A machine learning surrogate model for fast approximation of simulated microwave ablation zones

Microwave ablation (MWA) is a minimally invasive therapy for liver, lung, and kidney tumors. Computational modeling using finite element methods (FEMs) can simulate ablation zones but is computationally expensive and unsuitable for interactive use. This study develops and evaluates a machine learning surrogate model to rapidly approximate 50 °C isothermal contours from FEM simulations of MWA, enabling faster computational exploration. A feedforward neural network was trained on 2625 FEM simulations incorporating tumor geometry and dielectric properties. Predictions were compared with FEM-derived ablation zones, demonstrating high accuracy and strong correlation (R> 0.95), with error distributions centered near zero. The proposed framework enables rapid approximation of electrothermal simulation outputs, significantly reducing computational cost while maintaining accuracy within the simulated parameter space. However, the model is trained and evaluated exclusively on simulation data and therefore reflects interpolation within this predefined domain. This study represents a simulation-based proof-of-concept, and the proposed model is intended to approximate the outputs of a simplified computational framework rather than directly predict clinical outcomes. Further validation using experimental and clinical data is required before practical applicability can be established. The proposed approach should be interpreted as a surrogate model of the underlying simulation framework rather than a direct predictor of clinical outcomes.