A Deep Learning Framework for Retrospective Motion Detection and Conditional Correction (MDCC): Application to Multi-Gradient-Recalled Echo Quantitative MRI
Year: 2026
Type: article
Abstract: Multi-echo gradient-recalled-echo (mGRE) MRI supports quantitative neuroimaging by enabling voxel-wise T2*/R2*/R2t* mapping for quantitative assessment of tissue damage and magnetic susceptibility imaging. Subject motion degrades these acquisitions by introducing inconsistencies due to phase-encoding steps collected at different head positions. Although the gradient echoes associated with a given phase-encoding line are acquired at practically the same head position, differences in contrast, signal-to-noise ratio, and B0 field-dependent phase evolution cause the resulting artifacts to appear differently across echoes and can bias downstream parametric maps. Existing learning-based retrospective correction approaches often discard complex phase information, treat multiple gradient-recalled echoes independently, and apply correction indiscriminately to both motion-corrupted and clean data. We propose Motion Detection and Conditional Correction (MDCC), a retrospective image-domain pipeline that couples a complex-input volume-level Multi-Echo Motion Detection (ME-MoD) model and a complex-valued Multi-Echo Motion Corrector (ME-MoC) in an iterative detect-correct-re-evaluate loop, where correction is invoked only on volumes flagged as motion corrupted. ME-MoD reweights each echo with a phase-preserving paired complex gate, followed by a shared residual slice encoder, and a selected volume-level aggregator. ME-MoC is a fully complex-valued, factorized U-Net with echo-axis multi-head self-attention at the bottleneck, trained with a T2*-consistency physics loss. ME-MoD attained an AUC of 0.95 and sensitivity of 0.91 on the held-out detection set. At the cohort-level analysis, ME-MoC improved downstream R2* fitting results relative to the motion-corrupted input and 2D U-Net and achieved the lowest numerical R2* root-mean-square error (RMSE) among the evaluated methods, including modest numerical improvement over baseline.
Authors: Gulfam Ahmed Saju(Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft), Jeramy Lewis(Software), Satya V. V. N. Kothapalli(Formal analysis, Investigation), Cihat Eldeniz(Investigation, Methodology), Dmitriy A. Yablonskiy(Conceptualization, Methodology, Project administration, Resources, Supervision, Writing – review & editing)
Institution: Washington University in St. Louis
Language: English
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
Estimated APC paid: $4,690
Awards: Deep-Learning-Augmented Quantitative Gradient Recalled Echo (DLA-qGRE) MRI for in vivo Clinical Evaluation of Brain Microstructural Neurodegeneration in Alzheimer Disease, In vivo Identification of Pre-Atrophic Brain Neurodegeneration in Prodromal Alzheimer Disease with Quantitative Gradient Recalled Echo MRI, In vivo Identification of Pre-Atrophic Brain Neurodegeneration in Prodromal Alzheimer Disease with Quantitative Gradient Recalled Echo MRI
OpenAlex ID: W7235166121
