PaperlayerView exact revision on arXiv

Source-linked AI summary

Coronary Mask Guided Registration for Continuous Time 4D Cardiac CT Dataset Construction

Yuang Wang, Shuo Wang, Changyu Chen, Dufan Wu, Pengfei Jin, Yunqiang An, Yang Gao, Bin Lu, Dongrui Dai, Muge Du, Yan Yan, Dong Li, Liang Li, Li Zhang, Zhiqiang Chen

arXiv:2608.28712v1eess.IVcs.CV

TL;DR

Clinical multiphase cardiac CT can contain severe non-reference-phase motion artifacts, especially in the RCA, limiting pseudo-ground-truth availability for 4D research. CMGR registers a low-artifact reference volume to other phases, guides registration with RCA masks, and interpolates deformation fields to arbitrary times. It produced artifact-reduced, motion-preserved continuous-time sequences with strong RCA results, competitive whole-heart motion, and plausible intermediate transitions.

  • Problem

    Clinical multiphase reconstructions may contain severe motion artifacts outside ED or ES, limiting pseudo-ground-truth construction for 4D cardiac CT research.

  • Method

    CMGR uses a low-artifact reference phase, RCA-mask-guided registration, reference-volume warping, and deformation-field interpolation to construct continuous-time sequences.

  • Results

    CMGR showed superior RCA-motion and shape performance, competitive whole-heart motion, reduced motion artifacts, and generally plausible intermediate transitions.

  • Takeaways & Limitations

    The resulting patient-specific continuous-time dataset is generally suitable as pseudo ground truth for cardiac CT system simulations and reconstruction-algorithm development.

  • Takeaways & Limitations

    CMGR cannot capture cardiac-valve opening and closing motion, and may slightly over-compress or over-stretch structures adjacent to a substantially displaced RCA.

Abstract

from arXiv · show

Objective: Clinical cardiac CT multiphase reconstructions generally provide acceptable image quality in end-diastole (ED) or end-systole (ES) phases, but in other phases may exhibit motion artifacts, especially in the right coronary artery (RCA). This limits ground-truth availability in 4D cardiac CT imaging research. We aim to construct a 4D cardiac CT dataset that is generally suitable to serve as pseudo ground truth. Methods: We propose Coronary Mask Guided Registration (CMGR) to produce a motion-preserved, artifact-reduced, and continuous-time 4D cardiac CT sequence from the clinical multiphase reconstruction of each patient. For artifact reduction, CMGR uses the ED or ES phase as the reference phase and warps the reference volume with deformation fields to produce the sequence. For motion preservation, CMGR registers the reference phase to each non-reference phase of the multiphase reconstruction. To capture the motion of both the RCA and other cardiac structures in each registration, CMGR regularizes RCA masks and incorporates them into image-domain registration. Time-continuity is achieved by interpolating the deformation fields for non-reference phases to arbitrary times. Results: CMGR outperformed representative image-domain registration methods in capturing RCA motion and providing reasonable RCA shape, and showed competitive performance in capturing whole-heart motion. Additionally, CMGR reduced motion artifacts from clinical multiphase reconstructions, and intermediate CMGR frames generally provided plausible transitions between discrete cardiac phases. Conclusion: CMGR provides an effective approach for constructing continuous-time 4D cardiac CT datasets. Significance: The dataset can be used in system design simulations and in reconstruction algorithm development, thereby facilitating advances in cardiac CT imaging.

I. INTRODUCTION

CMGR addresses motion artifacts and limited temporal resolution in clinical multiphase cardiac CT by constructing motion-preserved, artifact-reduced, continuous-time sequences for pseudo-ground-truth dataset use. It combines RCA-mask guidance with image-domain registration and deformation-field interpolation.

