One shape, a full cycle
A new preprint describes a model that generates a full cycle of movement for both ventricles from a single end-diastolic mesh—a computer representation of the heart’s surface at one point in the cycle. In its benchmark tests, the method recorded the lowest reported biventricular geometric errors.
For the end-diastolic-only setup, average symmetric surface distance (ASSD) was 1.49 ± 0.34 mm, the 95th-percentile Hausdorff distance (HD95) was 3.77 ± 1.06 mm, and vertex-wise root mean square error (vRMSE) was 3.31 ± 1.03 mm. These are measures of how closely the generated and reference heart surfaces matched, with lower values indicating closer matches.
A measurable edge over the benchmark
Against RePCM, a competing method, the proposed model had lower values on all three biventricular measures: ASSD was 1.49 mm rather than 1.69 mm, HD95 was 3.77 rather than 4.27 mm, and vRMSE was 3.31 rather than 3.61 mm. The reported p values were below 0.01 for ASSD and HD95 and below 0.05 for vRMSE, based on subject-level paired comparisons.
The model also followed the reference contraction-and-relaxation patterns. Agreement for left-ventricular ejection fraction (LVEF), a measure of how much blood the ventricle ejects during a beat, was r = 0.90 with a mean absolute error of 8.1%. For right-ventricular ejection fraction (RVEF), the corresponding figures were r = 0.76 and 9.6%. The supplied analysis reports no confidence intervals for these estimates.
A benchmark built from public MRI data
The study converted three public cine cardiac MRI datasets—ACDC, M&Ms and M&Ms-2—into unified-topology biventricular surface-mesh sequences. The in-distribution cohort contained 666 subjects: 194 classified as normal, 187 with dilated cardiomyopathy, 175 with hypertrophic cardiomyopathy and 110 with right-ventricular abnormality. Five additional diagnostic groups were excluded from training for an out-of-distribution evaluation.
The data were split at the patient level into training, validation and test sets in a 7:1:2 ratio. For each test subject, every method generated 20 independently generated motion samples, and the reported metrics were averaged within subject across those samples.
At its core, the framework groups the heart surface according to how regions move, then uses a conditional generator in a learned internal representation to create region-specific motion adapted to phenotype, the subject’s observed condition.
The details of the design mattered
Component tests did not show that dividing the surface into more regions automatically improved the result. The motion-derived 16-region partition had the best reported biventricular performance, while the 32-region version did not retain those gains. The configuration without multi-hop regional attention bias had higher vRMSE than the reported configuration: 3.51 ± 1.08 mm versus 3.31 ± 1.03 mm. No inferential test was reported for these ablations.
Across the five diagnostic groups held out from training, overall error distributions stayed within a comparable range across the two datasets used for that analysis, although the CIA group showed greater variability. The study did not report an inferential comparison between the out-of-distribution groups.
An optional branch that used additional motion descriptors produced lower biventricular ASSD and vRMSE than the end-diastolic-only condition: 1.46 versus 1.49 mm for ASSD, and 3.21 versus 3.31 mm for vRMSE. That branch was separate from the primary setup and depended on motion information beyond the starting end-diastolic mesh.
What the result does—and does not—establish
The benchmark’s reference sequences were made by fitting a statistical shape model to segmentation masks. That means the comparison may inherit errors from segmentation, template fitting and temporal smoothing, so the reported accuracy is tied to those fitted reference meshes.
The work is identified as arXiv:2608.19738v1, and the authors say the code will be released publicly upon acceptance of the manuscript. The findings therefore remain a preprint methods result: they show performance under the reported benchmark settings, not improved patient outcomes or established clinical benefit.
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