoC is a fully complex-valued (2+1)D U-Net with echo-axis multi-head self-attention at its bottleneck, trained with complex fidelity, magnitude alignment and T2*-consistency terms; on 20 synthetic moderate-motion volumes it reached cohort-mean SSIM 0.953 and PSNR 36.73 dB, significantly better than the 2D U-Net on all four metrics (Holm-adjusted p < 0.001), with differences from LEARN not significant after multiplicity correction.
The payoff lands on downstream parametric maps: R2* maps fitted from corrected multi-echo images sit closer to ground truth than those from the corrupted input or the 2D U-Net. Motion does not merely produce ghosting and blur; it propagates into R2t*, R2′ and quantitative susceptibility maps as systematic bias and voxel-level outliers, distorting the quantities the acquisition was meant to measure. Across the 20 synthetic test volumes, ME-MoC reduced R2* RMSE by 16.1% (95% CI 11.6–20.8%) relative to the corrupted input and by 30.5% (95% CI 26.5–34.8%) relative to the 2D U-Net, both Holm-adjusted p < 0.001; the mean paired improvement over LEARN was 0.04 s⁻¹ (95% CI −0.15 to 0.25, p = 0.841), a numerical rather than statistically established difference.
Perspective
The demonstrated scope is ten-echo mGRE data acquired at a single center on a Siemens 3 T scanner with a 32-channel coil: detection is evaluated on 58 real-motion volumes, while reference-based correction and R2* metrics are computed only on 20 synthetic moderate-motion volumes, with real motion limited to two cases lacking paired motion-free references. It suits researchers and engineers working on quantitative MRI motion correction and complex-valued networks; the authors propose future validation on larger prospective datasets across scanners and motion conditions.
The authors note that the motion simulator rotates the existing complex phase without recomputing orientation-dependent B0 changes, does not model through-plane translations or explicit slab-encoding corruption, and places replacements mainly in peripheral k-space, so synthetic results correspond to light-to-moderate simulated motion. The experiments do not directly evaluate corrected-phase fidelity, and all image metrics are computed on magnitude images; MDCC relies on reconstructed complex multi-echo images, which clinical workflows often do not retain, so magnitude-only archives would require retraining. The authors also report a larger R2* positive bias for ME-MoC than LEARN (1.31 ± 0.75 versus 0.43 ± 0.34 s⁻¹), and that its advantage shrinks under light motion, where LEARN better preserved several intermediate and late echoes.
