SAR to Temperature: Preliminary Study of a U-Net-Based Model for Fast Temperature Variation Prediction in Biological Tissues Under Microwave Exposure

Hyperthermia is a clinically established therapeutic technique aimed at elevating tissue temperature to enhance the effectiveness of cancer treatment, often employed alongside traditional chemotherapy or radiotherapy. To ensure safe and targeted heating, accurate thermal modeling is imperative, as it helps clinicians determine how much heat is delivered to the treated region. Conventional methods that solve Pennes' bioheat transfer equation using finite difference or other numerical schemes are typically computationally intensive, limiting their use in real-time scenarios. To overcome these constraints, we propose a deep learning framework leveraging a U-Net convolutional neural network architecture to predict thermal distributions directly from specific absorption rate (SAR) data, which represent the microwave energy absorption within tissues. Through the incorporation of a linear output layer, our model delivers high-fidelity temperature estimations, offering a faster alternative to numerical methods. It achieves a 15-fold acceleration over the finite difference approach while maintaining an error below 0.003°C over 6 s of heating in the core region. The signal-to-noise ratio (SNR) of our predictions exceeds 25 dB, underscoring the accuracy and robustness of this technique. This innovative approach holds promise for enhancing treatment planning and monitoring of hyperthermia procedures in biomedical applications. By reducing computation times and maintaining precise temperature control, this methodology can be integrated into clinical workflows for improved patient safety.