TetHeart: 4D Cardiac Mesh Recovery

• TetHeart is a 4D heart mesh recovery framework that uses deep deformable tetrahedra and a hybrid explicit-implicit representation to reconstruct patient-specific cardiac anatomy.
• It employs an attentive slice-adaptive 2D–3D feature assembly, integrating both full-stack cine CMR and sparse intra-procedural slices for robust volumetric feature extraction.
• A two-stage weakly supervised motion learning scheme enhances dynamic reconstruction accuracy, delivering clinically relevant metrics and real-time performance at 12 FPS.

TetHeart most commonly denotes an end-to-end deep learning framework for patient-specific 4D heart mesh recovery from cardiac MRI, designed to operate on both offline full-stack cine CMR and intra-procedural sparse 2D slices. In its primary contemporary usage, it is described as the first end-to-end framework that unifies full 4D multi-structure heart mesh recovery from both offline full-stack acquisitions and intra-procedural sparse-slice observations, using deep deformable tetrahedra, an explicit-implicit hybrid representation, attentive slice-adaptive 2D-3D feature assembly, and a two-stage weakly supervised motion learning scheme requiring only keyframe annotations (Chen et al., 15 Sep 2025).

1. Name, scope, and disambiguation

In the current literature, “TetHeart” is not a fully unambiguous label. The principal usage refers to the 2025 framework for 4D heart mesh recovery across full-stack and sparse cardiac MRI (Chen et al., 15 Sep 2025). A related 2024 MedTet-derived technical description also presents “TetHeart” as a deformable-tetrahedral-mesh system for real-time 4D heart reconstruction from sparse intraoperative data, including 3D volumes, a few 2D MRI slices, or even 1D signals (Chen et al., 2024). Separately, a 2018 time-series paper uses “TetHeart” to denote the profile of 4-node horizontal-visibility-graph motifs applied to heartbeat-rate analysis rather than geometric heart reconstruction (Xie et al., 2018).

This terminological overlap matters because the three usages address different objects. The 2025 and 2024 works are concerned with geometric and kinematic reconstruction of cardiac anatomy through tetrahedral representations, whereas the 2018 usage concerns network-motif statistics on RR-interval sequences. A common misconception is therefore to treat every occurrence of “TetHeart” as a cardiac mesh-recovery framework; in fact, context determines whether the term denotes a 4D geometric model or a heartbeat-dynamics descriptor (Chen et al., 15 Sep 2025).

2. Deep deformable tetrahedra and static shape recovery

TetHeart’s geometric core is a hybrid explicit-implicit representation based on deep deformable tetrahedra. The heart model is discretized as

$$\mathcal{G}=(\{\mathbf{v}_i\},\{s_i\},\mathcal{T}),$$

where $\mathbf{v}_i\in\mathbb R^3$ are tetrahedral grid vertices, $s_i$ are signed distance function values, and $\mathcal{T}$ is the tetrahedron connectivity. The explicit surface is obtained as the zero-level set of the SDF through differentiable marching tetrahedra. Given the end-diastolic initialization $\mathcal{G}^0$, temporal evolution is represented by deforming the vertices while keeping the SDF values and connectivity fixed: $$\mathcal{G}^t=\bigl(\{\mathbf{v}_i^t\},\{s_i\},\mathcal{T}\bigr), \qquad \mathbf{v}_i^t=\mathcal{D}(\mathbf{v}_i^0,\mathbf{O}^t),$$ with $\mathcal{D}$ a learned deformation operator driven by observations $\mathbf{O}^t$ (Chen et al., 15 Sep 2025).

Static reconstruction begins from an end-diastolic full CMR stack $\mathbf{I}^0\in\mathbb R^{D\times H\times W}$. A graph convolutional network predicts per-vertex offsets $\Delta\mathbf{v}_i$ and SDF values $\mathbf{v}_i\in\mathbb R^3$0 from trilinearly interpolated static features, yielding an initial patient-specific tetrahedral mesh $\mathbf{v}_i\in\mathbb R^3$1. The deformed vertex coordinates satisfy $\mathbf{v}_i\in\mathbb R^3$2, and the resulting mesh captures patient-specific anatomy at ED (Chen et al., 15 Sep 2025).

