Deep Learning-Based Temperature Distribution Estimation for Microwave Hyperthermia Applications

Accurate temperature measurement is essential in microwave hyperthermia to render cancer treatment efficient and safe. Traditional imaging methods are hampered by low resolution and include complex inversion schemes or are noise-sensitive. In this paper, we propose a U-Net deep learning architecture that is capable of predicting 2D temperature distributions directly from complex-valued differential scattered fields. The model is trained and tested on synthetically generated datasets that simulate real breast tissues under various heating conditions. Performance is compared under different signal-to-noise ratio (SNR) conditions, including SNR $_{\text {train }}=30 \mathrm{~dB}$ and 40 dB, with test performances ranging from $\mathrm{SNR}_{\text {test }}=10 \mathrm{~dB}$ to 60 dB. The results show relative errors of $\mathbf{1. 3 7 \%}$ and $\mathbf{1. 5 5 \%}$ for models trained at 40 dB and 30 dB, respectively. Besides, our method is more accurate and reliable compared to conventional inversion techniques. These findings show the potential of data-driven temperature estimation models in clinical hyperthermia systems.