Spatiotemporal estimations of temperature rise during electroporation treatments using a deep neural network.

The nonthermal mechanism for irreversible electroporation has been paramount for treating tumors and cardiac tissue in anatomically sensitive areas, where there is concern about damage to nearby bowels, ducts, blood vessels, or nerves. However, Joule heating still occurs as a secondary effect of applying current through a resistive tissue and must be minimized to maintain the benefits of electroporation at high voltages. Numerous thermal mitigation protocols have been proposed to minimize temperature rise, but intraoperative temperature monitoring is still needed. We show that an accurate and robust temperature prediction AI model can be developed using estimated tissue properties (bulk and dynamic conductivity), known geometric properties (probe spacing), and easily measurable treatment parameters (applied voltage, current, and pulse number). We develop the 2-layer neural network on realistic 2D finite element model simulations with conditions encompassing most electroporation applications. Calculating feature contributions, we found that temperature prediction is mostly dependent on current and pulse number and show that the model remains accurate when incorrect tissue properties are intentionally used as input parameters. Lastly, we show that the model can predict temperature rise within ex vivo perfused porcine livers, with error <0.5 °C. This model, using easily acquired parameters, is shown to predict temperature rise in over 1000 unique test conditions with <1 °C error and no observable outliers. We believe the use of simple, readily available input parameters would allow this model to be incorporated in many already available electroporation systems for real-time temperature estimations.