An arXiv preprint describes an AI model that synthesizes a full cycle of motion for both ventricles from a single end-diastolic (ED) mesh, a three-dimensional representation of the heart’s starting shape. In the study’s ED-only comparison, it reported the lowest overall biventricular geometric errors among the methods tested.
The geometric comparison
The central comparison used three measures of geometric agreement between generated and reference meshes: average symmetric surface distance (ASSD), the 95th-percentile Hausdorff distance (HD95) and vertex-wise root mean square error (vRMSE). The proposed method reported a biventricular ASSD of 1.49 ± 0.34 millimetres, HD95 of 3.77 ± 1.06 millimetres and vRMSE of 3.31 ± 1.03 millimetres. The values are reported as means with standard deviations, and all three were the lowest overall in the comparison.
Compared with RePCM, the proposed method was reported to have lower ASSD, at 1.49 millimetres versus 1.69; lower HD95, at 3.77 versus 4.27 millimetres; and lower vRMSE, at 3.31 versus 3.61 millimetres. The paper reported p < 0.01 thresholds for ASSD and HD95 and p < 0.05 for vRMSE.
How the model was tested
The framework uses a motion variational autoencoder to reconstruct full-cycle sequences, then trains a rectified-flow generator in the learned latent space, a compressed representation of the motion. In the primary setup, generation was conditioned on the ED mesh rather than on a full motion sequence.
The evaluation drew on three public cine cardiac MRI datasets—ACDC, M&Ms and M&Ms-2—which were converted into unified-topology biventricular mesh sequences. The reported in-distribution cohort table contained 666 subjects in total: 120 from ACDC, 286 from M&Ms and 260 from M&Ms-2. Its diagnostic breakdown listed 194 NOR, 187 DCM, 175 HCM and 110 RVA; the dataset and diagnostic figures are overlapping breakdowns rather than totals to be added together.
All sequences were resampled to 25 frames, and patients were split at the patient level into training, validation and test sets in a 7:1:2 ratio. Each method generated 20 stochastic samples for every test subject, with metrics averaged within subject. The full-cycle reference sequences were produced by fitting a statistical shape model to segmentation masks.
Functional agreement was weaker on the right
The functional results were less consistent for the right ventricle than for the left. For left-ventricular ejection fraction (LVEF), the reported correlation was r = 0.90 and the mean absolute error (MAE) was 8.1%. For right-ventricular ejection fraction (RVEF), correlation was r = 0.76 and MAE was 9.6%. The synthesized and reference functional values therefore tracked more closely for the left ventricle than for the right.
In additional out-of-distribution (OOD) tests, the authors reported broadly comparable overall error distributions across the groups. Some groups, particularly CIA, showed greater error variability.
Additional tests of the design
A separate test changed how motion was divided into regions. The 16-region motion-derived partition reported biventricular ASSD of 1.49 ± 0.34 millimetres, HD95 of 3.77 ± 1.06 millimetres and vRMSE of 3.31 ± 1.03 millimetres. The AHA-like partition reported 1.59 ± 0.44, 4.11 ± 1.74 and 3.62 ± 1.43 millimetres for the same measures, while the 32-region partition reported 1.59 ± 0.38, 4.19 ± 1.36 and 3.62 ± 1.21 millimetres. The 16-region option had the best reported biventricular ablation performance.
Another ablation compared a configuration without MRAB with the Lhop = 4 configuration. Without MRAB, biventricular vRMSE was 3.51 ± 1.08 millimetres, versus 3.31 ± 1.03 millimetres for Lhop = 4. The Lhop = 4 row also reported ASSD of 1.49 ± 0.34 millimetres and HD95 of 3.77 ± 1.06 millimetres.
The paper also reported a separate optional configuration that used additional motion descriptors. Its biventricular ASSD was 1.46 millimetres and vRMSE was 3.21 millimetres, compared with 1.49 and 3.31 millimetres, respectively, for the ED-only configuration. Because this branch requires the extra descriptors, it was analyzed separately from the primary ED-only setting.
The reference sets the boundary of the claim
How the reference was built matters to the interpretation. The study used full-cycle mesh sequences fitted with a statistical shape model as the reference standard, and those fitted references may inherit errors from segmentation, template fitting and temporal smoothing. The reported differences therefore describe agreement with fitted reference meshes.
No funding or conflict-of-interest statement is reported in the document. The authors state that the code will be released publicly upon manuscript acceptance.
Paper data and sources
Original title: Learning to Beat: Phenotype-Guided Latent Flow with Regional Motion Priors for Biventricular Motion Synthesis
Authors: Xuan Yang, Xiaohan Yuan, Hao Li et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text