CMotion2Infarct-Net: 3D Cardiac Infarct Reconstruction

• The paper introduces CMotion2Infarct-Net, a model that reconstructs personalized 3D myocardial infarct geometry by explicitly modeling cardiac motion from cine MRI.
• It employs a three-stage framework that integrates 4D biventricular mesh reconstruction, cine-LGE registration for supervision, and motion-centric segmentation to derive simulation-ready outputs.
• Performance evaluations show competitive Dice and Recall metrics, with ablation studies highlighting that explicit motion features significantly enhance infarct localization.

CMotion2Infarct-Net is an infarction reconstruction model introduced in “Personalized 3D Myocardial Infarct Geometry Reconstruction from Cine MRI with Explicit Cardiac Motion Modeling” (Lyu et al., 21 Jul 2025). It is designed to reconstruct personalized 3D myocardial infarct geometry from standard cine MRI without using contrast-enhanced late gadolinium enhancement (LGE) MRI at inference time. The framework first reconstructs a 4D biventricular mesh from multi-view cine MRIs via the automatic deep shape fitting model biv-me, then uses CMotion2Infarct-Net to explicitly utilize motion patterns within this dynamic geometry to localize infarct regions. The method is motivated by the role of infarct geometry in patient-specific cardiac modeling, including structural remodeling, electromechanical dysfunction, arrhythmia risk, and the broader objective of a cardiac digital twin (Lyu et al., 21 Jul 2025).

1. Clinical setting and problem formulation

CMotion2Infarct-Net addresses reconstruction of high-fidelity 3D infarct geometry in post-myocardial infarction patients from non-contrast cine MRI. In the formulation reported for the method, LGE MRI remains the clinical gold standard for infarct or scar detection, but it requires gadolinium contrast injection, introduces side effects and discomfort, increases scan complexity and time, and often still consists of sparse 2D slices, so 3D reconstruction is limited (Lyu et al., 21 Jul 2025).

The framework does not perform direct LGE segmentation and does not synthesize LGE images first. Instead, it reconstructs a 4D biventricular mesh from cine MRI, uses LGE only to generate training supervision through registration and scar projection, and then learns to predict infarct regions directly from cine-derived motion and morphology. This design places motion, rather than enhancement, at the center of infarct localization. The paper explicitly links this to the observation that myocardial infarction is associated with abnormal regional motion and wall thickening changes (Lyu et al., 21 Jul 2025).

A central implication of this formulation is that infarct reconstruction is transferred from an image-domain contrast problem to a dynamic geometry problem. This suggests that the method is intended not merely as a segmentation substitute, but as a mechanism for deriving simulation-ready infarct representations from routine cine acquisitions.

2. End-to-end reconstruction framework

The reported framework has three stages. Stage A is 4D biventricular reconstruction from cine MRI. Stage B is cine-LGE registration and scar projection. Stage C is CMotion2Infarct-Net itself (Lyu et al., 21 Jul 2025).

In Stage A, the preliminary mesh reconstruction step uses biv-me to reconstruct a 4D biventricular surface mesh from multi-view cine MRI. The input consists of a stack of short-axis slices plus 2-, 3-, and 4-chamber long-axis views. The output is a structured 4D biventricular mesh with point correspondence across time, which is described as crucial for motion modeling (Lyu et al., 21 Jul 2025).

In Stage B, manual LGE scar labels are registered onto cine and projected to the 3D mesh to create ground-truth infarct labels for supervised training. Because cine and LGE are acquired separately and differ in resolution, field of view, and respiratory motion, they are spatially misaligned. The reported procedure identifies corresponding slices along the Z-axis, performs in-plane rigid and non-rigid registration, and aligns LGE to cine, especially to end-diastolic phase cine (Lyu et al., 21 Jul 2025).

In Stage C, CMotion2Infarct-Net takes the reconstructed 4D mesh and predicts infarct regions by explicitly modeling spatial anatomy, inter-phase motion, and wall thickness. The full framework is therefore not a conventional voxel-wise pipeline. It operates on a patient-specific dynamic cardiac representation derived from cine MRI and uses projected LGE information only during supervision (Lyu et al., 21 Jul 2025).