  • Motivation: Clinical multiphase reconstructions often show severe motion artifacts outside ED or ES, particularly around the RCA.This complicates construction of pseudo ground truth for cardiac CT research.
  • Motivation: Existing strategies either retain artifacts, manually deform mainly the RCA, rely on phantoms, or require unavailable raw projection data.Reprojection and post-processing methods can reduce artifacts but typically preserve the original number of cardiac phases.
  • Method: CMGR uses a minimally artifact-affected reference phase, registers it to non-reference phases, and warps the reference volume with the resulting deformation fields.The approach directly processes each patient’s clinical multiphase reconstruction.
  • Method: Interpolating non-reference deformation fields enables cardiac CT frames at arbitrary times rather than only the discrete clinical phases.This design targets continuous-time sequence construction.
  • Method: CMGR regularizes RCA masks and incorporates them into image-domain registration to capture both RCA and broader cardiac motion.The masks enforce connected trunks, reasonable shapes, plausible centerlines, and consistent branches.
  • Results: Experiments showed strong RCA-motion and shape performance, competitive whole-heart motion, reduced motion artifacts, and generally plausible intermediate transitions.CMGR was compared with representative image-domain registration approaches.

II. METHODS

CMGR constructs a continuous-time 4D cardiac CT sequence by selecting a minimally artifacted reference phase, registering it to non-reference phases, interpolating deformation fields, and warping the reference volume.

  • CMGR targets a motion-preserved, artifact-reduced, and continuous-time 4D cardiac CT sequence from each patient’s clinical multiphase reconstruction.
  • The phase with minimal motion artifacts, typically near ED or ES, serves as the reference phase; ED is preferred when both phases have comparable quality.
  • CMGR registers the reference phase to every non-reference phase to estimate whole-heart deformation fields and interpolates them to arbitrary times.
  • The continuous-time sequence is generated by warping the reference volume with the interpolated deformation field.
  • RCA masks are incorporated into image-domain registration to capture RCA motion while preserving motion of other cardiac structures.

A. Stage 1: Derive Plausible RCA Masks

Stage 1 derives plausible RCA masks by repairing disconnected trunks, regularizing mask geometry, and preparing masks for subsequent motion registration.

  • A. Stage 1: Derive Plausible RCA Masks: Stage 1 derives RCA masks for reference and non-reference phases, whose segmentations may be degraded by thin anatomy, occlusion, or motion artifacts.
  • A. Stage 1: Derive Plausible RCA Masks: Regularization connects the RCA main trunk, regularizes mask shape, and propagates reference-phase geometry to non-reference phases.
  • 1) Connect RCA Main Trunk:: Patient-specific proximal and distal reference-phase endpoints identify the RCA fragments that may require connection.
  • 1) Connect RCA Main Trunk:: The closest fragment boundaries are connected by a minimum-cost path on a cost map combining voxel intensity and RCA probability.
  • 1) Connect RCA Main Trunk:: A tubular region around the selected path is merged with the fragments to produce an RCA mask with a connected main trunk.
  • 2) Regularize RCA Mask Shape:: Each RCA mask is converted into a sphere-union representation using a skeletonized centerline and locally estimated vessel radii.

2) Regularize RCA Mask Shape:

CMGR regularizes RCA shape and propagates reliable reference geometry to non-reference phases before RCA-motion registration.

  • 2) Regularize RCA Mask Shape:: Sphere-union masks represent RCA geometry as unions of spherical regions centered on centerline points with associated radii.
  • 3) Propagate RCA Mask Geometry:: Reference-phase RCA geometry is generally more reliable than non-reference geometry degraded by motion artifacts, distorted centerlines, unreliable radii, or abnormal branches.
  • 3) Propagate RCA Mask Geometry:: Geometry propagation uses deformation fields guided by shape-regularized masks, with the same registration strategy later reused in Stage 2.
  • 3) Propagate RCA Mask Geometry:: Displacement along the reference RCA centerline is smoothed before transforming the centerline to each non-reference phase.
  • 3) Propagate RCA Mask Geometry:: Reference-phase local radii are assigned to the propagated centerline under an assumption of negligible cross-phase radius variation, producing final non-reference RCA masks.
  • 3) Propagate RCA Mask Geometry:: The propagated non-reference masks and the reference mask guide Stage 2 RCA-motion capture.

