Simulation Dataset and Code for: Cross-Site RF Bone Fracture Detectability in Simulation — Cylindrical FDTD Modelling, Site/Protocol Domain Shift, and Domain Adaptation

This repository contains the complete simulation dataset and Python source code accompanying the paper "Cross-Site RF Bone Fracture Detectability in Simulation: Cylindrical FDTD Modelling, Site/Protocol Domain Shift, and Domain Adaptation" (submitted to Computers in Biology and Medicine, 2026). DATASET The file unified_L4096.h5 (HDF5 format, 237 MB) contains 706 paired healthy/fractured bone simulations generated using a custom 2.5D cylindrical FDTD engine. Each simulation pair includes 21-channel time-domain signals (7 healthy, 7 fractured, 7 difference channels, each 4096 samples), 63 hand-crafted electromagnetic features, and metadata (bone site, fracture type, gap width, healing stage, carrier frequency, dataset phase). Simulation parameters: two anatomical sites (tibia at 2 GHz, femur at 1 GHz), five fracture types, seven gap widths (0.5-8.0 mm), seven healing stages, six operating frequencies, and patient-variation perturbations. CODE All Python scripts required to reproduce the classification results in the paper are included: - run_experiments.py: classical baseline experiments (SVM, Random Forest) - run_dl_experiments.py: deep learning baseline experiments - run_coral_experiments.py: CORAL domain adaptation experiments - run_nontransductive_check.py: non-transductive CORAL sensitivity check - generate_freq_sweep_fig.py: Figure 11 (frequency sweep analysis) - models_classical.py, models_dl.py: model architectures - physics_losses.py: physics-informed loss terms - dl_trainer.py: PyTorch training loop - data_loader.py: HDF5 data loading and feature extraction - evaluate.py: metrics and operating-point selection - config.py: hyperparameters and paths - requirements.txt: Python dependencies REQUIREMENTS Python 3.9+, PyTorch, scikit-learn, h5py, numpy, pandas, matplotlib. See requirements.txt for exact versions. CITATION Please cite the companion paper (Computers in Biology and Medicine, 2026, under review) when using this dataset or code.