Technical Summary: Physiology and Anatomy Aware Inverse Inference of Myocardial Infarction for Cardiac Digital Twin

Problem Statement

Accurate localization of myocardial infarction (MI) is critical for risk stratification and therapeutic planning. While Late Gadolinium Enhancement Magnetic Resonance Imaging (LGE-MRI) is the gold standard, it is resource-intensive, costly, and carries safety risks. Electrocardiography (ECG) is widely accessible but lacks the resolution to quantify detailed infarct characteristics such as size, spatial extent, or transmurality.

Existing inverse inference methods that attempt to localize MI from ECG data face two primary limitations:

1. Oversimplified Scar Morphology: They often model scars as regular ellipsoids, failing to capture the realistic, irregular morphology and border zones (BZ) found in ischemic progression.
2. Incomplete Electrophysiological Modeling: They primarily focus on ventricular depolarization (QRS complex), neglecting repolarization (T-wave) dynamics, which reduces sensitivity to subtle ECG variations and limits the interpretability of infarct-induced electrophysiological changes.

The authors formulate MI localization as an inverse problem within a Cardiac Digital Twin framework, aiming to bridge the gap between simulation and reality to enable noninvasive, comprehensive MI assessment using cine MRI and multi-lead ECG.

Methodology

The proposed framework operates in two stages: Anatomy-Aware Stochastic Infarct Synthesis and Inverse Inference via PAA-Net.

1. Anatomy-Aware Stochastic Infarct Synthesis

To generate a realistic virtual cohort, the authors construct simulation-ready biventricular meshes from multi-view cine MRI.

• Mesh Construction: Surface meshes are reconstructed from SAX and LAX segmentation labels, followed by volumetric mesh generation using Iso2Mesh. Fields for ventricular coordinates (Cobiveco), fiber orientations, and electrode placements are assigned.
• Stochastic Scar Generation: Unlike previous methods using regular shapes, this approach generates irregular scars mimicking ischemic transmural progression.
  • A spatially correlated noise field $N(x)$ is generated via Gaussian filtering.
  • A dynamic thresholding function defines the scar core based on the transmural coordinate $T_m$: $Scar = \{x|N(x) > \tau_{base} + \lambda(1 - T_m(x))\}$. This penalizes scar formation toward the epicardium, simulating transmural progression.
  • A border zone (BZ) is created morphologically using a K-D tree search, reclassifying viable nodes within a 2mm Euclidean radius of the scar core.
• Virtual Cohort: The synthesis covers 17 scenarios varying in location (7 types), size (2 types), and transmurality (subendocardial vs. transmural).

2. Electrophysiological Modeling

The authors simulate full QRS-T waveforms using a reaction-Eikonal model:

• Depolarization: Local activation times are computed by solving the Eikonal equation with an orthotropic conduction velocity tensor.
• Repolarization: Action Potential Duration (APD) heterogeneity is implemented via a spatial gradient field accounting for apicobasal and transmural variations.
• Signal Generation: Multi-lead ECGs are generated using a pseudo-ECG formulation, capturing both QRS and T-wave morphologies. Chronic MI remodeling is simulated by scaling conduction velocities and increasing APD in scar and border zones.

3. Physiology and Anatomy Aware Network (PAA-Net)

The inverse inference model jointly encodes 3D myocardial geometry and multi-lead ECGs.

• Geometric Encoding: The mesh point cloud is encoded using PointNet++, utilizing 3D Cartesian and Cobiveco ventricular coordinates.
• Physiological Encoding (A-SE Module): An Anatomical Squeeze-and-Excitation module is introduced to exploit the spatial topology of ECG electrodes. It computes temporal energy for each lead and concatenates it with a spherical lead position prior ($\theta, \phi$). A gating mechanism recalibrates lead weights, dynamically prioritizing leads with the most informative anatomical views.
• Temporal Feature Extraction: A BiLSTM module processes the weighted lead signals to extract temporal features.
• Multi-Modal Fusion (FiLM): The Feature-wise Linear Modulation (FiLM) mechanism uses the global ECG feature as a condition to modulate the point cloud features via affine transformations ($\gamma \cdot F_{pc} + \beta$). This allows the network to selectively emphasize or suppress geometric channels based on electrophysiological context.
• Loss Function: The network is trained with a composite loss including segmentation cross-entropy, compactness loss, and penalties for right ventricle structure and size constraints.

Key Contributions

1. Realistic Scar Synthesis: A novel stochastic strategy that generates irregular scars with border zones, mimicking ischemic transmural progression across diverse locations, sizes, and transmurality types.
2. Full QRS-T Simulation: The construction of a virtual cohort simulating both depolarization and repolarization phases, addressing a gap in existing literature that often neglects T-wave dynamics.
3. PAA-Net Architecture: A novel network that jointly models spatial point cloud information and ECG signals. It introduces the A-SE module for lead-wise anatomical weighting and FiLM for physiological-geometric fusion, enhancing both localization accuracy and interpretability.

Experimental Results

The framework was evaluated on a dataset of 130 cine MRI cases (2,210 simulated samples) and an external validation set of 17 real-world cases.

• Performance Metrics: PAA-Net achieved a mean Dice score of 0.7391 for Scar and 0.5503 for Border Zone (BZ), outperforming a baseline VAE-based method (DiceScar: 0.7064, DiceBZ: 0.5159).
• Generalizability: In external validation on unseen real-world data, PAA-Net significantly outperformed the VAE baseline, achieving a Dice Scar score of 0.2284 vs. 0.1540, and an AHA Score of 0.4490 vs. 0.3607.
• Ablation Studies: The removal of either the A-SE module or the FiLM mechanism resulted in decreased performance, confirming that both lead-wise anatomical weighting and physiological modulation are critical for accurate scar reconstruction.
• Interpretability: Saliency map analysis demonstrated that the model attends to clinically relevant features, specifically the QRS complex and ST-T segments, with attention patterns varying according to infarct location (e.g., septal vs. lateral).

Significance and Claims

The paper claims to be the first study to simulate full QRS-T complexes and integrate them into an inverse MI inference framework. The authors assert that their approach:

• Bridges the Domain Gap: By synthesizing realistic, irregular scars with border zones, the method better mimics real-world ischemic progression compared to regular ellipsoid models.
• Enhances Sensitivity: By incorporating repolarization (T-wave) and APD heterogeneity, the model captures subtle ECG variations that depolarization-only models miss.
• Improves Interpretability: The PAA-Net architecture provides a mechanism to understand how specific ECG leads and morphological features contribute to the localization of infarcted regions.
• Enables Noninvasive Assessment: The framework offers a pathway for comprehensive MI assessment in resource-limited settings by combining accessible cine MRI and ECG, reducing reliance on LGE-MRI.

The authors conclude that while the current framework demonstrates robust performance, future work will focus on enhancing ECG simulation realism, integrating AI-based generative models, and extending the synthesis module to simulate different time phases of MI progression.