Estimating Spatially Resolved Radiation Fields Using Neural Networks

TL;DR

The paper tackles the slow real-time capability of Monte Carlo radiation transport in interventional radiology by introducing surrogate neural-network models that reconstruct spatial fluence and local spectra from synthetic Geant4-based datasets. It compares NeRF-like fully connected networks with a 3D U-Net, incorporating spectrum-aware encodings and FiLM conditioning, across increasingly complex datasets. Key contributions include three publicly available datasets, architectural insights, and practical design recommendations for accurate, voxel-level radiation-field reconstructions. The work advances real-time visualization and training tools for IR dosimetry, with pathways toward uncertainty handling and dynamic geometry in future work.

Abstract

We present an in-depth analysis on how to build and train neural networks to estimate the spatial distribution of scattered radiation fields for radiation protection dosimetry in medical radiation fields, such as those found in interventional radiology and cardiology. We present three different synthetically generated datasets with increasing complexity for training, using a Monte-Carlo Simulation application based on Geant4. On those datasets, we evaluate convolutional and fully connected architectures of neural networks to demonstrate which design decisions work well for reconstructing the fluence and spectra distributions over the spatial domain of such radiation fields. All our datasets, as well as our training pipeline, are published as open source in separate repositories.

Figures (5)

• Figure 1: Distribution of relative voxel fluences inside datasets DS-01 and DS-02 for range $\Phi^{3}_{\mathrm{\gamma}} \geq 0.0$ on the X-axis, exhibiting a strong imbalance towards $\Phi^{3}_{\mathrm{\gamma}} \leq 10^{-2}$.
• Figure 2: Distribution of relative voxel fluences inside datasets DS-01 and DS-02 for range $\Phi^{3}_{\mathrm{\gamma}} \geq 10^{-1}$ on the X-axis, exhibiting a much weaker imbalance in that range.
• Figure 3: Distribution of relative voxel fluences inside dataset DS-03 for range $\Phi^{3}_{\mathrm{\gamma}} \geq 0.0$ on the X-axis, exhibiting a strong imbalance towards $\Phi^{3}_{\mathrm{\gamma}} \leq 10^{-5}$.
• Figure 4: Distribution of relative voxel fluences inside dataset DS-03 for range $\Phi^{3}_{\mathrm{\gamma}} \geq 10^{-1}$ on the X-axis, exhibiting a strong imbalance towards $\Phi^{3}_{\mathrm{\gamma}} \leq 20^{-1}$ and a small peak at $\Phi^{3}_{\mathrm{\gamma}} = 1$.
• Figure 5: Schematic of the fully connected architecture used as the backbone for SRBFNet, SPERFNet and PBRFNet outputting a spectrum and a fluence for each voxel location.