Researchers report that a method called Coronary Mask Guided Registration, or CMGR, captured the movement of the right coronary artery more accurately than three other registration methods in heart CT tests. The strongest gains appeared in artery-focused measures in both simulated data and a small, high-quality clinical subset, while whole-volume image measures were more mixed.
The work tackles the task of building 4D cardiac CT, in which three-dimensional images show the heart changing over time. Its goal was to turn each patient’s multiphase reconstruction into a continuous-time sequence that preserved motion and reduced artifacts well enough to serve as pseudo ground truth, a reference sequence for testing other methods.
CMGR aligns the images by choosing an ED or ES phase as the reference, registering that phase to every non-reference phase with guidance from an RCA mask, and interpolating the resulting deformation fields to create frames at arbitrary times. In plain terms, the mask tells the alignment process where the right coronary artery is, while the deformation field maps how image locations shift from one phase to another.
From phantom to hospital scans
To test the idea, the authors used male and female beating-heart representations from the XCAT phantom. They reconstructed each representation at 20 cardiac phases, spaced 5% apart through the cycle, and used the 45% phase as the reference.
Clinical evaluation covered 25 cases, including 16 from Fuwai Hospital and 9 from PLA General Hospital. Image quality determined the reference: ED was used in 21 cases, while ES was used in the remaining 4. The clinical work had institutional review board approval from both hospitals.
CMGR was compared with FFD, DARTEL and UGICON, and all tested methods used the same reference phase before registering it to each non-reference phase.
Where CMGR gained ground
In XCAT, CMGR had the best RCA-focused results. Its Dice score, an overlap measure for the artery masks, increased by 0.03 to 0.34 compared with the other methods. Hausdorff distance, which captures the largest boundary mismatch, decreased by 44% to 81%, while mean surface distance, an average boundary gap, decreased by 37% to 91%. For the whole volume, RMSE, an overall image-difference measure, decreased by 0.5 to 2.9 HU, the CT image-intensity unit. SSIM and LPIPS were similar across methods, and the reported standard deviations for Hausdorff distance and mean surface distance were smaller for CMGR.
The method was also checked at a temporal density ten times greater than the original sampling. Metrics averaged across the denser phases were slightly better than metrics from the original phases, and intermediate CMGR volumes generally formed plausible transitions between adjacent points in the sequence. The paper also notes slight overfitting at the original motion-corrupted phases.
Clinical checks were narrower
The clinical quantitative test was narrower. It used four high-quality ED and ES pairs, with ES treated as ground truth and ED registered to ES. That provided a controlled reference for numerical comparison, but ground-truth-dependent testing did not cover all 25 clinical cases.
On those four pairs, CMGR again had the strongest RCA-focused results. Dice increased by 0.10 to 0.53 over the alternatives, while Hausdorff distance decreased by 36% to 40% and mean surface distance decreased by 83% to 90%. Whole-volume RMSE ranked second, with SSIM and LPIPS again similar across methods.
The researchers also checked RCA cross-sectional shape with two measures, FOR and NC. CMGR had the best values for both and the smallest standard deviations, pointing to more stable shapes in this test. When the measures were calculated over phases sampled at ten times the original density, their values were comparable with the original-phase results.
Visual review covered all 25 clinical cases at the original phases. Two radiologists and three cardiac CT researchers received anonymized, randomly shuffled DARTEL, UGICON and CMGR sequences.
CMGR received the best rank scores from both observer groups across all four measures: RCA motion, RCA geometry, chamber and aorta motion, and chamber and aorta structure. The reported p-values were below 0.001 for the two RCA measures and below 0.02 for the chamber-and-aorta measures.
Clinical visualizations showed reduced motion artifacts in non-RCA structures, including the left coronary artery, right ventricle, left atrium and aorta.
The boundaries of the result
The method has clear limits. The authors report that CMGR cannot capture valve opening and closing reliably. They also report that a large RCA displacement may slightly over-compress or over-stretch adjacent cardiac structures.
The clinical numerical comparison rested on four high-quality pairs, even though the wider clinical sample contained 25 cases from two hospitals. The observer findings came from two radiologists and three cardiac CT researchers. The paper reports generally plausible transitions between adjacent CMGR volumes, but that visual plausibility was not a physical validation of every generated frame.
This is a preprint identified as arXiv version 1 dated 28 Aug 2026. The work was supported by the National Natural Science Foundation of China through grants 12327809 and 62031020.
Paper data and sources
Original title: Coronary Mask Guided Registration for Continuous Time 4D Cardiac CT Dataset Construction
Authors: Yuang Wang, Shuo Wang, Changyu Chen et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text