Motion-robust whole-heart 5D MRI using Expectation-Maximization-Guided Binning Correction and Outlier Rejection (EMORe)

Background: Free-breathing volumetric CMR data are acquired over several minutes. These methods rely on data binning (e.g., using self-gating (SG)) to resolve cardiac and respiratory motion. However, data binning can be inaccurate, particularly in the presence of irregular breathing, arrhythmias, or bulk motion, leading to residual motion artifacts that limit diagnostic utility.¹ To address this, we propose using expectation–maximization² (EM)-guided binning-correction and outlier-rejection framework (EMORe) for reconstruction. In prior work (SCMR’25), we applied EMORe to a 2D digital phantom study,³ but this method has not been validated for in vivo volumetric CMR. Here, we validate EMORe’s effectiveness for whole-heart 5D MRI reconstruction against standard compressed sensing (CS).

Methods: EMORe addresses imperfect cardiorespiratory binning and motion corruption of data readouts by integrating adaptive inter-bin correction and outlier rejection within an EM framework. EMORe iteratively performs two steps until convergence: the E-step and the M-step (Fig. 1). In the E-step, probabilistic (soft) bin assignments are refined by correcting misassignment of valid readouts and rejecting corrupted readouts to an outlier bin. In the M-step, the image estimate is improved using the refined soft bin assignments.

Thirteen healthy volunteers (age 20-68 years; 6 female) with ferumoxytol infusion were scanned on a 3T system for 5 minutes using 3D Cartesian sampling with SG.⁴ Three volunteers were instructed to cough during the final 30 s to provide a controlled evaluation of EMORe’s ability to discard corrupted readouts. Data readouts were retrospectively sorted into 20 cardiac and 4 respiratory bins using cardiac and respiratory signals extracted from SG. Binned data were reconstructed using EMORe and CS. Qualitative assessment used blinded artifact scoring (1—non-diagnostic to 5—excellent) of sagittal cine pairs at end-expiration and end-inspiration by two experts; quantitative assessment used blood–myocardium edge sharpness,⁵ measured on a single sagittal slice where the blood pool was fully enclosed by the myocardium.

Results: Across 26 cine pairs (13 × two respiratory phases), EMORe significantly improved the averaged blind artifact score from two experts versus CS (mean±SEM: 4.06±0.08 vs 3.62±0.15; paired t-test p< 0.001). Edge sharpness was also higher with significance using EMORe (0.72±0.028 vs 0.69±0.023 (CS); paired t-test p< 0.01). In Fig 2, frames and temporal profiles from three representative volunteers show reduced motion artifacts and improved delineation of the blood–myocardium boundaries. Aggressive outlier rejection during the known coughing (volunteer #1) and potential deep breath (volunteer #2), shown with corresponding respiratory signals in Fig. 3, supports EMORe’s ability to reject motion-corrupted data.

Conclusion: EMORe enables motion-robust whole-heart 5D MRI using adaptive, probabilistic, retrospective data binning correction and outlier rejection during reconstruction.

Fig. 1: Schematic overview of the proposed EMORe framework. In the E-step, we refine bin participation of readouts to valid motion bins and an outlier bin, given the prior bin participation and current image estimate. (Right) In the M-step, we improve the image estimate using the refined bin participation. Both steps are repeated until convergence, resulting in motion-compensated images.
Fig. 2: Representative sagittal cine frames at end-expiratory and end-inspiratory phases for three volunteers reconstructed using CS and EMORe. Corresponding temporal profiles along the x–t and y–t dimensions are presented for qualitative assessment of temporal consistency and boundary definition. Arrows highlight the comparison of motion artifacts suppressed by EMORe. (†Volunteer instructed to simulate coughing during the final 30 seconds of the scan.)
Fig. 3: Self-gating-based respiratory surrogate signals (blue) from the three volunteers in Fig. 2 overlaid with the assignment percentage to the outlier bin for the corresponding readout (vertical red bars). Aggressive outlier rejection during the instructed coughing interval (horizontal black bar, volunteer #1) and potential deep breath (magenta arrow, volunteer #2) supports EMORe’s ability to reject motion-corrupted (outlier) data.