B. Stage 2: Register the RCA

Stage 2 uses DARTEL-based multi-resolution registration guided by plausible RCA masks to capture RCA motion while maintaining an invertible deformation model.

  • B. Stage 2: Register the RCA: Stage 2 applies a multi-resolution registration strategy built on DARTEL to capture RCA motion using Stage 1’s plausible masks.
  • 1) DARTEL:: DARTEL models deformation with a stationary velocity field and optimizes it using similarity and regularization losses.
  • 1) DARTEL:: The DARTEL formulation integrates the velocity field into an invertible deformation that warps the moving volume toward the fixed volume.
  • 1) DARTEL:: Direct DARTEL registration may miss RCA motion when displacement is large or motion artifacts are severe, motivating mask-guided registration.
  • 1) DARTEL:: RCA guidance highlights masked regions with a large constant intensity while retaining other cardiac structures for registration context.
  • 1) DARTEL:: RCA-highlighted reference and non-reference volumes are downsampled with max-pooling and registered from lower to higher resolutions.
  • 1) DARTEL:: Warping the original reference volume with the resulting field produces RCA-aligned volumes with plausible RCA position and geometry for Stage 3.

C. Stage 3: Produce Whole-Heart Deformation Fields

Stage 3 augments RCA-motion deformation fields with whole-heart motion while preserving RCA position and geometry. It aligns non-RCA structures, protects the RCA during registration, and composes the resulting fields for sequence generation.

  • Stage 3 produces whole-heart deformation fields that preserve RCA motion and geometry while better capturing other cardiac structures.
  • For each non-reference phase, registration aligns non-RCA structures from the RCA-aligned volume with the corresponding clinical reconstruction while preserving RCA position and geometry.
  • An expanded, isotropically dilated RCA region receives a constant intensity in both volumes, creating strong correspondence that discourages RCA deformation during DARTEL registration.
  • The whole-heart deformation field combines the Stage 2 RCA-motion field with the deformation field capturing remaining cardiac motion.
  • The resulting fields for all non-reference phases and the reference volume generate a motion-preserved, artifact-reduced, continuous-time 4D cardiac CT sequence.

III. EXPERIMENTS

CMGR was evaluated on simulated XCAT and clinical cardiac CT data against representative image-domain registration methods using phase-resolved geometric, image-quality, and motion-related assessments.

  • Data and reconstruction: CMGR evaluation used male and female beating-heart XCAT simulations with a 1-second cardiac cycle and 20 reconstructed phases at 5% intervals.Cone-beam CT acquisition used 5 gantry rotations per second and 800 views per rotation.
  • Data and reconstruction: Clinical evaluation included 25 cases from Fuwai Hospital and Chinese PLA General Hospital acquired on GE and Siemens CT scanners.Fuwai cases contained 16–20 phases at 5% intervals, while PLA cases contained 10 phases at 10% intervals.
  • Compared methods: The study compared CMGR with FFD, DARTEL, and UGICON using the same reference phase and reference-to-non-reference registration setup.
  • Evaluation: XCAT evaluation used ground truth at original phases and at 10× denser phases, computing Dice, HD, and MSD for RCA masks.

1) XCAT Evaluation:

The evaluation combines quantitative clinical assessment, visualizations, observer rankings, and statistical tests to examine motion capture and anatomical plausibility.

  • Clinical quantitative evaluation: Clinical ground-truth-dependent evaluation was restricted to four high-quality ED–ES pairs, using ES as ground truth and assessing RCA and whole-volume metrics.
  • Clinical quantitative evaluation: NC and FOR were computed on RCA masks for all 25 clinical cases at both original and 10× denser cardiac phases.
  • Visual evaluation: Fig. 2 compares original reconstructions, tested methods, and ground truth across cardiac phases, with magnified boxed regions shown under a [-250, 450] HU display window.
  • Visual evaluation: Fig. 3 visualizes CMGR intermediate frames using fixed boxed regions for each view and the same [-250, 450] HU display window.
  • Observer study: The observer study evaluated RCA motion, RCA geometry, chamber and aorta motion, and chamber and aorta structure using anonymized method sequences.
  • Observer study: Rank scores converted best, medium, and worst assessments to 3, 2, and 1, followed by Friedman and one-sided Wilcoxon tests.

