Prepare the CTA
Normalize orientation and resolution, localize the heart, then crop a 128³ input with the REPRISE-compatible intensity protocol.
Original image stays localCovariance-guided mesh · JOINT TPS 600
From CTA.
To a volumetric
heart.
An agent-skill workflow turns local CTA into a labelled tetrahedral heart mesh. Automatic environment setup. Dataset-aware preprocessing. One shared deformation.
9 predicted surfaces24 template parts1.52M tetrahedra
REPRISE3 / CARDIAC CYCLEMICCAI 2026
Surface mesh · volumetric mesh · tetrahedral cutaway
TEN PHASES, SIDE BY SIDE
One cycle.
Ten fixed perspectives.
Ten evenly spaced acquired phases from the same REPRISE3 series, arranged in two rows of five. Each heart retains its phase-specific shape and rotates in its own panel.
REAL INFERENCE OUTPUT
Compare the acquired phases.
We screened three complete sequences and reconstructed all twenty phases. This gallery displays phases 0%, 10%, …, 90% with a shared camera scale and the same viewing angle in every panel.
10 / 20acquired phases displayed / reconstructed
0nonpositive tetrahedra in these exports
Independent single-phase Covariance-guided mesh inference and joint TPS 600. Surface views of the actual volumetric meshes; geometry stays fixed within each panel. The actual boundary-triangle edges are visible. Only the cameras rotate. No temporal interpolation, remeshing or geometry smoothing.
LOOK BENEATH THE SURFACE
Not just a boundary.
An interior.
Move through the cutaway to see the original tetrahedral elements inside the heart walls. The camera orbits and moves closer; the mesh itself is unchanged.
LV wallRV wallLA wallRA wallAortaPA
MMWHS example. The cut removes complete tetrahedra behind a moving plane; no remeshing or display smoothing.
Inspect the scaled Jacobian map +
Fengming Turbo-64 · common range [0, 1]. Element quality alone does not establish anatomical or simulation validity.
BEYOND THE DEVELOPMENT DATA
New CTA.
The same frozen model.
Explore reconstruction examples from MMWHS and ImageCAS with dataset-aware preprocessing and the same frozen Covariance-guided mesh weights.
EXTERNAL CTA / MMWHSOne template,
One template,
new anatomy.
A labelled whole-heart reconstruction viewed from changing angles. No MMWHS fine-tuning was used for this example.
- Surface predictor
- Covariance-guided mesh
- Registration
- Joint TPS 600
- Nonpositive tetrahedra
- 0
EXTERNAL CTA / IMAGECASLocal data.
Local data.
A portable pipeline.
Native-image preprocessing followed by the Covariance-guided mesh model and shared volumetric deformation. No ImageCAS fine-tuning was used for this example.
- Surface predictor
- Covariance-guided mesh
- Registration
- Joint TPS 600
- Nonpositive tetrahedra
- 0
These are selected qualitative development examples. A separate surface evaluation covers 10 MMWHS and 10 ImageCAS cases; these examples are not a 20-case final-volume validation.
DOWNSTREAM ELECTROPHYSIOLOGY
from Geometry
to Electrophysiology Simulation
An openCARP continuous-current simulation on the anatomical volumetric template illustrates a downstream application of cardiac volume meshes.
Atria & conduction cable
The left view shows saved atrial and cable membrane voltage.
Myocardial tetrahedral cutaway
The right view reveals electrical propagation within the volume mesh.
One preserved voltage scale
Vm in mV, displayed from −100 to +40. The grey-to-colour mapping is retained from the source video.
Supplied anatomical-template diagnostic: right-atrial stimulation, native samples from 0 to 360 ms, fixed geometry, no temporal interpolation.
IMAGE → SURFACE → VOLUME
A shared field.
A coherent whole.
Predict the surface
The Covariance-guided mesh network and frozen weights predict nine cardiac surfaces, including the Aorta and pulmonary artery.
No default retrainingDeform the volume
Geometric correspondences drive one joint TPS field across all template parts, with Aorta continuation and a tetrahedron determinant guard.
One shared deformationMICCAI 2026 / THE METHOD
Covariance-guided.
