A direct learning approach for detection of hotspots in microwave hyperthermia treatments

Abstract: This paper presents a computational study for detecting whether the temperature values of the breast tissues are exceeding a threshold using deep learning (DL) during microwave hyperthermia (MH) treatments. The proposed model has a deep convolutional encoder-decoder architecture, which gets differential scattered field data as input and gives an image showing the cells exceeding the threshold. The data are generated by an in-house data generator, which mimics temperature distribution in the MH problem. The model is also tested with real temperature distribution obtained from electromagnetic-thermal simulations performed in commercial software. The results show that the model reaches an average accuracy score of 0.959 and 0.939 under 40 dB and 30 dB signal-to-noise ratio (SNR), respectively. The results are also compared with the Born iterative method (BIM), which is combined with some different conventional regularization methods. The results show that the proposed DL model outperforms the conventional methods and reveals the strong regularization capabilities of the data-driven methods for temperature monitoring applications.