Introduction
Inverse problems in medical imaging, such as Electrocardiographic Imaging (ECGI), involve reconstructing internal physiological activity from surface-level measurements. A primary challenge in ECGI is mapping electrical potentials from the irregular 3D geometry of the torso back to the surface of the heart. Traditional deep learning approaches, primarily Convolutional Neural Networks (CNNs), are optimized for Euclidean grids (pixels/voxels) and struggle to generalize when anatomical geometries—such as heart orientation or chest shape—vary between patients. This paper introduces a non-Euclidean framework designed to learn the underlying physics of signal propagation by explicitly accounting for 3D geometry.
Methodology: Geometry-Aware Graph Networks
The proposed architecture moves beyond grid-based convolutions by treating anatomical surfaces as graphs. The core of the system is a non-Euclidean encoding-decoding network utilizing several specialized components:
- Spatial-Temporal Graph Convolutional Neural Networks (ST-GCNNs): These layers process data across the mesh nodes while simultaneously accounting for temporal dynamics, such as the progression of a heartbeat.
- Spline-Based Kernels: By using continuous B-spline bases, the model's convolutions become "geometry-aware," adapting to the local curvature and connectivity of the mesh rather than relying on fixed grid weights.
- Hierarchical Graph Composition: To perform downsampling (pooling) on 3D meshes without losing physical structure, the authors employed specialized mesh coarsening techniques that preserve the topology of the organs.
- The Bipartite Latent Graph: This is the core innovation. The model encodes the torso signal into a latent space and then uses a bipartite graph—which connects two different sets of nodes—to map those signals to the heart's latent space. This mapping is explicitly conditioned on the relative 3D positions of the heart and torso nodes, effectively "learning" the physics-based forward operator in a way that can adapt to any geometry.
Experimental Setup and Datasets
The framework was evaluated using a combination of synthetic simulations and clinical data:
- Synthetic Simulations: 2,700 training sets were generated using the Aliev-Panfilov model to simulate diverse electrical activation patterns and cardiac scars.
- Generalization Benchmarks: The model was tested on heart rotations (x, y, and z axes) and entirely novel heart-torso geometries not present in the training set.
- Clinical Validation: In-vivo 120-lead body-surface potential data from two patients with myocardial scars were used to assess real-world applicability, with results compared against invasive voltage mapping.
Results and Discussion
The geometry-dependent model demonstrated significant improvements in both robustness and accuracy compared to standard Euclidean baselines.
Generalization to Unseen Rotations
In tests where the heart was rotated beyond the training distribution, the proposed model maintained high correlation coefficients and low mean square errors. While standard "flat" models failed to adapt to the shifting physical relationship between the heart and torso, the bipartite graph successfully adjusted the mapping based on the new spatial coordinates.
Visual Accuracy and Clinical Relevance
Qualitative results highlight the model's ability to accurately localize activation origins and scar tissue. As shown in the comparison below, the geometric approach produces reconstructions that closely mirror the ground truth, whereas baseline models often produce blurred or misplaced signals.
While powerful, the model is computationally intensive, requiring approximately three days to train on an NVIDIA Tesla T4. However, this investment yields a model capable of handling the anatomical variability of real patients with reasonable accuracy.
Conclusion
By integrating spatial-temporal graph convolutions with a geometry-conditioned bipartite mapping, this research provides a robust solution for inverse problems on non-Euclidean surfaces. The model's ability to generalize to novel patient anatomies without retraining marks a significant step toward clinical deployment of deep learning in ECGI. Future work will explore fine-tuning strategies for patient-specific optimization and the application of this framework to other domains, such as EEG-based brain activity reconstruction.
Summary updated