Real‐Time Prediction of Coupled Electric and Temperature Fields in Radiofrequency Ablation: A Physics‐Integrated Neural Network Approach

Radiofrequency ablation (RFA) is a widely used minimally invasive technique for solid tumors. Real-time feedback on the thermoelectric effects induced by RFA is very important for precise, personalized treatment. Current computer simulations help predict the electrical and thermal phenomena related to RFA, their high computational cost limits practical clinical use, especially for real- time monitor. In this study, a physics-integrated neural network-based model was proposed to predict the coupled real-time electric and temperature fields during the treatment. This approach, similar to traditional simulation methods, simulates the physical changes during the radiofrequency ablation process with input of treatment parameters. The combined deep learning model consists of a DeepONet network predicting the electrical potential distribution and a coupled ConvLSTM network forecasting the temperature distribution over time. The networks were trained using results from the thermoelectric coupling FEM model, and validated through bio-mimic phantom experiments. The DeepONet network achieves a mean absolute error (MAE) of 0.0241 and a mean maximum relative error (MRE) of 1.44%. The coupled ConvLSTM network achieves an MAE of 0.0286, an MRE of 3.46%, and a Dice score of 0.9334 for areas above 45°C. The model developed can provide coupled temperature and electric field predictions for a 120-s RFA process with varying properties in less than 1 s. This rapid prediction method is expected to be integrated with control calibration algorithms in the future, enabling the acquisition of real-time three-dimensional temperature fields and facilitating more precise temperature control.