Template-connected.
HeartVolMesh reconstructs a topology-consistent cardiac surface from CTA, then warps a fixed tetrahedral template while retaining its connectivity and vertex correspondence.
Why a volume mesh?
In-silico cardiac studies need an interior discretization as well as a surface. Thin walls and multi-structure junctions make faithful geometry and consistent connectivity especially valuable.
The core idea
Represent every template vertex as an anisotropic Gaussian kernel. Image-conditioned graph deformation predicts displacement and Cholesky-parameterized covariance, followed by template-driven volumetric registration.
Beyond isolated coordinates
Each vertex stores [x, y, z] plus six covariance parameters. The graph retains topology while learned anisotropy expresses spatially varying, direction-dependent geometric support.
Covariance-guided matching
The paper's surface-learning objective uses a covariance-aware negative log-likelihood with edge, Laplacian and normal regularization:
ℓ(x, y; Σ) = ½ (x − y)ᵀ Σ⁻¹ (x − y) + β log det Σ
This describes the paper's learning objective. The agent-skill workflow uses frozen inference weights and does not train by default.
Volume → graph
A 3D CNN extracts multi-scale features. Sampling at the current vertex positions brings image information into a GNN over the mesh graph.
Coarse → fine
Progressive graph refinement predicts vertex displacements and Cholesky parameters for anisotropic covariance.
Boundary → interior
The fitted surface motion is propagated to every tetrahedral-template vertex, preserving the fixed connectivity. The pipeline uses one shared joint TPS field.
THE PAPER / COMPARISONS & EVALUATION
Geometry.
Measured and compared.
The original project page's scientific figures and reported cohort results are preserved below. These paper experiments are separate from the selected reconstruction and EP examples above.
Dataset & training
Multi-centre in-house data: 900 patients / 4,000 temporal instances. Patient-level split: 800 patients / 3,600 training instances and 100 patients / 400 validation instances.
Paper configuration: NVIDIA A100 80GB, Adam, 100 epochs, batch size 1, 128³ input crop and 300³ deformation-field resolution.
Evaluation protocol
Surface structures: LA, LV, RA, RV and LV myocardium. Volumetric boundaries: LA, LV, RA and RV myocardium.
Symmetric Chamfer distance (CD), HD95 and normal consistency (NC), plus inverted elements, minimum scaled Jacobian and minimum dihedral angle.
0.0%inverted elements across 400 validation instances
0.0416 ± 0.0152minimum scaled Jacobian
5.56° ± 1.41°minimum dihedral angle
Surface meshes · Ours 3p6
| Structure | CD ↓ | HD95 ↓ | NC ↑ |
|---|---|---|---|
| LA | 2.5 | 2.2 | 89.5% |
| LV | 3.0 | 2.8 | 88.7% |
| Myo | 4.0 | 3.6 | 87.6% |
| RA | 2.9 | 2.8 | 88.3% |
| RV | 3.3 | 3.1 | 87.3% |
Volumetric boundaries · Ours 3p6
| Structure | CD ↓ | HD95 ↓ | NC ↑ |
|---|---|---|---|
| LVMyo | 3.6 | 3.2 | 95.3% |
| RVMyo | 3.1 | 2.8 | 94.2% |
| LAMyo | 3.0 | 3.0 | 87.7% |
| RAMyo | 3.0 | 2.9 | 92.2% |
Boundary localization
Covariance-guided supervision improves tolerance around ambiguous and thin anatomical regions.
Mesh correspondence
Deforming a fixed template retains connectivity and supports cross-case vertex correspondence by construction.
Template flexibility
The chosen template controls resolution and element specification for downstream FEM/CFD research. Application-specific validity still needs assessment.
MICCAI 2026 / CITE THIS WORK
Lin, Fengming, et al. "HeartVolMesh: Cardiac Volumetric Mesh Reconstruction via Covariance-Guided Graph Deformation." International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer International Publishing, 2026.
NVIDIA ACADEMIC PROGRAM / RESEARCH SUPPORT
Fengming Lin is funded by the NVIDIA Academic Program as Co-I.