A standardized and reproducible pipeline for fast multi-view digitally reconstructed radiograph (DRR) generation

Abstract Digitally reconstructed radiographs (DRRs) are synthetic projections derived from computed tomography (CT) data and are increasingly used in 3D synthetic image generation. However, variability in preprocessing and projection frameworks limits reproducibility and comparability across studies. We present a standardized pipeline for DRR generation from CTs ensuring geometric consistency, intensity normalization, and reproducible parameter control. The pipeline includes preprocessing with resampling, orientation normalization, and intensity windowing; optional rigid coregistration; and DRR generation using the GPU-accelerated TIGRE cone-beam model. CT volumes are resampled to 1 mm isotropic resolution, clipped to defined HU windows, and projected using the Beer–Lambert attenuation model. DRRs are generated at uniformly spaced angles from 0° to 180°. The pipeline was evaluated on 266 spine CTs from the VerSe2020 dataset using four HU windows and two projection geometries. Mean processing time was 1.12 s for preprocessing and 0.32 ± 0.03 s for generating five DRRs. Bone-windowed CTs produced realistic projections, while high-density bone windows yielded sharper cortical contours. Image realism was similar between projection geometries. This pipeline enables standardized DRR generation with consistent preprocessing, flexible geometry, and rapid computation, supporting reproducible research in synthetic imaging, registration benchmarking, and radiographic simulation.