Learned Finite Element-based Regularization of the Inverse Problem in Electrocardiographic Imaging

1. Problem Statement

Electrocardiographic Imaging (ECGI) aims to non-invasively reconstruct the electrical activity of the heart (epicardial potentials) from body-surface potential measurements.

• The Challenge: The inverse problem is severely ill-posed. Small measurement errors or noise lead to large, unstable variations in the reconstructed cardiac potentials.
• Current Limitations:
  • Classical regularization methods (e.g., Tikhonov, Total Variation) primarily rely on spatial smoothing.
  • They often fail to exploit the temporal structure of cardiac dynamics, which is physiologically critical for capturing wave-like propagation patterns.
  • Existing deep learning approaches often lack interpretability or are not integrated into rigorous Finite Element Method (FEM) frameworks required for complex, unstructured anatomical meshes.

2. Methodology

The authors propose a Learned Finite Element Regularization framework that combines spatial regularization with a data-driven temporal prior.

A. Variational Model

The inverse problem is formulated as minimizing an energy functional $J(u, z) = G(u, z) + R_\theta(u)$, where:

• $G(u, z)$ is the data fidelity term (measuring the difference between simulated body-surface potentials and actual measurements).
• $R_\theta(u)$ is the learned spatiotemporal regularizer.

B. The Learned Regularizer (FoE)

Instead of fixed penalties, the authors employ a Fields-of-Experts (FoE) framework adapted for PDE-constrained inverse problems:

• Structure: The regularizer is a sum of potential functions $\phi_i$ applied to the responses of linear operators $L_i$.
• Operators: The operators $L_i$ are designed to capture both spatial and temporal features:
  1. Zero-order term: $\epsilon_\theta u$ (stability).
  2. Spatial term: Surface gradient $\nabla_{\Gamma_H} u$ (spatial regularity).
  3. Temporal term: A learned temporal cross-correlation $K_i[u]$ using 1D kernels $k_i$.
• Potential Functions: The potentials $\phi_i$ are constructed using Moreau envelopes of the $\ell_\infty$ norm. This ensures the functions are smooth, differentiable, and non-convex (allowing for complex pattern learning) while remaining bounded from below.
• Learning: The kernels $k_i$, potential parameters, and weighting coefficients are learned via a bi-level optimization process. The inner loop solves a Gaussian denoising problem, and the outer loop updates the regularizer parameters to minimize the reconstruction error.

C. Finite Element Discretization

The method is rigorously discretized on unstructured cardiac surface meshes:

• Space: $P1$ finite elements on adaptive surface meshes.
• Time: Uniform temporal grid with $P1$ basis functions.
• Handling Non-Uniformity: Since spatial gradients on unstructured meshes are piecewise constant (discontinuous), while the potential functions require pointwise evaluation, the authors introduce interpolation operators ($P_{sp}$ and $P_{temp}$) to map piecewise constant/linear functions into a common finite element space for evaluation.
• Forward Operator: Discretized using standard FEM for the torso volume conductor, mapping epicardial potentials to body-surface potentials.

3. Key Contributions

1. Theoretical Convergence (Mosco-Convergence)

The paper provides the first theoretical proof of Mosco-convergence for inverse problems regularized by learned FoE-type energies in finite element spaces.

• Result: As the mesh size $h \to 0$, the discrete energy functionals converge to the continuous functional.
• Implication: This guarantees that the minimizers of the discrete problem converge to the minimizer of the continuous problem, ensuring the method is mathematically sound and stable under mesh refinement.

2. Scalable Optimization Algorithm

• The authors developed an accelerated gradient descent algorithm with objective-based restarts.
• They utilize implicit differentiation (via the torchdeq library and Broyden algorithm) to compute gradients through the equilibrium point of the inner optimization problem, enabling efficient training of the neural network parameters without unrolling the entire solver.

3. Integration of Learning and FEM

The work successfully bridges the gap between "black-box" deep learning and rigorous PDE-constrained optimization, creating a regularizer that is:

• Interpretable: Based on explicit operators and potentials.
• Physiologically Plausible: Captures complex spatiotemporal activation patterns.
• Mesh-Agnostic: Works on unstructured, realistic cardiac geometries.

4. Results

The method was evaluated on synthetic 2D epicardial data (simulated using a reaction-diffusion model with scar tissue) and compared against:

• TIK: Spatiotemporal Tikhonov regularization.
• TV: Spatiotemporal Total Variation.
• CMFoE: Convex version of the learned regularizer.
• MFoE: Non-convex (learned) version of the regularizer.

Performance Findings:

• Denoising: The learned MFoE approach significantly outperformed TIK and TV in terms of $L_2$ error. It effectively suppressed noise while preserving sharp edges and wave-like propagation features that TIK blurred and TV sometimes distorted.
• Inverse Problem: In the severely ill-posed ECGI reconstruction task:
  • MFoE achieved the lowest reconstruction errors across all Signal-to-Noise Ratios (SNR).
  • It produced physiologically plausible reconstructions with reduced non-physical oscillations compared to classical methods.
  • The learned temporal interactions helped stabilize the inversion process, capturing activation patterns more accurately than handcrafted priors.
• Robustness: The ranking of performance (MFoE > CMFoE > TV > TIK) remained consistent even as the problem difficulty increased from denoising to the full inverse problem.

5. Significance

• Bridging Theory and Practice: This work demonstrates that data-driven priors can be integrated into rigorous FEM frameworks without sacrificing mathematical convergence guarantees.
• Clinical Potential: By improving the accuracy and physiological plausibility of ECGI reconstructions, this method could lead to better diagnosis of arrhythmias and more effective therapeutic planning (e.g., ablation).
• Generalizability: The framework of learned spatiotemporal regularization on unstructured meshes is applicable to other inverse problems in medical imaging where complex spatiotemporal dynamics are present.

Limitations & Future Work:

• Current implementation assumes uniform timesteps; extending to adaptive/variable timesteps would increase complexity.
• Learning spatial convolutions on manifolds remains computationally expensive, so the current approach relies on fixed spatial operators and learned temporal filters.