Hybrid learning/numerical framework for fast and robust electric field simulation in irreversible electroporation

Irreversible electroporation (IRE) represents a promising non-thermal ablation modality for the treatment of deep-seated tumors. However, its clinical efficacy is critically dependent on the accurate, patient-specific distribution of the electric field. While advanced numerical solvers offer physically rigorous simulations, their computational demands render them impractical for intraoperative use, thereby limiting real-time treatment adaptation. This study seeks to develop a clinically oriented workflow designed to enable rapid and reliable electric dose mapping during IRE procedures, thereby enhancing treatment precision and patient outcomes.

We propose a hybrid learning/numerical framework that combines the speed of neural networks with the precision of classical solvers. A convolutional neural network generates rapid approximations of electric potential fields based on electrode configurations and tissue properties, which are then refined through a lightweight iterative numerical correction to enforce physical consistency. The framework is designed to integrate seamlessly into clinical workflows, accommodating intraoperative imaging and segmentation updates.

Evaluations conducted on both synthetic and clinical datasets, including 15 patient cases, using high-resolution domains (100×100×100 voxels at 1 mm³ resolution), demonstrate the model's robustness to variations in electrode configurations and heterogeneous tissue conductivities, two critical factors for personalized IRE treatment planning. Under homogeneous tissue conductivity assumptions, the hybrid solver achieves a 15-fold acceleration in computation time while preserving dosimetric accuracy. In non-homogeneous settings, the method not only surpasses conventional solvers in accuracy but also maintains a computational speedup exceeding twofold.

This work addresses a key barrier to the clinical adoption of numerical simulation in IRE by enabling near-real-time, patient-specific dosimetry. The proposed framework not only accelerates computation but also ensures the reliability required for clinical deployment, supporting more adaptive and precise tumor ablation.