Skip to main content
Back to timeline
arXivSource publication:

Cardiac cine MRI synthesis now traces motion: CardioTrace couples deformation fields and segmentation to curb drift over long sequences

Synopsis

CardioTrace ties cardiac cine MRI synthesis to explicit motion tracing: one branch predicts residual frames relative to the end-diastolic reference, another estimates time-resolved deformation fields from that reference, and segmentation labels of the left ventricle, right ventricle and myocardium at end-diastole and end-systole constrain anatomy. On the public M&M and ACDC cardiac cine MRI datasets it improves perceptual quality and temporal coherence over representative GAN- and diffusion-based baselines, while better preserving ventricular morphology through the cardiac cycle.

Source-provided article image: CardioTrace: Motion-Aware Cine Synthesis via Coupled Tracing and Segmentation
arXiv

Interpretation

Cardiac cine MRI synthesis can now emit an inspectable cardiac motion trajectory alongside the image sequence, rather than only frames that look realistic one at a time. Earlier GAN-based methods modeled temporal trajectories in latent space and reported high pixel-level similarity, but lacked explicit spatial or anatomical constraints, so synthesized sequences could show implausible deformations and motion artifacts. On the public M&M and ACDC cardiac cine MRI datasets, CardioTrace beats representative GAN and diffusion baselines including Pix2Pix, CycleGAN, Med-DDPM and TraceTrans on most metrics, with clearer myocardial boundaries and fewer motion artifacts in end-systolic frames.

The construction splits one generation into complementary appearance and motion pathways, then locks them together with segmentation and consistency constraints. The pixel-translation branch predicts cine frames directly in a residual formulation, while the flow-field branch estimates displacement fields from the reference phase to each time point, giving interpretable frame-to-frame motion correspondences. The segmentation branch supervises left ventricle, right ventricle and myocardium at end-diastole and end-systole, and a cross-branch consistency objective makes synthesized appearance dynamics follow the explicit motion guidance; on M&M, removing the flow branch, the segmentation branch or the meta-conditioning each degrades performance.

Heterogeneity across multi-center data is handled by injecting prompt-style conditioning rather than training a separate model per center. Optional patient meta-data enter as prompt-style conditioning: categorical attributes are embedded and concatenated with normalized numerical attributes, then applied to the encoder bottleneck through feature-wise linear modulation. Meta-conditioning is enabled on M&M and disabled on ACDC, and the ablation shows performance drops further when it is removed.

Perspective

The demonstrated scope is the two public short-axis cardiac cine MRI datasets M&M and ACDC, with an end-diastolic volume as input, sequences uniformly sampled between end-diastole and end-systole to a fixed length, and evaluation by the four image-similarity metrics SSIM, PSNR, MAE and NMI plus qualitative comparison.

The method section notes that registration-based consistency approaches suffer error accumulation over longer cine sequences, producing temporal drift and growing anatomical misalignment across frames, which is what CardioTrace is designed against, but the public material gives no quantitative curve as a function of sequence length. The ablation shows performance drops when meta-conditioning is removed, indicating prompt-style conditioning contributes to robustness on heterogeneous data, with its specific reach still bounded by the M&M results.