DEEP LEARNING FRAMEWORK FOR GLOBAL MOTION CORRECTION IN CINE CARDIAC MRI WITH EXPLICIT DISPLACEMENT FIELD ESTIMATION

Abstract

Cardiac cine magnetic resonance imaging (CMR) is the clinical reference standard for assessing ventricular function, myocardial motion, and valvular abnormalities. However, its diagnostic reliability is often compromised by global motion artifacts, including translation, rotation, and contraction/expansion, which degrade frame-to-frame alignment and impair quantitative accuracy. Traditional correction strategies, such as ECG gating, breath-hold acquisitions, and retrospective registration, remain limited by patient compliance and sensitivity to irregular rhythms, motivating the need for advanced data-driven approaches.In this work, we present a deep learning framework for global motion correction in cine CMR that integrates dual displacement field estimation, a global motion transformation layer, and a hierarchical feature encoding–decoding network with frequency-domain channel attention. The model predicts bidirectional motion fields between moving and fixed frames, enforces temporal consistency, and explicitly corrects for cardiac-specific global displacements. Evaluation was performed on the CMRxRecon2024 dataset, comprising 180 cine CMR subjects with 12 motion states per scan. Experimental results demonstrate that the proposed method achieves a normalized root mean squared error (NRMSE) of 0.098, outperforming the state-of-the-art Coarse-to-Fine Diffusion baseline (NRMSE = 0.1225). These findings confirm that explicitly modeling global cardiac motion within a diffusion-informed encoder–decoder architecture substantially improves reconstruction quality and temporal alignment. The proposed framework advances cine CMR motion correction and has the potential to enhance downstream clinical assessment of ventricular function and valvular pathology.