3. Mesh reconstruction and supervision generation

The biv-me component is described as an automatic 4D heart reconstruction pipeline with three stages: view selection, segmentation, and cardiac geometric fitting. A ResNet50 identifies useful cine views. nnU-Net segments the left ventricular cavity, right ventricular cavity, and left ventricular myocardium, after which 2D contours are extracted from selected views. The contours are then merged in world coordinates, and a surface template mesh is fitted to them using an iterative diffeomorphic registration algorithm. The fitting is done in two steps: an implicit linear least-squares alignment followed by explicit diffeomorphic refinement. The displacement is constrained so that the Jacobian determinant remains positive, preserving topology and bijectivity (Lyu et al., 21 Jul 2025).

The reported reconstruction quality for biv-me is an average ASD of $2.55 \pm 0.452$ mm on 49 randomly selected subjects against manual contours, together with good agreement in volume change curves over the cardiac cycle. These values are presented as evidence that the 4D mesh captures physiologically plausible motion (Lyu et al., 21 Jul 2025).

Ground-truth infarct generation is based on aligned LGE scar annotations projected onto the mesh. Because sparse 2D slices produce sparse scar labels, the annotations are densified using Gaussian sampling along the Z direction: $\mathcal{N}(\mu,\sigma^2)$ with $\mu$ equal to the original Z-coordinate and $\sigma = 3$ mm. The augmented points are then mapped onto the 5 nearest vertices of the 3D heart surface mesh using a KDTree nearest-neighbor search. For this study, scars are projected only onto the LV endocardium, following prior work (Lyu et al., 21 Jul 2025).

This supervision strategy is consequential for interpreting the method. The “ground truth” infarct geometry is itself a registered and projected label representation, not a directly acquired dense 3D scar volume. The paper later identifies this as a limitation (Lyu et al., 21 Jul 2025).

4. Architecture and motion-centric representation

CMotion2Infarct-Net receives a sequence of 4D biventricular surface meshes,

$\{M_t\}_{t=1}^N ,$

and focuses on the LV myocardium. The preprocessing stage constructs a hybrid input from the LV endocardial mesh plus the corresponding epicardial point cloud. In addition to geometry, the model uses two engineered features: a motion feature $F_{\text{motion}}$, derived from first-order inter-phase differences, and a thickness feature $F_{\text{thick}}$, which encodes wall thickness (Lyu et al., 21 Jul 2025).

The architecture has three main parts. The preprocessing module transforms the mesh input into features suitable for learning, forming a graph or point-cloud representation of the LV surface. The spatio-temporal feature extraction module combines a Graph Neural Network, which extracts structural or spatial features from the mesh geometry, with a two-layer LSTM, which models temporal dependencies across the cardiac cycle. The attention-based segmentation module contains a fully connected layer, a temporal attention layer, max pooling and mean pooling for global spatial aggregation, and a multi-layer perceptron for final node-wise infarct prediction. The temporal attention is reported to use a Transformer with 4 attention heads to capture long dependencies and complex temporal interactions (Lyu et al., 21 Jul 2025).

The training objective is a regularized mesh segmentation loss: $$\mathcal{L}(M_{infarct}, \hat{M}_{infarct}) = \mathcal{L}_{BCEweighted} + \lambda_{Tversky} \mathcal{L}_{Tversky}(\alpha, \beta)$$ where $M_{infarct}$ denotes ground-truth infarct and $\hat{M}_{infarct}$ the prediction. The paper sets $\mathcal{N}(\mu,\sigma^2)$0, $\mathcal{N}(\mu,\sigma^2)$1, and $\mathcal{N}(\mu,\sigma^2)$2. Weighted binary cross entropy is used to address imbalance, while Tversky loss controls the false-positive and false-negative tradeoff. Because infarcts occupy a small fraction of LV nodes, the choice $\mathcal{N}(\mu,\sigma^2)$3 is explicitly aligned with penalizing false negatives more strongly (Lyu et al., 21 Jul 2025).

The architecture therefore encodes a specific hypothesis: infarct localization from cine MRI is recoverable from the joint structure of anatomy, motion, and thickness when those signals are expressed on a temporally corresponding 4D cardiac mesh.

5. Data, implementation, and empirical performance

The method is evaluated on 205 paired LGE and multi-view cine MRI scans from 126 post-MI patients across multiple centers. The cine acquisition comprises short-axis balanced steady-state free precession cine MRI plus 2-, 3-, and 4-chamber long-axis cine views, with short-axis sequences containing 8 to 17 slices across 25 frames. The data split is random by patient with no leakage: 150 training scans, 10 validation scans, and 45 test scans, equivalent to 93 training patients, 6 validation patients, and 27 test patients (Lyu et al., 21 Jul 2025).

