Cardio-Synk-Net: A Self-Supervised Dl Framework For Physiologically Consistent Motion Correction In Cine Cardiac Mri

Abstract

Motion artifacts from cardiac and respiratory sources severely degrade cine cardiac MRI quality, compromising diagnostic accuracy in clinical settings. This paper proposes CARDIO-SYNK-Net, a deep learning framework for non-rigid motion correction directly in k-space, addressing limitations of rigid-body and registration-based methods. The framework introduces a Spatio-Temporal Consistency Metric (STCM) to quantify motion-induced k-space inconsistencies via coil-combined magnitude variance and phase decorrelation, enabling adaptive weighting of corrupted segments. CARDIO-SYNK-Net employs a hybrid 3D convolutional-temporal attention architecture trained on synthetic MRXCAT phantoms simulating realistic myocardial contraction, torsion, and respiratory drift. It estimates dense, time-resolved deformation fields from 4D k-space inputs, integrated into an iterative NUFFT reconstruction for motion-compensated images. Evaluated on the STONE dataset (210 subjects), the method outperforms state-of-the-art techniques, achieving R²=0.990±0.026, DSC=0.935±0.051, HD=4.02±3.27 mm, and clinical score=4.81±0.28. CARDIO-SYNK-Net enables robust, reference-free correction of complex physiological motion, paving the way for motion-robust cardiac MRI pipelines.