Prediction of spatial X-ray dose distribution using pix2pix-based model

We develop a deep learning model based on an extended pix2pix framework to predict 3D spatial distributions of X-ray dose to realize real-time monitoring of medical staff's radiation exposure. Utilizing conditional generative adversarial networks, the model processes 3D voxel grids to estimate X-ray dose distributions rapidly, addressing the limitations of Monte Carlo simulations in real-time applications. Training employed simulated datasets generated via a Monte Carlo code: PHITS. The trained model achieves an approximately 150,000-fold speedup compared to the Monte Carlo simulations. While the model predicts distributions with characteristics similar to the true values, errors increase in regions shielded by the objects.