“There is no threshold dose below which radiation can be considered completely harmless.”
— Hermann J. Muller

Although originally formulated for radiation in general, this statement applies to X-ray imaging
as well, given its ionising nature. Even low-dose diagnostic procedures can damage DNA and
carry cumulative cancer risk, which has led radiologists to minimise patient exposure whenever possible. These safety constraints limit the acquisition of large, diverse imaging datasets needed for developing, validating, and benchmarking modern medical imaging systems. They also restrict the number of projections that can be acquired in modalities such as computed tomography (CT), requiring accurate volumetric reconstruction from fewer scans and thereby increasing the technical demands on reconstruction algorithms.

Because acquiring realistic X-ray images is limited by patient safety, modern approaches first use sparse CT scans to reconstruct an accurate volumetric model of the object of interest. From this model, synthetic images can be generated through novel view synthesis, enabling large-scale offline datasets for machine learning, system testing, and model validation without additional radiation exposure. Beyond dataset creation, these synthetic projections can be produced in real time for interactive applications such as digital twins, used in virtual physician training and integration testing. However, generating high-fidelity synthetic images in real time remains challenging, given the substantial computational requirments of the algorithm.

This thesis investigates ways to accelerate X-ray simulation using graphics processing unit (GPU) implementations. Two techniques were developed: one based on voxelised models and another using Gaussian mixture models (GMMs). The approaches were evaluated in terms of visual fidelity and rendering performance, achieving ≈300 frames per second for voxel-based simulation and ≈40 frames per second for GMM-based simulation. Both techniques significantly reduce
computation time compared to baseline CPU implementations, while maintaining realistic image quality suitable for virtual testing, physician training, and AI data generation.

These results demonstrate that GPU acceleration can enable real-time synthetic X-ray simulation, supporting scalable dataset creation and interactive applications while maintaining strict adherence to radiation safety principles.