Deep Learning-Based Electric Field Prediction in Realistic Breast Models for Microwave Medical Applications.
Authors
Affiliations (1)
Affiliations (1)
- Department of Electronics and Communication Engineering, Istanbul Technical University, Sarıyer, Istanbul 34469, Turkey, Istanbul, 34469, Turkey.
Abstract
Microwave medical systems for breast imaging and hyperthermia require iterative estimation of electric field distributions within heterogeneous tissues. Breast anatomy exhibits substantial geometric and compositional heterogeneity, with considerable variations in shape, tissue distribution, and dielectric properties across different cases. Accurate field estimation therefore requires high-resolution full-wave electromagnetic simulations, which are computationally demanding for time-sensitive clinical workflows. This study proposes a deep learning (DL)-based surrogate modeling approach for rapid prediction of electric field distributions in heterogeneous breast tissues under a fixed antenna configuration. The dataset consists of high-resolution dielectric property maps derived from Magnetic Resonance Imaging (MRI)-based breast models at 2.4 GHz. An anatomically consistent augmentation strategy was applied to increase data variability while preserving realistic tissue distributions, expanding the dataset to 4,878 breast slices. Corresponding electric field distributions were generated using a finite element method (FEM)-based solver as ground-truth labels for training and evaluating ResNet and U-Net architectures. A masked loss function was also employed to focus learning on heterogeneous breast tissue regions. Performance was evaluated on 504 test samples. ResNet achieved an average Signal-to-Noise Ratio (SNR) of 10.01 ± 2.84 dB and a Normalized Mean Squared Error (NMSE) of 0.130 ± 0.113, whereas U-Net achieved 10.83 ± 2.68 dB and 0.105 ± 0.087, respectively. The masked-loss strategy further increased the average U-Net SNR by 0.12 dB. The results demonstrate that DL-based surrogate models can accurately reconstruct electric field distributions while avoiding repeated full-wave simulations, supporting rapid field prediction for microwave imaging and hyperthermia treatment-planning workflows.