A related system description grounded in MedTet states the same representation in more constructive terms. There, $\mathbf{v}_i\in\mathbb R^3$3 defines tetrahedral-grid vertices, $\mathbf{v}_i\in\mathbb R^3$4 the tetrahedra, and $\mathbf{v}_i\in\mathbb R^3$5 samples a continuous SDF on the grid. Within each tetrahedron, the SDF is barycentrically interpolated, and the zero-isosurface is extracted with Marching Tetrahedra. The sampling grid is uniform inside a bounding cube, with a representative resolution of $\mathbf{v}_i\in\mathbb R^3$6, corresponding to approximately $\mathbf{v}_i\in\mathbb R^3$7 tetrahedra (Chen et al., 2024). This makes explicit why TetHeart is not merely a surface-mesh method: geometry is encoded volumetrically, while surfaces remain differentiably recoverable.

3. Slice-adaptive 2D–3D feature assembly

TetHeart’s central mechanism for handling sparse observations is the Attentive 2D–3D Feature Assembler, or AFA. The purpose of AFA is to compute volumetric features from arbitrary sets of 2D slices while keeping the downstream mesh-recovery architecture unchanged across offline and online settings. Static and temporal inputs are first encoded by a 2D CNN: $\mathbf{v}_i\in\mathbb R^3$8 Each 2D feature map is then spatially registered into the reference 3D volume using DICOM metadata (Chen et al., 15 Sep 2025).

Motion features are assembled through localized multi-head attention,

$\mathbf{v}_i\in\mathbb R^3$9

where $s_i$0 retrieves for each voxel the $s_i$1 nearest slices and, within each slice, the $s_i$2 nearest pixels. The reported setting is $s_i$3 and $s_i$4, and a learnable 3D positional embedding is added to the keys. Static features $s_i$5 are computed analogously by self-attending $s_i$6. Both feature volumes are then refined by a U-Net-style encoder to produce final feature grids $s_i$7 and $s_i$8 (Chen et al., 15 Sep 2025).

Two architectural decisions are emphasized in the reported ablations. First, AFA requires both image and spatial cues for sparse generalization. Second, the 2D encoder and 3D U-Net weights are shared between static and motion branches, transferring spatial knowledge from the shape-reconstruction pathway to the motion pathway (Chen et al., 15 Sep 2025). This directly addresses another misconception: TetHeart does not reconstruct sparse-slice motion by naively stacking or zero-filling slices alone; it uses explicit slice-adaptive attention and shared volumetric encoding.

4. Motion deformation, distillation, and weak supervision

TetHeart is trained with a two-stage weakly supervised motion learning scheme tailored to the fact that many CMR datasets provide only ED and ES annotations. Stage I trains the static shape branch on labeled frames by minimizing

$s_i$9

with $\mathcal{T}$0, $\mathcal{T}$1, and $\mathcal{T}$2. Here $\mathcal{T}$3 denotes marching tetrahedra, $\mathcal{T}$4 is Chamfer distance, $\mathcal{T}$5 is an L1-SDF loss, and $\mathcal{T}$6 is cross-entropy (Chen et al., 15 Sep 2025).

Stage II trains the motion branch by pairing an unlabeled frame $\mathcal{T}$7 with a labeled keyframe $\mathcal{T}$8. The model predicts $\mathcal{T}$9, deforms it to $\mathcal{G}^0$0 using full slices, and to $\mathcal{G}^0$1 using sparse slices. The motion objective is

$\mathcal{G}^0$2

where the feature-level distillation term is

$\mathcal{G}^0$3

This distillation co-trains full-slice and sparse-slice settings by forcing sparse-slice motion features to match full-slice motion features (Chen et al., 15 Sep 2025).

The reported implementation uses a modified nnU-Net with five levels and channel widths 32–320 as the 2D encoder, a five-level 3D U-Net shared by static and motion branches, 8-head attention in AFA, a 3-layer GCN with 128 channels for shape reconstruction, an initial tetrahedral grid at $\mathcal{G}^0$4 resolution, and a deformation network consisting of 2 iterations of GCN $\mathcal{G}^0$5 GRU with hidden size 128 $\mathcal{G}^0$6 MLP offset prediction. Optimization uses SGD with learning rate 0.01, momentum 0.99, and weight decay $\mathcal{G}^0$7; training lasts 300 epochs for the static branch and 150 for the motion branch on a single NVIDIA V100 (Chen et al., 15 Sep 2025).

A related MedTet-based formulation describes the online motion model as two iterative vertex updates per frame, combining graph convolution, GRU state propagation, and MLP-predicted shifts, with additional smoothness and tetrahedral volume-preservation energies in the total loss (Chen et al., 2024). This suggests a broader methodological family in which TetHeart-style systems combine tetrahedral geometry, recurrent graph deformation, and weak supervision.

