Abstract:
Accurate localization of metallic fragments embedded within biological tissues is critical in med- ical and trauma-related scenarios. This work presents a simulation-driven approach that com- bines Finite-Difference Time-Domain (FDTD) modeling with machine-learning (ML) techniques to estimate fragment depth using a 60GHz wideband radar system. A two-dimensional FDTD model was developed to simulate electromagnetic propagation through layered tissue structures containing metallic and vascular inclusions. Operating across a 4GHz bandwidth centered at 60GHz, the radar setup captured reflected Ez-field waveforms at multiple observation points to emulate realistic returns. Time-domain signals were transformed into high-dimensional feature vectors via spectrograms and Fast Fourier Transforms (FFT). These features trained ML regressors for depth prediction. A Random Forest (CPU) established a low-cost, interpretable baseline, while a GPU-accelerated XGBoost (RGBoost) model exploited parallelism to handle larger, more complex datasets and shorten training times. Both models achieved high localization accuracy, validating the end-to-end simulation pipeline. Future work will extend the framework to larger anatomical regions (hand, abdomen, thigh, shoulder, head), incorporate full-body modeling, and integrate a real 60GHz radar for experimental validation. The study lays the groundwork for a real-time, non-invasive imaging system with biomedical and defense applications.
