RadField3D: A Data Generator and Data Format for Deep Learning in Radiation-Protection Dosimetry for Medical Applications

TL;DR

This work introduces RadField3D, an open-source Geant4-based Monte-Carlo application that generates spatially resolved, voxelized radiation-field datasets for dosimetry in interventional radiology. It is complemented by RadFiled3D, a fast binary data format with a Python API designed for machine-learning workflows and rapid data loading. The authors validate RadField3D against ISO 4037-1 measurements using two configurations and phantoms, deriving a conversion factor $S_c$ to map relative simulated quantities to absolute measurands such as air kerma rate. They implement a voxel-scoring strategy with dynamic statistical-error tracking via Welford's online variance and a voxel-voxel energy-distribution framework, enabling robust termination criteria and data usability for ML. Overall, RadField3D and RadFiled3D aim to improve reproducibility, interoperability, and speed in generating dosimetric datasets, thereby supporting real-time dose estimation methods and data-driven approaches in radiation protection for medical applications.

Abstract

In this research work, we present our open-source Geant4-based Monte-Carlo simulation application, called RadField3D, for generating threedimensional radiation field datasets for dosimetry. Accompanying, we introduce a fast, machine-interpretable data format with a Python API for easy integration into neural network research, that we call RadFiled3D. Both developments are intended to be used to research alternative radiation simulation methods using deep learning.

Figures (9)

• Figure 1: Wireframe render of one single photon track, crossing a scatter phantom geometry, with colored track-components. Red: Direct beam component (Beam). Green: Scatter component inside the patient (Patient). Blue: Scatter component outside the patient (Scatter).
• Figure 2: Schematic top-down view of our measurements setup. The parameters for each of our measurement configuration, A and B, are annotated at each value. Some parameters like the X-ray tube distance and the opening angle of the primary beam were fixed during all measurements. The phantom was a cylindrical water barrel during configuration A that was exchanged with a male Alderson phantom torso for configuration B.
• Figure 3: Render of the reconstructed CT-scan of our male Alderson phantom torso as it was used in experiment B.
• Figure 4: Schematic view of our RadFiled3D Format structure consisting of channel and layer blocks, preceded by a metadata block. For better readability, the metadata block only shows the metadata elements partially. Please note, that channel and layer blocks can be repeated.
• Figure 5: Experiment A.1: $19.5\,\mathrm{cm}$ distance of the detector from the center of the water phantom.