MARVEL

A cardiac CT scan is trusted only in the phases where the heart happens to be still. ECG gating aims the exposure window at mid-diastole, but the heart is a moving object photographed by a rotating gantry, and when a patient's rate is high or their rhythm irregular the chosen window is not quiet after all. The result is blur along the coronary arteries and chamber walls that can hide a lesion or manufacture one, and the rest of the cycle is discarded.

MARVEL (Motion-Aware Reconstruction Via Embedded Learning of motion prior) treats the motion as part of the reconstruction problem rather than as noise to be gated away. Instead of assuming the anatomy is static during the scan, it estimates a cardiac motion vector field and carries that field into the back-projection, so each measured ray is integrated along the trajectory the tissue actually followed. Motion-compensated reconstruction is not a new idea in CT; what has kept it out of routine use is that the motion field has classically been recovered by iterative optimization for each scan. MARVEL replaces that per-case search with a trained network, MP-Net, that predicts the field in a forward pass.

It comes from Ziheng Deng, Jun Zhao and colleagues in the School of Biomedical Engineering at Shanghai Jiao Tong University, with radiologists from Shanghai East Hospital, published in IEEE Transactions on Medical Imaging in 2026.

#Key Features

• Reconstruction-oriented motion estimation: An embedded learning strategy places the reconstruction step inside MP-Net's training objective, so the predicted motion vector field is optimized for the image it produces rather than for agreement with a reference deformation.
• Mixed spatiotemporal convolutions: The network reads 4D pre-reconstructed images — one volume per cardiac phase — mixing convolution across space and time to extract motion from a stack of images that are themselves still artifact-laden.
• Whole-heart, all-phase imaging: Artifact reduction is applied across the entire heart and across the cycle rather than at one gated phase, lifting the conventional phase restriction on cardiac CT and yielding a time-resolved volume.
• Hybrid rather than end-to-end: Only motion estimation is learned; the reconstruction around it stays model-based, keeping the forward model's interpretability and its flexibility across scan geometries.
• Drop-in to standard workflows: MARVEL consumes ordinary axial cardiac acquisitions with no modified protocol, and the released checkpoint is applied to new scans without retraining or per-case fitting.

#Technical Details

MP-Net takes the 4D image series produced by conventional short-scan reconstruction at a sequence of cardiac phases and outputs the motion vector field describing how the heart deforms across it. Because that field is learned as a prior over cardiac motion rather than fitted to one patient, a single set of weights transfers to scans it has never seen. The field then parameterizes a motion-aware back-projection, and the corrected volume follows from the same model-based machinery a conventional reconstruction uses.

The published evaluation combines simulated and clinical datasets with qualitative and quantitative assessment and blinded reader studies against established motion-correction methods. The authors' public demonstrations run one checkpoint across two vendors' clinical data: 320-row axial scans from a Canon Aquilion ONE, each covering roughly a whole cardiac cycle, and dynamic reconstructions from a UIH uCT 960+. Because the correction is genuinely three-dimensional, the recovered motion stays consistent in coronal and sagittal views, and in one case it also resolves respiratory motion from a failed breath-hold.

#Applications

The immediate use is diagnostic coronary CT in the patients who currently get the worst images — high or irregular heart rates, where a gated reconstruction blurs the vessel wall. The authors' own example is a proximal right coronary artery with mixed plaque and positive remodelling causing moderate stenosis, where the corrected image preserves lumen and plaque morphology the uncorrected short-scan reconstruction loses. Recovering the full cycle also turns a static angiogram into a 4D study of chamber and coronary motion at CT's spatial resolution, a regime usually ceded to cine cardiac MR — and since the input is data the scanner already acquired, it applies retrospectively.

#Impact

MARVEL's contribution is less a new network than a demonstration that the motion field in motion-compensated cardiac CT can be amortized into fixed weights without giving up the model-based reconstruction it feeds — per-case optimization being the cost that has kept these methods in the literature rather than the reading room. The release is correspondingly practical: inference code, a pretrained model and sample data, with the GitHub repository under Apache-2.0. Two limits are worth stating plainly. The composition and scale of the training data are not described in the public materials, so how far the learned motion prior generalizes beyond the demonstrated scanners is a question for external validation. And the artifacts ship through a shared cloud drive rather than a versioned release, making the code and weights less durable than the paper they accompany.