IV. RESULTS

Across XCAT and clinical evaluations, CMGR reduced artifacts, captured RCA motion and geometry more reliably, remained competitive for whole-heart motion, and generated plausible intermediate transitions.

  • 0.03 to 0.34 higher Dice, 44% to 81% lower HD, and 37% to 91% lower MSD were achieved by CMGR on RCA-focused metrics than by other methods.CMGR also showed substantially smaller HD and MSD standard deviations.
  • 0.5 to 2.9 HU lower RMSE with similar SSIM and LPIPS demonstrated competitive CMGR performance for whole-heart motion.
  • Intermediate CMGR frames generally provided plausible transitions between adjacent original phases, with slightly better metrics than original-phase frames.The improvement was attributed to intermediate frames averaging out some overfitting to motion-corrupted reconstructions.
  • In XCAT visualizations, CMGR substantially reduced RCA and LV-corner motion artifacts and aligned better with ground truth for RCA position and shape.
  • CMGR intermediate frames reduced RCA artifacts at 20% and 25% and generally aligned with corresponding ground-truth frames.

B. Results for Clinical Data Evaluations

Clinical evaluations found that CMGR improved RCA motion capture and shape stability, while maintaining competitive whole-heart motion performance. It also reduced artifacts in RCA and non-RCA structures and produced plausible intermediate-frame transitions.

  • Quantitative Results: A 0.10–0.53 Dice increase, 36%–40% HD decrease, and 83%–90% MSD decrease showed CMGR’s strongest RCA-focused performance on ED and ES pairs.CMGR also exhibited substantially smaller Dice and MSD standard deviations.
  • Quantitative Results: CMGR achieved the second-best RMSE and similar SSIM and LPIPS, indicating competitive whole-heart motion capture.
  • Clinical Data Evaluation: CMGR achieved the best FOR and NC with the smallest standard deviations, while 10× denser frames retained comparable average RCA-shape metrics.
  • Visualization Results: Compared with original reconstructions and other methods, CMGR reduced RCA motion artifacts and preserved more plausible RCA positions, shapes, and anatomical courses across phases.The comparisons include axial and coronal views, with competing methods showing shape distortions at some phases.
  • Visualization Results: Intermediate CMGR frames at 15% and 20% reduced RCA artifacts and provided plausible axial and coronal transitions between cardiac phases.
  • Visualization Results: CMGR also reduced motion artifacts in the left coronary artery, right ventricle, left atrium, and aorta.
  • Observer Study: Observer evaluations gave CMGR the best performance across four metrics, with statistically significant advantages for RCA and chamber-and-aorta assessments.RCA Motion and RCA Geometry had p-values below 0.001; Chamber & Aorta Motion and Structures had p-values below 0.02.

3) Observer Study Results:

CMGR produced motion-preserved, artifact-reduced, continuous-time cardiac CT sequences from multiphase reconstructions without raw projection data. The approach was generally suitable for pseudo-ground-truth dataset construction, but its motion modeling has defined limitations.

  • Observer Study Results: CMGR produced sequences without access to raw projection data, while capturing RCA motion and shape, whole-heart motion, and motion continuity.
  • Observer Study Results: The resulting dataset was generally suitable as pseudo ground truth and retained patient-specific cardiac anatomy and motion with generally acceptable motion-artifact levels.
  • Limitations: CMGR cannot reliably capture cardiac-valve opening and closing, so a valve may remain open or closed throughout the generated cardiac cycle.The valve state is largely determined by the reference phase.
  • Limitations: When RCA displacement is large relative to the reference phase, adjacent cardiac structures may be slightly over-compressed or over-stretched.This limitation is associated with the SVF-based RCA deformation model and long RCA trajectories.
  • Future Work: Future work will apply CMGR to more patients and use the resulting dataset for generative-model-based reconstruction algorithms and stationary-CT simulations.
Loading 2608.28712v1…