5. Evaluation, ablation, and runtime characteristics

TetHeart is trained on a unified public cohort comprising ACDC with 100 patients, M&Ms with 375, and M&Ms-2 with 360, for a total of 835 sequences with ED/ES ground truth. It is externally evaluated on CHUV interventional MRI and on 4DM, which provides full myocardium labels for every frame (Chen et al., 15 Sep 2025).

On the Unified Test set in the few-slice regime, specifically 1/5 slices, TetHeart reports Myo CD $\mathcal{G}^0$8, LV CD $\mathcal{G}^0$9, RV CD $$\mathcal{G}^t=\bigl(\{\mathbf{v}_i^t\},\{s_i\},\mathcal{T}\bigr), \qquad \mathbf{v}_i^t=\mathcal{D}(\mathbf{v}_i^0,\mathbf{O}^t),$$0, and Myo Dice $$\mathcal{G}^t=\bigl(\{\mathbf{v}_i^t\},\{s_i\},\mathcal{T}\bigr), \qquad \mathbf{v}_i^t=\mathcal{D}(\mathbf{v}_i^0,\mathbf{O}^t),$$1. In the same table, Ours-Mesh reports Myo CD $$\mathcal{G}^t=\bigl(\{\mathbf{v}_i^t\},\{s_i\},\mathcal{T}\bigr), \qquad \mathbf{v}_i^t=\mathcal{D}(\mathbf{v}_i^0,\mathbf{O}^t),$$2, LV CD $$\mathcal{G}^t=\bigl(\{\mathbf{v}_i^t\},\{s_i\},\mathcal{T}\bigr), \qquad \mathbf{v}_i^t=\mathcal{D}(\mathbf{v}_i^0,\mathbf{O}^t),$$3, RV CD $$\mathcal{G}^t=\bigl(\{\mathbf{v}_i^t\},\{s_i\},\mathcal{T}\bigr), \qquad \mathbf{v}_i^t=\mathcal{D}(\mathbf{v}_i^0,\mathbf{O}^t),$$4, and Myo Dice $$\mathcal{G}^t=\bigl(\{\mathbf{v}_i^t\},\{s_i\},\mathcal{T}\bigr), \qquad \mathbf{v}_i^t=\mathcal{D}(\mathbf{v}_i^0,\mathbf{O}^t),$$5, while Ours-SDF reports Myo CD $$\mathcal{G}^t=\bigl(\{\mathbf{v}_i^t\},\{s_i\},\mathcal{T}\bigr), \qquad \mathbf{v}_i^t=\mathcal{D}(\mathbf{v}_i^0,\mathbf{O}^t),$$6, LV CD $$\mathcal{G}^t=\bigl(\{\mathbf{v}_i^t\},\{s_i\},\mathcal{T}\bigr), \qquad \mathbf{v}_i^t=\mathcal{D}(\mathbf{v}_i^0,\mathbf{O}^t),$$7, RV CD $$\mathcal{G}^t=\bigl(\{\mathbf{v}_i^t\},\{s_i\},\mathcal{T}\bigr), \qquad \mathbf{v}_i^t=\mathcal{D}(\mathbf{v}_i^0,\mathbf{O}^t),$$8, and Myo Dice $$\mathcal{G}^t=\bigl(\{\mathbf{v}_i^t\},\{s_i\},\mathcal{T}\bigr), \qquad \mathbf{v}_i^t=\mathcal{D}(\mathbf{v}_i^0,\mathbf{O}^t),$$9. Clinical MAE at 1-slice on M&Ms is reported as $\mathcal{D}$0 ml for LVESV, $\mathcal{D}$1 for LVEF, $\mathcal{D}$2 ml for RVESV, and $\mathcal{D}$3 for RVEF, each improving on the corresponding Ours-Mesh values of $\mathcal{D}$4 ml, $\mathcal{D}$5, $\mathcal{D}$6 ml, and $\mathcal{D}$7 (Chen et al., 15 Sep 2025).