Implementation is reported in PyTorch using an AMD EPYC 7K62 CPU and an NVIDIA GeForce RTX 4090 GPU, with Adam optimization, weight decay $\mathcal{N}(\mu,\sigma^2)$4, an initial learning rate $\mathcal{N}(\mu,\sigma^2)$5, and a schedule in which the learning rate is multiplied by 0.7 every approximately 800 iterations. The reported efficiency figures are about 9 min per subject for biv-me, 2 hours for 600 epochs of CMotion2Infarct-Net training, and 5 sec per case for inference (Lyu et al., 21 Jul 2025).

For infarct reconstruction, the evaluation metrics are Dice, Recall, ASD in millimeters, and Generalized Dice. The paper emphasizes Recall and Generalized Dice because infarct occupies only about 8.3% of LV nodes. On the test set, CMotion2Infarct-Net achieves Dice $\mathcal{N}(\mu,\sigma^2)$6, Recall $\mathcal{N}(\mu,\sigma^2)$7, ASD $\mathcal{N}(\mu,\sigma^2)$8 mm, and G Dice $\mathcal{N}(\mu,\sigma^2)$9. Inter-observer variation on 15 cases is reported as Dice $\mu$0, Recall $\mu$1, ASD $\mu$2 mm, and G Dice $\mu$3 (Lyu et al., 21 Jul 2025).

The paper interprets these values as showing reasonable agreement with manual delineation. A plausible implication is that the method is closer to a clinically informative reconstruction system than to expert-equivalent annotation replacement, especially given the gap to inter-observer variability.

6. Ablation evidence, limitations, and relation to adjacent infarct segmentation work

The ablation study identifies explicit motion modeling as the dominant contributor. Removing $\mu$4 causes the largest performance drop, reducing Dice to $\mu$5, Recall to $\mu$6, worsening ASD to $\mu$7 mm, and lowering G Dice to $\mu$8. Removing $\mu$9 also degrades performance, to Dice $\sigma = 3$0, Recall $\sigma = 3$1, ASD $\sigma = 3$2 mm, and G Dice $\sigma = 3$3. Removing temporal attention yields Dice $\sigma = 3$4, Recall $\sigma = 3$5, ASD $\sigma = 3$6 mm, and G Dice $\sigma = 3$7. The paper states that the strongest evidence is that explicit motion modeling materially improves infarct reconstruction (Lyu et al., 21 Jul 2025).

The reported failure modes include overprediction and false positives, with particularly challenging cases in atria-adjacent scars, apical regions, or poorer image quality. The limitations identified in the paper are that supervised labels depend on expert annotation, the ground truth is projected from sparse 2D LGE rather than a direct volumetric scar scan, the current method reconstructs infarct on the LV surface or endocardium rather than a true 3D volumetric tetrahedral scar model, infarct morphology varies substantially, and scar projection is simplified to the LV endocardium (Lyu et al., 21 Jul 2025).

Relative to prior infarct imaging methods in the supplied literature, CMotion2Infarct-Net occupies a distinct methodological category. “Cascaded Convolutional Neural Network for Automatic Myocardial Infarction Segmentation from Delayed-Enhancement Cardiac MRI” (Zhang, 2020) is a DE-MRI segmentation method that uses a 2D U-Net followed by a 3D U-Net for automatic segmentation of myocardium, infarction, and no-reflow, whereas CMotion2Infarct-Net is not a delayed-enhancement segmentation pipeline and does not infer infarct from enhancement contrast. “Symmetry-Enhanced Attention Network for Acute Ischemic Infarct Segmentation with Non-Contrast CT Images” (Liang et al., 2021) likewise targets a different organ system and imaging modality, using symmetry-based alignment and contralateral attention for acute ischemic infarct segmentation in non-contrast head CT. This suggests that CMotion2Infarct-Net is best understood as a cardiac, surface-based, motion-driven infarct reconstruction framework rather than as a conventional voxel-domain lesion segmenter.

The paper positions the model as a step toward contrast-free, patient-specific infarct modeling. It explicitly states that future work will extend the method from a surface-based representation to a 3D volumetric tetrahedral model, with the aim of enabling more direct use in personalized simulations and digital-twin workflows (Lyu et al., 21 Jul 2025).