Author: Yankun Lang, Lei Ren, Dario B. Rodrigues 👨🔬
Affiliation: University of Maryland School of Medicine, Department of Radiation Oncology, University of Maryland School of Medicine 🌍
Purpose:
HTP of microwave (MW) phased-array systems determine MW antenna settings to maximize energy absorption (SAR in W/kg) in tumor. Conventional HTP algorithms calculate SAR based on electromagnetic field-tissue interactions. This process is slow, computationally demanding, preventing real-time optimizing the MW antenna settings to achieve optimal SAR distribution. To address this challenge, we developed a deep learning method for fast and accurate SAR prediction, which we will test in a brain tumor model.
Methods:
Our deep learning model uses an encoder-decoder architecture, leveraging four cross-attention blocks within the encoder to integrate antenna phases with brain electrical properties and its 3D coordinates. The model matches SAR predictions with ground-truth using L2 loss. Our dataset consisted of 201 samples generated by changing the tissue properties, brain tumor position, and antenna phases. The hyperthermia applicator is comprised of three rings with 24 antennas each, where the external rings are synchronized, resulting 48 unique phases. We randomly selected 160 data for training, 21 data for validation, and the rest 20 data for testing. The SAR ground truth for model training was computed on a human head model with a 2-cm brain tumor using a finite-element-based software (COMSOL Multiphysics).
Results:
The prediction was evaluated by root-mean-squared-errors (RMSE) and mean-absolute-errors (MAE). The RMSE achieved a mean value of 3.27 W/kg, while the MAE achieved 1.56 W/kg, where the mean SAR is 7.30 W/kg. The structural-similarity-index-measure (SSIM) achieved a mean value of 0.89. Notably, our deep learning algorithm requires <1s for prediction, much faster than 10 minutes required by conventional algorithms.
Conclusion:
Our study demonstrated high efficiency and accuracy of our deep learning method for SAR prediction in HTP. This enhanced efficiency creates an opportunity for real-time adaptive HTP to optimize SAR and consequently temperature during hyperthermia treatments, which is correlated with improved clinical outcomes.