In the full-slice setting, TetHeart reports Myo CD $\mathcal{D}$8, LV CD $\mathcal{D}$9, RV CD $\mathbf{O}^t$0, and Myo Dice $\mathbf{O}^t$1. The same section states that TetHeart outperforms all methods on clinical MAE, with LVESV $\mathbf{O}^t$2 ml and LVEF $\mathbf{O}^t$3. In zero-shot evaluation on 4DM, without retraining and using a model trained on ACDC+M&Ms+M&Ms-2, TetHeart achieves myocardium CD of $\mathbf{O}^t$4 mm$\mathbf{O}^t$5 for $\mathbf{O}^t$6full slices, whereas an “Oracle” trained directly on 4DM yields $\mathbf{O}^t$7 mm$\mathbf{O}^t$8 (Chen et al., 15 Sep 2025).

The CHUV interventional experiments report physiologically plausible motion under rest and exercise, with heart-rate increases of 20–30 bpm and volume–time curves matching expected cardiac-cycle shortening. Inference runs at approximately 12 FPS on a V100, with the paper noting potential further acceleration via lightweight encoders or quantization (Chen et al., 15 Sep 2025).

Ablation results further clarify which components are essential. On M&Ms myocardium CD, the full model obtains $\mathbf{O}^t$9 for full-slice/5-slice/1-slice settings. Removing AFA leads to failure on sparse inputs, though the full-slice score is $\mathbf{I}^0\in\mathbb R^{D\times H\times W}$0. Removing positional embedding yields $\mathbf{I}^0\in\mathbb R^{D\times H\times W}$1; removing image features in the query produces $\mathbf{I}^0\in\mathbb R^{D\times H\times W}$2; removing the shared encoder gives $\mathbf{I}^0\in\mathbb R^{D\times H\times W}$3; and removing distillation gives $\mathbf{I}^0\in\mathbb R^{D\times H\times W}$4. The reported interpretation is that AFA with both image and spatial cues is essential for sparse generalization, shared 3D encoder weights transfer shape knowledge to motion, and distillation stabilizes few-slice performance without hurting full-slice accuracy (Chen et al., 15 Sep 2025).

6. Relation to adjacent tetrahedral meshing work and alternate non-geometric usage

TetHeart sits within a broader body of work on tetrahedral cardiac modeling, but it is distinct from static volumetric meshing methods. HeartVolMesh, for example, addresses patient-specific tetrahedral cardiac meshes from CTA by lifting template surface vertices to anisotropic Gaussians, predicting per-vertex displacements and Cholesky-parameterized covariances with a 3D CNN–GNN, and warping a fixed tetrahedral template through staged alignment, coherent point drift, rasterized deformation fields, and deformation propagation. Its volumetric mesh preserves connectivity and correspondence by construction, and experiments report gains in surface mesh accuracy and volumetric mesh fidelity over deformation-only baselines (Lin et al., 5 Jul 2026). This places HeartVolMesh adjacent to TetHeart in representation choice, but in a different imaging modality and task setting: static CTA meshing rather than 4D CMR motion recovery.

The alternate 2018 usage of “TetHeart” is conceptually separate. There, the term refers to tetradic motif profiles of horizontal visibility graphs derived from heartbeat time series. For RR-interval data, the method maps a scalar sequence $\mathbf{I}^0\in\mathbb R^{D\times H\times W}$5 to a graph in which two nodes are connected when the corresponding samples satisfy a horizontal-visibility condition. The analysis considers six connected induced subgraphs on four chronologically ordered nodes, defines motif frequencies $\mathbf{I}^0\in\mathbb R^{D\times H\times W}$6, and applies these frequencies to distinguish healthy, congestive-heart-failure, and atrial-fibrillation subjects. Reported group averages for windows of length $\mathbf{I}^0\in\mathbb R^{D\times H\times W}$7 are approximately $\mathbf{I}^0\in\mathbb R^{D\times H\times W}$8 for healthy subjects, $\mathbf{I}^0\in\mathbb R^{D\times H\times W}$9 for CHF, and $\Delta\mathbf{v}_i$0 for AF, with two-sample $\Delta\mathbf{v}_i$1-tests on $\Delta\mathbf{v}_i$2 and $\Delta\mathbf{v}_i$3 giving $\Delta\mathbf{v}_i$4 (Xie et al., 2018).

The coexistence of these usages creates a modest nomenclature controversy rather than a methodological one. In cardiac imaging and geometric modeling, TetHeart generally denotes a tetrahedral 4D reconstruction framework centered on sparse-to-dense observation fusion (Chen et al., 15 Sep 2025). In nonlinear time-series analysis, TetHeart denotes a heartbeat motif-profile method on horizontal visibility graphs (Xie et al., 2018). Precision therefore requires specifying the associated paper or task whenever the term appears in technical discourse.